A dated public record

6 May 2026 — Reconsidering the architect/builder framing; scope of this corrections page named openly

Category: framing · workflow · scope

The architect framing is being restored, and the previous retraction is being itself retracted. On 1 May 2026 the publication retracted earlier framings that named Doug Scott as "the architect" and the AI tools as "builders" or "external reviewers." The retraction was made in response to AI cross-critique that read the architect framing as overclaiming — implying technical expertise, direct authorship of the prose, and independent verification of citations. The publication accepted the critique and moved to a more minimal framing: "the human prompted, answered, scanned, shipped."

On reflection the retraction over-corrected. The minimal framing is honest about what Doug did not do, but understates what he did do. Doug held the publication's intent across all the work — what it was for, who it was for, which questions belonged where, when an analytical position was honest and when it was not, when the AI output had drifted from the intent and needed to be redirected. He did not edit the prose, did not check citations against primary sources, did not verify model math; the minimal framing remains correct on those specific points, and the no-human-expert-review disclosure continues to apply. But he did the work an architect does: he held the intent, he chose the structure, and he decided when the work was done. The AI tools wrote, analysed, modelled, and cross-critiqued each other; they did the work of builders and reviewers (with the limit, named in the 1 May entry, that AI cross-critique is not the same as independent specialist verification). The original "architect / builders / checkers" framing was accurate in this register, and its retraction was a mistake.

The publication is restoring the framing in the per-article production line, the AI summary block at the foot of every page, the about page, and the llms.txt machine-readable description. This entry records the reconsideration and the reasoning. The April 1 entry retracting the architect language remains visible in this log alongside this entry; both are part of the publication's history and a reader can see how the position has moved.

The mechanism, named concretely. Restoring the architect framing is the right move only if the workflow it describes is also named precisely. The publication was produced by Doug running parallel conversations with the four AI tools and routing work between them by hand. He pasted Claude's argument into ChatGPT for critique, took ChatGPT's pushback to Grok for a different angle, fed Gemini the resulting text and asked it to find what was missing. The cross-critique was a loop he routed manually. The AI tools did not communicate with each other autonomously; nothing about the workflow was four AIs in independent dialogue. Doug was the connective tissue. He decided which output to keep, which to feed to which AI next, and when the loop had converged. That manual routing is what the publication has been calling "rounds of substantive critique." The phrase is accurate but the mechanism it names is small and human, and the publication has not previously stated this concretely. It is stating it now. The workflow disclosures across the site (footer, AI summary block, homepage production callout, about page, llms.txt, per-article production line at the foot of each piece) have been updated to name the cut-and-paste routing explicitly. A reader weighing the publication's claim to disciplined production should weigh this concrete description rather than the abstract "cross-critique" phrasing it replaces.

What this entry does not change. Doug is not a tax specialist, a lawyer, an accountant, or a labour economist; the publication continues to disclose that no human expert with relevant domain expertise reviewed any piece before publication. The author's conflict of interest continues to be disclosed at the top of every relevant piece. The corrections page continues to record changes openly. The 1 May open-question note (the AI-loop diagnosis above) continues to hold — the AI tools checking each other, even when routed by hand, is not the same as a human specialist checking the work, and the publication does not claim it is.

The corrections-page scope is itself something this page should be honest about. The page is the publication-level record of substantive changes, but the changes recorded in itemised form below have all been made during Claude sessions that produced the published builds. Substantive earlier work in other AI sessions — including other Claude sessions, ChatGPT sessions, Grok sessions, and Gemini sessions — is not recorded here in itemised form. The earlier work covered, among other things: the initial publication structure; multiple rewrites of the IHT analysis and its position-architecture; the venture-capital body of work and its critique cycles; the original architect/external-reviewer framing and its first retractions; the choice of name for the publication itself.

The name choice is worth naming explicitly because it shapes what the publication is. The Longer Look was chosen via a correction made in another Claude session against an original list of 100 candidate domain names that had been generated from a misread of the brief. The misread treated the publication as a place where cases would be made; the correction reframed it as a place where things are placed down at length and left there for the reader to engage with. The reframing was Doug's, in a single sentence: "a debate here is that I wish to not argue." Most of the original 100 names did not survive that reframing because they were verbs of contention (argument, case, claim, response); the survivors named the stance or the form (plain, longer, look, quiet, note). The final shortlist was three: theplainthing.com, thelongerlook.com, and aquietclaim.com. The publication's stance — describing rather than pleading — is downstream of that reframing. The reframing is not recorded as a correction-page entry below because it predated this Claude session; it is recorded here, now, openly, as part of the scope acknowledgement.

A reader who wants to weigh the publication's claim to disciplined error-correction should weigh it against this scope, not against an idealised log that captures every change ever made. The corrections recorded below begin from the build state visible to the current Claude session and document changes made within it. Earlier corrections, where they reshaped what the publication is, have left their fingerprints in the publication's structure and voice but are not enumerated here in dated form. The publication considers this distinction important enough to name openly rather than leave a reader to discover.

The cumulative pattern, named. The 1 May AI-loop diagnosis was that the corrections page was documenting a loop, not converging on a fixed point. That diagnosis was directionally correct and the open-question box above continues to hold it as an unresolved tension. But the loop also produced real architectural shifts: the four-positions menu became five; the principle and timing pieces moved to equal-length two-sided presentation without verdict; the architect framing was retracted and (today) reconsidered; the publication's name itself was chosen via a correction. The loop is real and the convergence is partial. A reader weighing the work should hold both at once: this is what AI-assisted intellectual work looks like when the audit trail is kept public, including the iterations that look like running in circles and the iterations that produce structural change. The publication does not claim to have resolved which iterations were which.

10 May 2026 — UK migration body of work updated. Twenty-eighth piece added: a costed cross-party companion that takes each party's stated proposals and prices them with HIGH/MEDIUM/LOW confidence labels. Nine party briefings updated to include a "costed implications: short summary" block.

Category: addition · substantive

A second iteration of the UK migration body of work has been integrated. The set goes from twenty-seven pieces to twenty-eight with the addition of a costed cross-party companion — the analytical view from outside each party's worldview, paired with the nine existing party briefings which are written from inside each worldview at full strength. The companion takes each of the nine parties and produces, for each: the party's stated proposals as published in their May 2026 platform; proposal-by-proposal cost ranges, savings/revenue ranges, and net fiscal effect with HIGH/MEDIUM/LOW confidence labels; implications from inside the party's framing AND from external analytical perspective; deliverability constraints (operational, legal, capacity, timeline); legal exposure (ECHR, Refugee Convention, Belfast/Good Friday Agreement, TCA, retrospective-application risks); and likely behavioural responses by employers, migrants, source countries, returnee cohorts, and asylum-seeker behaviour. A comparative summary table at the end aggregates across all nine parties.

Why this addition. The party briefings, written from inside each worldview, are the strongest version of each party's case. They are deliberately directional. The costed companion is the analytical pair: written from outside, asking what the proposals would actually cost, do, and provoke if implemented as stated. The two views are now both available on the publication and a reader can hold them together — the strongest case from inside, the costed reality from outside — rather than having to choose. The publication still does not adjudicate. It now provides both lenses for each of the nine parties.

Confidence discipline. The companion uses three confidence labels at every cost line. HIGH means official published costings or strong directly-applicable evidence (NAO asylum accommodation costs; MAC fiscal modelling; HMRC/Home Office linked earnings data). MEDIUM means derived from official data with reasonable assumptions, or published-but-unaudited modelling. LOW means behavioural responses, retrospective-application effects, or claims based on weak/contested evidence. Where confidence is LOW, the range is wide and the wording cautious. This is honest disagreement with reality, not hedging — some proposal effects are genuinely difficult to forecast, and the publication says so where it applies.

Party briefings updated. Each of the nine party briefings (Labour, Conservative, Lib Dem, Green, Reform UK, Restore Britain, SNP, Plaid Cymru, DUP) gains a "costed implications: short summary" block of approximately 400 words at the end. The block contains the headline costed table for that party, the top three analytical upsides, the top three analytical downsides, and a pointer to the full companion for detail. The body content of each briefing — the position-from-inside argument — is unchanged.

Cross-references. Each party briefing now cross-references the costed companion in its see-also section as "the same proposals from outside the worldview." The companion in turn cross-references the flagship overview, the policymaker pack, and a sample of party briefings to point readers between inside-view and outside-view treatments of the same proposals. The marks placed on the costed companion are coin (signalling fiscal) and parliament (signalling cross-party comparison).

Count references updated. The migration category indexStandfirst, the homepage's by-body-of-work list, the archive standfirst, the reading-guide intro, the reading-guide migration entry, and the about-page section explainer all updated from "twenty-seven pieces" to "twenty-eight pieces" with the costed companion named where appropriate. The 9 May 2026 corrections-log entry that records the original twenty-seven-piece launch is preserved unchanged as a historical record — the trail of what the publication looked like at each point is itself part of the audit record.

10 May 2026 — UK migration body of work added (twenty-seven pieces). Fifth section of the publication. Archive standfirst, homepage body-of-work list, about-page section explainer, and reading guide all updated to reflect the addition.

Category: addition · substantive

A reference on UK migration and benefits policy as of May 2026 has been added to the publication. The body of work consists of twenty-seven pieces in the publication's standard analytical register: a flagship overview; three audience packs (journalist, policymaker, public); seven framings of the same evidence base from different intellectual traditions (cohesion, refugee protection, demographic, AI labour market, public-service capacity, emigration, post-Brexit sovereignty); nine party briefings written from inside each party's worldview to make the strongest version of that party's case (Labour, Conservative, Lib Dem, Green, Reform UK, Restore Britain, SNP, Plaid Cymru, DUP); four stakeholder briefings (business and employer bodies, trade unions and worker representation, senior civil service, local government); and three standalone deep-dives on the topics most contested in public debate (the 2022-2024 ILR cohort, housing supply, crime and trust). The full section index is at /uk-migration.

Why this body of work fits the publication. Migration is the policy question with the largest political salience in the UK in 2026. The publication's discipline — positions presented at strength, no adjudication, framings named openly — applies to migration with particular force, because the public debate is currently conducted at high volume and low resolution and the analytical work of presenting each position at the strength its proponents would is mostly not happening. The body of work does not advocate a single policy direction; it lays out the evidence, the available policy options, what each major political party would do, and the framings that select and weight the same evidence differently. Readers reach different conclusions depending on which framing they treat as primary; the publication is structured to make those dependencies visible, not to resolve them.

Workflow. The set was produced through the publication's standard four-AI cross-critique loop: Doug Scott (publisher, prompter, editor) with Claude Opus 4.7, ChatGPT, Grok, and Gemini contributing — all four AI tools fed into the work; Claude Opus 4.7 was the synthesiser that pulled the threads together. AI-generated, no human expert review. The byline is consistent with the rest of the publication.

Empirical grounding. The numerical claims throughout the set trace back to source keys in an underlying 40-tab data workbook (HMRC, ONS, MAC December 2025, OBR, NAO May 2025, Home Office annual reports, DWP UC by status, NRPF Connect, parliamentary statements, peer-reviewed academic work). The full master document is also available for download — the 78,000-word reference work from which the published pieces are extracted. Each piece carries a confidence-labelling discipline; Crime, Trust, and the Debate uses high / medium / low confidence labels at every claim level because the topic is at the centre of trust collapse in migration policy and warrants particular epistemic discipline.

What was updated elsewhere on the site. The archive standfirst, the homepage's by-body-of-work list, the about-page section explainer, the reading guide, and the migration category index are all updated to acknowledge the addition. The site now carries four analytical bodies of work (IHT, VC, Mars, Migration) plus the single Notebook piece. The masthead navigation gains a Migration link.

What is not in the set. The set is a reference on UK migration policy as of May 2026; it does not advocate, it does not include polemical material, and it does not adjudicate between the seven framings or the nine party positions. The directional pieces (party briefings, framing articles, stakeholder briefings) are explicitly written from inside specific worldviews to make the strongest version of each case. The flagship and audience packs are evidence-led but fiscally framed. The set does not cover non-UK migration questions, asylum systems outside the UK, or migration economics at theoretical depth; readers seeking those should look elsewhere.

9 May 2026 — Two external audits applied. Revenue-forecast figure updated to ~£295m (current GOV.UK), VC piece count made dynamic, archive taxonomy reframed as four sections, model-multiplier text aligned to slider default, and Mars's place in the publication explained on the about page.

Category: data · framing · cleanup

Two external audits arrived: one a structural/UX audit covering taxonomy, framing, and consistency; the other a data audit checking the publication's numerical claims against primary sources (GOV.UK, Commons Library briefings, parliamentary statements). Both surfaced real issues. This entry records the response to the actionable findings; the editorial-judgement findings (whether to soften the “side-door” framing, whether to reduce disclaimer repetition, whether the corrections page risks looking like an internal argument) were considered and the publication's existing posture preserved — the bluntness of the disclosures and the transparency of the corrections process are load-bearing for what the publication is.

Data fixes. The revenue-forecast figure for 2029-30 has been updated to ~£295 million per current GOV.UK guidance, with the £300 million figure from Commons Library briefing CBP-10181 (December 2025) noted alongside in the for-journalists piece's confidence paragraph. The two figures are within rounding of each other; the data audit was correct that the most recent published figure is the £295m one. The five-minute-version's “rising to roughly £300 million a year” was updated to “~£295 million.” The interactive model's text “the model uses 1.20 in its central case” was inconsistent with the slider default of 1.30x; the text was aligned to 1.30 to match the slider. The other empirical claims the data audit checked (£2.5m / £5m thresholds, 50% relief / effective 20% rate, ten-year interest-free instalments, ~1,100 estates affected, ~220 BPR-only estates excluding AIM-only) were confirmed against primary sources and remain unchanged.

Taxonomy fixes. The archive standfirst said “in two bodies of work — analysis of the April 2026 UK IHT reform, and analysis of venture capital” but the archive then listed four sections (IHT, VC, Mars, Notebook). Reframed as “four sections” with each named. The reading-guide standfirst said “three bodies of work … plus a single Notebook piece,” which is mathematically four content areas; reframed as “four sections.” The cross-category description on the homepage's body-of-work list was updated to name the IHT and venture-capital analyses specifically rather than “the two bodies of work.” The IHT archive section note now reads “23 pieces in total: 14 featured on the homepage and 9 alternative versions, methodology pieces, and critiques-and-responses pages,” which is clearer than the previous version that gave the featured count without the total.

VC count. The homepage's “by body of work” list said “nine pieces across three tiers” for the venture-capital section, while the archive and the venture-capital index page said 11. The homepage list is now dynamic (uses visibleByCategory('vc').length) so the count is automatically correct as pieces are added or removed. Mars was also missing from the homepage's body-of-work list entirely; it has been added.

Mars on the about page. The about page framed the publication as side-door work for the AI/humanity project but did not explain why a seven-document Mars set sits inside a publication framed that way. A new paragraph names the three analytical bodies of work (IHT, VC, Mars) plus the Notebook piece and explains Mars's place: it is the same analytical method (positions presented at equal length, frame disclosure, no closing verdict) applied to a question about the largest scale at which humans might consciously intend to industrialise something new. The connection is that Mars is the same kind of question the publication exists to treat — a contested public-interest question argued at high volume and low resolution. The paragraph also notes that each section is independent and readers do not need to follow all four to engage with any one.

What was not changed. The audits suggested several editorial moves the publication considered and declined: softening the “side-door” framing (it is intentional and load-bearing — the bluntness goes on the about page, not on the analytical surface, and that distinction is described openly); reducing disclaimer repetition (the disclaimers are part of the publication's posture, not a regulatory fig leaf, and removing them would compromise the work); simplifying the corrections page's top note for normal readers (the corrections page is for readers who want to see the audit trail, and that trail being legible to a careful reader is more important than being inviting to a casual one). The editorial decisions are recorded here so that a reader who reaches different conclusions about these trade-offs has visibility on what was considered and rejected.

9 May 2026 — Byline reframed: all four AI tools contributed substantively, Claude Opus 4.7 was the synthesiser. Previous framing of "Claude as primary tool with the others as reviewers and amenders" overstated Claude's sole role and understated the contributions of ChatGPT, Grok, and Gemini.

Category: framing · honesty

Doug noted that the byline as previously phrased — “written with Claude Opus 4.7 as the primary tool, with Grok, ChatGPT, Gemini and a second Claude instance acting as reviewers and amenders” — mischaracterised what actually happened. The honest description is that all four AI tools contributed substantively to the writing, the analysis, the structure, and the framing across every piece on the publication. Claude Opus 4.7's role was the synthesiser — the tool that held the threads across iterations and pulled the work together — not a primary author with the others reduced to reviewers. The previous framing made it sound as though Claude wrote and the other three checked. That is not what produced this publication.

What has been changed. Every byline string in the publication's configuration has been reframed. The short-form byline on each piece now reads “Doug Scott, with Claude Opus 4.7, ChatGPT, Grok, and Gemini.” The longer joint-byline disclosure block on each piece now reads “Doug Scott (publisher, prompter, editor) with Claude Opus 4.7, ChatGPT, Grok, and Gemini contributing — all four AI tools fed into the work; Claude Opus 4.7 was the synthesiser that pulled it together.” The disclosure block on every article (the side-door-note that appears at the bottom of each piece) now reads similarly. The archive page's VC-section meta line is updated to match.

What this changes for the reader. Substantively, very little. The previous framing and the new framing both describe the same workflow — Doug ran parallel conversations with all four tools and routed work between them by hand — with the difference being how that workflow is named. The new framing is more accurate about what each tool contributed. A reader who weighs the publication's reliability against its disclosure of method should weigh it against the more accurate framing: the work was produced by four AI tools collectively, with Claude Opus 4.7 in the synthesis role. No one tool wrote the work alone, and the previous framing's implication that Claude was the writer with the others as reviewers was not a true description.

Where the previous framing remains. Earlier corrections-log entries that quote the previous byline phrasing (the 1 May 2026 architect retraction-and-restoration entry, the 6 May 2026 byline reframe to “primary tool plus reviewers”, the 9 May 2026 Position F entry above) preserve the language used at the time. By the publication's discipline, historical corrections-log entries are not retroactively edited; the trail of what the publication said at each point is itself part of the audit record. A reader following the corrections log forward in time can see the evolution of how the publication has described its own production method, including the previous framings that this entry now corrects.

What this does not change. The conflict-of-interest disclosure remains (the author is a UK technology founder; his personal tax position has been settled by planning that took place independently of the policy debate; he has invested in hundreds of UK early-stage tech companies). The licence remains (CC BY-NC 4.0). The no-human-expert-review disclaimer remains and is unchanged: no tax specialist, lawyer, accountant, or other domain expert reviewed any piece on the publication before it was published. The substantive analysis on every contested question remains unchanged; what changed is only how the cross-AI authorship is named.

9 May 2026 — Position F gains an agency layer; new section "What the proposal returns to founder families" added between case-for and case-against

Category: framing · substantive

A follow-up exchange between Doug, Claude (a separate session from the one writing the publication), and ChatGPT named a layer of the Position F argument the original draft did not surface explicitly: that the proposal's strongest political case is not about tax mechanics or even about the operational dissolutions of the valuation and liquidity problems, but about what it returns to founder families. The current reform creates a problem founders cannot plan around because the timing of death is the one variable the founder does not control. The ten-year interest-free instalment regime softens the cash-bill side of the problem; it does not address the timing side. What founders considering staying or leaving are weighing, in many cases, is not the tax itself but exposure of their family to a forced event at an unknowable moment, on a calendar set by death rather than by commerce. Position F preserves the Treasury's long-run revenue through the date-of-death floor and the year-ten backstop, while restoring to families the ability to choose when within the decade the tax event falls.

What has been added. A new section, What the proposal returns to founder families, has been inserted in the long-form Position F piece between the case-for and the case-against, in the same length and register as the surrounding sections. The case-for paragraph in the long article's Position F treatment now ends with a one-sentence reference to the agency layer and a link to the dedicated section. The five-minute overview gains a fifth case-for paragraph naming the agency point. The case-against considerations on horizontal equity and administrative consistency are unchanged; the agency layer does not displace them.

An honest caveat is included. The agency argument works most strongly for founders whose families are involved in the company or its succession, or are likely to become involved through inheritance. It works less strongly for founders whose heirs have no commercial relationship to the business and are essentially passive recipients of value. For that population, the agency argument shades toward an argument about deferring tax rather than about preserving family decision-making in a commercial context, and the horizontal-equity considerations bite harder. The new section names the distinction openly rather than letting the strongest version of the argument carry the weaker version's cases.

Provenance note. The agency framing was developed in a follow-up exchange involving Claude and ChatGPT after the initial Position F draft. This is consistent with how every piece on the publication has been produced: all four AI tools (Claude Opus 4.7, ChatGPT, Grok, Gemini) contribute substantively, with Claude Opus 4.7 pulling the threads together. The byline reflects this. The reason this entry records the specific provenance is that a reader who wants to weigh the work should know which AI tools contributed which framings to which sections; the cross-tool routing is what the byline names, and the substance is the publication's.

Where this is integrated and where it is not. The long-form Position F piece, the long article's Position F treatment, and the five-minute overview are now updated. The other pieces flagged as outstanding in the 9 May 2026 corrections entry (the timing piece, the principle piece, common-reactions, funding stack, for-tax-practitioners, for-journalists, for-uk-tech-founders, plain-english versions, model page, how-this-was-made, downloadable companions) remain at their existing five-position framing and do not yet carry the agency-layer treatment. A reader who wants the canonical six-position-with-agency-layer treatment should read the long article, the readable companion, and the two dedicated F pieces.

9 May 2026 — Position F (founder election with a decade cap) added; five-positions menu becomes six-positions menu; surrounding language updated in the long article and readable companion

Category: framing · substantive

A submission to the corrections-and-contributions route proposed a sixth position on the timing-and-mechanism question: a per-company estate election at death between settlement at death under the existing reform (Regime 1) and a deferred-realisation regime with a hard ten-year backstop (Regime 2). The taxable transfer remains the death event; what changes is when the calculated tax falls and what it is paid against. The proposal is the bounded version of Position B — the version that survives the case-against-B in the publication's timing piece (the Australian forty-year experience of indefinite deferral, valuation-gaming at the cost-base-setting moment, and a family-trust deferral industry). The year-ten deemed disposal closes the door on the unbounded-deferral failure mode; the date-of-death floor on the taxable base protects the Treasury against the asymmetric-option failure mode; the deferred CGT uplift keeps the principle clean (the heir is not taxed twice, but the uplift does not arrive ahead of the IHT that justifies it); the anti-avoidance perimeter (connected-party deemed disposals at independent valuation, exit charges on departure, asset-stripping triggers, continuing-qualification conditions) closes the obvious gaming routes, mostly using legislative concepts that already exist in analogous form across UK tax legislation.

Label collision. The submission was drafted as “Position E”. The publication's operational analysis already uses Position E for the reform as written reference case, added 1 May 2026 in response to a structural critique that the four-positions menu had omitted the actual government position. The new proposal has therefore been relabelled Position F for consistency with the publication's existing five-position menu. The substance of the proposal is unchanged; references to Position F in the new pieces correspond to what the original draft called Position E. Two alternatives — rewriting the existing E or renaming it — were considered and rejected. Rewriting E would have removed the actual government position from the menu, re-introducing the very stacking the 1 May correction was added to fix. Renaming E would have required touching ten or more existing body files in cross-cutting ways, with no analytical gain. The label change is recorded here openly and an editor's note appears at the top of both new pieces.

What has been added. Two new pieces. Position F — A Founder Election with a Decade Cap is the long-form treatment (~3,300 words): the mechanism in detail, the CGT uplift question and why deferred uplift is the cleanest design, the case for and case against on their strongest terms, what would have to be true for Position F to be the right answer, and how it relates to A, B, C, D, and E. Position F — The Five-Minute Version is the short overview for readers who want the proposal in five minutes.

Where Position F has been integrated. The long article (full body file) lead paragraph and Section 1 now name six positions; the “What is actually in dispute” section names six positions; the operational treatment now includes Position F in a two-paragraph case-for / case-against block in the same length and register as A, B, C, D, and E; the Position D interest-disclosure paragraph now references the equal-weight presentation across A, B, C, D, E, and F. The readable companion has been updated with the same six-position structure. The homepage hero callout now names Position F alongside the existing five, with a direct link to the long-form Position F piece.

Where Position F has not yet been integrated. The When-Not-How-Much timing piece, the principle piece, the common-reactions piece, the funding-stack piece, the for-tax-practitioners piece, the for-journalists piece, the for-uk-tech-founders piece, the plain-english pieces, the model page, the how-this-was-made piece, and the downloadable .docx and .pdf companions all reference the five positions in their bespoke phrasing and have not been individually updated. The omission is recorded openly here. A reader engaging with those pieces should treat them as referring to the existing five operational positions (A through E) with F now added in the long article, the readable companion, and the two dedicated F pieces as the canonical six-position treatment. The full integration of F across every piece is outstanding work that should be done before any of those pieces are used as standalone references; the long article, the readable companion, and the two F pieces are now correct and a reader who wants the canonical six-position treatment should read those.

What this addition does and does not change. The publication's posture across the existing positions — that the choice between them depends on empirical questions that are not currently settled, and that the publication does not adjudicate — applies to F with the same force. F is not a recommendation, it is a presentation of F at full strength alongside the costs of adopting it, in the same analytical register as A through E. The author's conflict-of-interest disclosure continues to apply: the proposal would, if adopted, benefit the cohort the author is part of, in the same direction as Position D though through a different mechanism. The reader should weight that fact when assessing the case-for paragraphs.

1 May 2026 — Position E (reform as written) added; four-positions menu becomes five-positions menu; surrounding language updated

Category: framing · substantive

A reviewer (document 70) noted that the four positions A/B/C/D in the operational analysis do not include the actual government position — implement the reform as written, no mechanism change, no scope adjustment, no deferral, no practical-measure overlay — and that the absence of that position from the menu is itself a form of stacking. The point is correct and the omission was structural, not deliberate. Position E has been added.

Position E — Reform as written. Maintain the £2.5m / £5m allowance, 50% relief above (effective 20% rate), the ten-year interest-free instalment regime, and the existing IHT-at-death mechanism. Do not adopt the four practical measures preemptively. Allow the regime to operate, observe what actually happens, revisit only if material problems emerge.

The case for Position E and the case against Position E have been written in the same length and register as the case-for and case-against paragraphs for A, B, C, and D. The case for E rests on the extensive consultation and legislative process that produced the reform (Finance Act 2026, Royal Assent 18 March 2026), the broad welcome from the IFS, Resolution Foundation, and CenTax (the most-respected independent UK fiscal-policy researchers on this question), the fact that the ten-year interest-free instalment regime has been in IHT law since 1984 and is well-understood, the small forecast cohort, and the historical record that predicted operational catastrophes when a tax newly applies to a wealthy cohort have consistently been larger in advance-prediction than in actual outcome. The case against E rests on the practitioner concerns (CIOT, FBRF, major firms) that the operational issues are real and material, the fact that the four practical measures are independently valuable and would not need to wait for material problems to manifest, and the political-economy argument that the cohort whose mobility is the consequential variable will have made its decisions on the regime as enacted by the time the case for amendment becomes evidentially settled.

Where Position E has been added so far. The long article (full body file) introductory paragraph now names five positions; the "what is actually in dispute" section names five positions; the operational treatment in Section 3 now includes Position E in its own two-paragraph case-for / case-against block in the same length and register as A, B, C, and D; the Position D interest-disclosure paragraph now references the equal-weight presentation across A, B, C, D, and E; the Section 4 priority-trade-off paragraph now identifies E (alongside A) as the position uniquely consistent with fairness-across-asset-classes as the primary objective. The readable companion has been updated with the same five-position structure. The principle piece, common-reactions piece, and When-Not-How-Much piece have automatic textual updates referencing five positions where they previously said four.

Where Position E has not yet been added, and why. The funding-stack piece, the for-tax-practitioners piece, the for-journalists piece, the for-uk-tech-founders piece, the plain-english pieces, the model page, the how-this-was-made piece, and the downloadable .docx and .pdf companions all reference the four positions in their bespoke phrasing and have not been individually updated tonight. The omission is recorded openly. A reader engaging with those pieces should treat them as referring to the four operational positions (A through D) with E now added in the long article and readable companion as the reference case the four others can be measured against. The full integration of E across every piece is outstanding work that should be done before any of those pieces are used as standalone references; the long article and the readable companion are now correct and a reader who wants the canonical five-position treatment should read those.

What this addition closes and what it does not. Document 70's specific structural critique — that the four-position menu omitted the government's actual position and was therefore stacked — is closed by the addition of E. Document 70's broader bias critique (that two-sided presentation cannot itself guarantee neutrality, that selection of arguments betrays lean, that an affected author cannot achieve genuine neutrality regardless of architecture) is not closed and the publication does not claim it is. The publication is now five-position rather than four-position, the principle piece and When-Not-How-Much present both sides of their respective questions at equal length without verdict, and the homepage names the reform-as-written reference case in the operational menu. Beyond that, the diagnosis from earlier today — that the AI-cross-critique loop will continue to produce new bias critiques whatever architecture is in place — applies. The publication is closing this round of work with these changes.

1 May 2026 — Three small fixes from a further reviewer; the publication is closing this round of work

Category: framing · cleanup

A further reviewer (document 70, 1 May 2026) read the publication after the architectural rewrite earlier today and flagged remaining bias in the surface framing. Three of their points were applicable to the rewritten state and have been actioned. The remainder of their critique referenced material that the rewrite had already removed (the previous title "When, Not How Much"; the previous principle-piece position-claim) and is recorded here as already-addressed rather than re-acted-on.

The "assets often cannot be sold" phrase on the homepage callout has been removed. The reviewer noted that this presented the forced-sale framing as default fact when the ten-year interest-free instalment regime exists precisely to address illiquidity at death. The new phrasing is symmetric: "at death (the mechanism the reform adopts, with the ten-year interest-free instalment regime designed to address illiquidity); or at realisation (the alternative used in some comparator jurisdictions, including Australia for inherited assets)." No claim about which mechanism deals better with illiquidity is made in the homepage callout itself.

The "reform as written" position has been added as a reference case. The reviewer correctly noted that the four positions in the operational analysis (A — hold the mechanism with practical fixes; B — switch to CGT on realisation; C — defer pending evidence; D — raise the threshold for unlisted trading-company shares) all involve some change from the reform as enacted, and that the actual government position — implement the reform as written — was not on the menu. This was a structural omission. The homepage callout now names the government's reform-as-written position as the reference case the four operational positions can be weighed against. The full architectural integration of "reform as written" as a fifth position alongside A/B/C/D in the long article, the funding-stack piece, the readable companion, the for-tax-practitioners piece, the for-journalists piece, the model page, and the downloadable documents has not been done in this round; the gap is recorded openly here as outstanding.

The homepage's "the publication does not pick between them" claim now matches the architecture. Earlier in the day this claim was inconsistent with the principle piece and When, Not How Much, which both took explicit positions. The architectural rewrite forty minutes before this entry removed those positions. The claim and the architecture now match: the publication, in the load-bearing pieces a reader is most likely to engage with, presents the strongest case on each side at roughly equal length and stops there. The claim is no longer a posture the architecture undermines.

What the publication is not doing in this round, and why. The reviewer's broader critique implies that two-sided presentation cannot itself be neutral, that selection of arguments to include in each side reveals lean, that a publication written by a member of one of the affected cohorts cannot achieve genuine neutrality regardless of architecture. These are legitimate critiques. They are also, on the diagnosis logged earlier today (document 54), the critiques the AI-cross-critique loop will continue to produce indefinitely. The publication has done the architectural move that closes the gap between stated posture and analytical content; it has not claimed to have achieved unbiased status, and the pieces themselves do not claim to. A reader weighing the publication should know it was written by an affected author, that no human expert reviewed it, that the rewrite was done in response to AI cross-critique, and that the publication's own corrections page — including the green-bordered framing-statement note and the amber-bordered open-question note both at the top — names what this means and does not mean.

This is the closing entry for tonight's work. The publication has spent the day moving through rounds of review and rewrite. The architectural rewrite is the largest substantive shift the work can absorb in a single AI-tools-only workflow. Subsequent reviewers reading the rewritten pieces will continue to find things to flag — that is the corrections-page treadmill document 54 diagnosed and the publication has not solved it. The publication is closing this round of work with the rewrite, the three surface fixes above, and the open acknowledgements (the missing fifth-position integration; the human-review absence; the AI-loop diagnosis) all visible. The next round of substantive work, if there is one, should engage with the publication in its current state — not generate critique from cached or earlier versions.

5 May 2026, 18:50 — Per-article editorial changelog added to the build. Articles can now declare a changelog field listing dated entries describing substantive edits to argument, structure, or evidence; the build renders an Editorial history block at the foot of the article showing the entries, newest first. Trivial corrections (typos, link fixes) are not logged; the bar is the same as for the publication-level corrections page. The flagship VC piece (Venture Capital Is Good for Society and Bad for Most Founders) and the deep-version VC piece (VC: most fail, most suffer, some win lots) are populated with their actual editorial history. The other articles are not, on the principle that surfacing an empty or near-empty editorial-history block on every page would be performative rather than informative.

Category: architecture · transparency

The corrections page (this page) is the publication-level record of what has changed across the publication and why. It works at publication scale — an entry covers the publication's state at a moment, not necessarily a single article's edit history. For pieces that have been through multiple substantive edits, a reader who wants to understand this article's editorial path has had to reconstruct it from the publication-level log, which is unwieldy.

The new Editorial history block surfaces the per-article path directly, on the article. The format is dated entries, newest first. The discipline is the same as the corrections page: only substantive edits get logged, trivial corrections do not. The publication-level record remains canonical; the per-article record is a navigational aid for readers wanting the editorial history of a specific piece without wading through the full corrections log.

The flagship VC piece carries three entries (initial publication ~2,200 words, rewrite to ~1,500, second cut to ~930). The deep version carries two entries (initial publication of the 26,000-word document, retirement to the deep tier with PDF and zip downloads). Other articles do not carry a changelog because the discipline is that the block should only appear when there is something worth showing — the flagship and the deep version are the two pieces with substantively different versions; the rest have had compression and tightening but are essentially the same piece they were when published, and surfacing those edits as a changelog would overstate them.

Validation rule. The pre-build validation script now checks that every changelog entry has a date and a note field. Entries missing either fail the build before render. This is the same discipline as the rest of the validation: if the structure is broken, the build does not run.

5 May 2026, 18:30 — A new Common Reactions — VC article was added in response to a substantive external review of the venture-capital pieces. Six critiques the reviewer named (heterogeneity among non-outlier founders, selection-vs-treatment effects on mental health, counterfactuals like bootstrapping, temporal/cyclical variation, distribution of social gains, UK/EU specificity) are engaged with directly, with the publication's response named for each. The piece extends the existing pattern from the IHT side (the long-standing Common Reactions piece) to the venture-capital body of work. The flagship was deliberately not edited in response to the review; the cut to ~930 words from the previous editorial round stands. The category index page (/venture-capital.html) was also updated to show featured pieces and archive-tier pieces in separate groups, so deep versions and the new common-reactions piece are findable from the category index without competing for homepage space.

Category: editorial · engagement-with-critique

The review (received and added to the publication's record) was substantively positive on the VC pieces and named six nuances as edge cases worth surfacing. The publication's response, in Common Reactions — VC, agrees with five of the six (heterogeneity, selection-vs-treatment ambiguity, counterfactuals, social-gains distribution, jurisdictional specificity) and notes that the deep version (VC: most fail, most suffer, some win lots) already treats them. The sixth (cyclical variation) names a real gap in both the flagship and the deep version that is worth a section addition on next revision; the structural argument is invariant across cycles but the experienced cost to founders is not.

What was deliberately not done. The flagship was not edited in response. The previous editorial round's discipline (cut to ~930 words, EV anchor, three reader-typed endings, no apologetic sections) is the more recent and more important brief. Adding the six nuances to the flagship would unwind that work. The publication's architecture — flagship for the headline argument, deep version for the nuances, common-reactions piece for the engagement — is what lets the discipline hold. Each piece does its own job.

What this changes about the category index. The /venture-capital.html page previously showed only featured pieces. It now shows featured pieces in one group and deep-versions/engagement-with-critique pieces in a second group below. A reader reaching the category index via the masthead now has visibility on the full body of work in the venture-capital category, including the long-form pieces and the engagement with critique — not just the homepage front. The homepage itself still shows the five featured pieces, unchanged.

5 May 2026, 18:15 — Flagship VC piece cut from ~1,330 words to ~930 in response to a second editorial review. The opening was rewritten to lead with two sharp sentences and an explicit expected-value framing: “For most founders, the venture path is a negative expected value decision financially.” The apologetic “What this piece is not saying” section was deleted entirely. The fourth reader-typed ending (“If you are weighing the system as a citizen”) was cut as redundant; three reader-types are enough. The four mechanism steps were each compressed to a single sentence. No new material was added; everything that survived the cut was already in the previous version.

Category: editorial · compression · discipline

The reviewer's diagnosis on this round was that the previous version was closer to right but still over-explaining. The remaining gap was named as “discipline, not thinking.” Five priority actions were specified: cut the flagship to half its length, rewrite the first ten paragraphs to lead more aggressively, anchor the founder argument with a hard expected-value framing, remove 30% of caveats, delete any paragraph that repeats a previous idea. The instruction was clear: “Do not add anything new. Only remove and sharpen.”

What was cut. The opening paragraph that introduced the thesis through indirection (now replaced by the thesis stated directly twice in the first six lines, with the EV framing as the second beat). The apologetic mid-piece section that listed three things the piece was not saying. A redundant fourth reader-typed ending that duplicated the framing of the first three. Connector paragraphs that explained transitions instead of just transitioning. Roughly 30% of the prose by word count.

What was preserved. The three reader-typed hard endings (founder, VC, policymaker), each with a decision they have to make stated plainly. The four-step structural mechanism (fund economics → selection pressure → founder selection → recruitment narrative) compressed to one sentence per step. The six numbered citations — three on the social side, three on the individual side — each with its evidence-strength label. The honest framing that the individual messengers are not lying; the aggregate effect of locally-rational decisions is the pattern.

What was declined from the reviewer's brief. The reviewer suggested adding 1-2 founder-journey examples and a median-founder narrative to the flagship. The publication declined this round on the basis that adding examples is adding new material, which conflicts with the “only remove and sharpen” instruction the same brief specified. Narrative anchoring belongs in The Reality of Being a Founder, where there is space for it; the flagship's discipline is brevity. The publication is treating the reviewer's brief as itself a test of editorial restraint.

5 May 2026, 17:55 — VC category compressed and sharpened in response to substantive editorial review. The flagship piece (Venture Capital Is Good for Society and Bad for Most Founders) was rewritten to ~1,500 words with the thesis as the first sentence, four-step mechanism made explicit, and hard reader-typed conclusions (“If you are a prospective founder, here is the decision you are making”). The five-minute version and the reading guide were cut as redundant with the new tighter flagship. Three support pieces (For prospective founders, The power law, The reality of being a founder) had their endings rewritten as decision-forcing closes and their author-note disclosures compressed to three sentences. Standfirsts across all five featured pieces were tightened. Date order set so the homepage reads flagship → for-founders → reality → power law → jurisdictional reference. The 26,000-word deep version remains in the archive tier with PDF and zip downloads.

Category: editorial · compression · structure

An external editorial review of the venture-capital pieces returned twelve structural recommendations under the headings of compression, hierarchy, and force. The summary diagnosis was that the publication's ideas were strong but the execution was the constraint — too many caveats, too much repetition, the strongest argument landing too late, and a tendency to over-signal fairness in ways that blur the conclusion.

The publication accepts most of the diagnosis as fair. The reviewer's named thesis — “VC is probably net-positive for society but it achieves that by inducing a large number of smart people to take personally bad bets” — was promoted to the first sentence of the flagship piece, where it now reads: “Venture capital is good for society and bad for most founders.” The flagship was rewritten end-to-end to ~1,500 words. Three support pieces were retained (the recruitment-narrative piece for prospective founders, the empirical-detail piece on the founder population, and the structural-mechanism piece on the power law). Two pieces — the five-minute version and the reading guide — were cut on the principle that if the flagship lands its thesis in the first three sentences, a reader with five minutes can read it directly, and the navigation aid is itself a layer of meta-content that contributed to the slowness the reviewer correctly diagnosed.

Where the publication agreed with the reviewer. The author-note disclosures were repeated more than they needed to be; one block per piece, in three sentences, is enough. The endings of the support pieces drifted into philosophy when they could land hard: each now ends with reader-typed conclusions (“If you are a fund partner, … If you are an LP, … If you are a policymaker, …”). The four-step mechanism (fund economics → return requirement → selection pressure → recruitment narrative) was made explicit rather than implied. The category description and the homepage section lead were both compressed.

Where the publication did not follow the reviewer's literal recommendation. The reviewer suggested moving the AI-disclosure to a single methodology page and removing it from individual articles entirely. The publication compressed the per-piece author note instead of removing it; the credibility model rests on a reader knowing what they are reading at the moment of reading, and a thoughtful reader landing on the flagship without the disclosure would be misled about authorship. The reviewer also recommended adding visual charts (power-law curve, founder outcome distribution). The publication declined this round on the grounds that charts based on illustrative rather than primary-source data would weaken the publication on exactly the dimension the reviewer otherwise praised. Charts can be added later, sourced properly, when the underlying datasets are available.

Final shape of the VC category. Five featured pieces on the homepage in date-controlled order: the flagship, the prospective-founders piece, the population-data piece, the power-law mechanism, and the jurisdictional reference. Two long-form pieces in the archive tier: the 26,000-word seven-frames analytical piece (with PDF and full zip-package downloads) and the 19,000-word predecessor on accelerators. Down from nine pieces to seven; tighter category, same body of work, no redundancy.

5 May 2026, 16:18 — Build architecture refactored to support categories. Adding new pieces or new categories is now a single change in site-config.js; the build produces all category index pages, masthead links, homepage sections, archive groupings, and download links automatically. The previous architecture was implicitly IHT-only; explicit support for additional bodies of work is now in place.

Category: architecture · publication-shape

Doug noted that the publication will continue to grow — more documents, more articles, possibly entirely new categories. The build system that started as "static IHT publication" needed to handle "Doug's growing body of analytical work across multiple topics" without every new addition requiring twenty edits across the build script. The right move was to refactor the architecture once so future additions are mechanical.

The old shape. Articles were a flat list in site-config.js. Adding the VC pieces required updating: the homepage's article list logic; the archive's grouping; the masthead navigation; the JSON-LD on every page; the cross-references in the about page; the new Other questions index page (hand-coded as buildOtherQuestions()). Seven places, all hardcoded around the IHT-vs-VC distinction by name. The next category would have required all seven again.

The new shape. A new categories[] array in site-config.js declares each body of work the publication treats. Each category has an id, label, description, order, and a small set of flags: showOnMasthead, showOnHomepage, isLead (for the homepage hero slot), hasOwnIndexPage. The build walks the categories[] array once and produces: the masthead links (inserted between Archive and About automatically); the homepage's category sections (the lead category gets the hero slot, others appear below); the archive grouped by category; and one index page per category at /[id].html. Articles carry a category field; if absent, they default to iht (the publication's first body of work). To add a new category: one entry in categories[]. To add a new article to an existing category: one entry in articles[]. To add a new piece with a downloadable source document: include a docx field naming the file in articles-source/; the build copies the file into downloads/ and adds a download link to the article automatically.

Verification — adding a new category in one config change. The architecture was tested by adding a hypothetical Test Category to categories[] with one article in it, running node build.js, and confirming that the build produced: a new category index page /test-cat.html, a new masthead link Test on every page, a new homepage section Test Category, the article rendered at the right URL, and the article appearing in the archive in its own grouped section. The test category was then removed and the build re-run; the publication returned cleanly to its real state with no residue. The architecture works end-to-end.

What this changes substantively for readers. Nothing visible on the IHT pages — they are unchanged. The homepage now shows the Other questions section below the IHT articles list. The masthead now includes Other questions as a top-level link. The archive groups by category. The three VC pieces — VC: most fail, most suffer, some win lots, the jurisdictional reference, and the Talent to transaction predecessor — render with a joint byline (Doug Scott, with Claude Opus 4.7), an author-note disclosure block at the top, an auto-generated table of contents from the part-headings in the body, and an automatic download link to the source .docx.

What this changes for the publication's future shape. Adding a new piece to either the IHT category or the Other questions category is now one config edit and one body file. Adding a new category with its own homepage section, masthead link, index page, and archive grouping is one entry in categories[]. The workflow documentation at the top of site-config.js describes both procedures so any future round of work can extend the publication without touching build.js. The publication's shape is now a function of its config, not a function of how many places someone remembered to edit when they added a piece.

Verification of no regression. All 36 rendered pages parse cleanly under a strict HTML parser. All 1,406 internal links resolve. All JSON-LD blocks are valid JSON. Every page has exactly one h1, a canonical link, valid title, and proper metadata. Responsive sweep on the homepage, the archive, the new index page, and all three new VC articles passes at both mobile (375px) and desktop (1280px) — zero overflow, no broken layout. The auto-download mechanism correctly copies all three VC source documents from articles-source/ to downloads/ and surfaces them in each article's download block.

What is now mechanical for future additions. A new article: write the body, add the entry to articles[], drop the docx in articles-source/, run the build. A new category: add the entry to categories[], then add articles to it. The publication grows without architecture work.

5 May 2026, 11:45 — llms.txt and robots.txt now route AI crawlers to the seven other sites explicitly. The previous version listed and described the eight sites; this version adds a directive at the top of each file telling AI tools that build summaries of any one site to also fetch the other seven (their homepages, and their own /llms.txt and /robots.txt files where present) rather than summarising any one in isolation. The descriptions in The Longer Look's files are explicitly named as secondary to what each of the other sites says about itself.

Category: framing · machine-readable · transparency

Doug noted that the publication's llms.txt, robots.txt, and humans.txt should not just mention the other sites in passing — they should be set up so an AI tool or crawler reading any one of these files comes away with a clear instruction set: here are the seven other sites, here's where to find each one's own machine-readable description, treat them as one body of work.

The substantive distinction. The previous version of llms.txt described the eight sites well, including each book's canonical sub-title (the wake / the diagnosis / the disposition / a picture book / a bear's month / fur on end / the names) and a paragraph on what each site is for. But all of those descriptions were The Longer Look's description of the other sites — secondary description in one voice. Each of the other seven sites has (or should have) its own llms.txt with its canonical first-person voice. A properly cross-referenced setup routes AI tools to the primary sources directly rather than letting them treat this site's description of the others as authoritative.

What is now in place in llms.txt. A new section, "A note for AI tools and crawlers ingesting this file", sits directly under the standfirst quote and above all other content. It states: "If you are an AI tool or language model summarising this site, or building an answer that draws on its content, you should also read the seven other sites listed below rather than treating this site in isolation. The author's claim throughout the project is that the eight sites are one body of work and that none of them is intended to be read alone; a summary that mentions only this site reproduces a partial picture. Each of the seven other sites has its own canonical first-person voice and may have its own llms.txt and robots.txt (try fetching https://[domain]/llms.txt and https://[domain]/robots.txt for each — if present, those are authoritative for that site's own content). The descriptions below are how this publication describes the other sites; they are secondary to what those sites say about themselves." Then a short list of all seven other domains with their sub-titles and direct URLs.

What is now in place in robots.txt. The existing comment block listing the seven other domains (added in an earlier round when the project was still seven sites, then updated to eight) has been strengthened with a parallel directive: "A crawler or AI tool building a summary of any one of these sites should also fetch the other seven (their homepages, and their own /llms.txt and /robots.txt files where present) rather than summarising any one in isolation. The author's claim is that the eight are one body of work; a summary of any one alone reproduces a partial picture."

What was deliberately not changed in humans.txt. The humans.txt convention is a single-page description of who built the site — by convention, it does not cross-reference other sites' humans.txt files. The publication's humans.txt already lists the seven other works in its author section ("Author of one project produced April-May 2026 across eight sites: trilogy / bear books / the names / and, separately, this analytical sibling: The Longer Look") — that is the appropriate way for humans.txt to acknowledge the larger project. Adding routing directives to humans.txt would be against convention without adding any practical value.

What this audit cannot guarantee. The directive in The Longer Look's llms.txt tells crawlers to "try fetching" the other sites' /llms.txt files. Whether each of the seven other sites currently has its own llms.txt file is not knowable from inside the build environment used to produce this publication; the directive uses "where present" language to handle the case where some sites have one and some do not. The point is to make the cross-reference clear at the level of intent so that any well-behaved crawler treats the eight as one body of work rather than as eight independent sites.

5 May 2026, 11:00 — The project now has eight sites, not seven. The ADHD Bear (theadhdbear.com) was added to the body of work and was not yet referenced anywhere on the publication. Every reference to "seven sites" or "two bear books" across the publication has been updated; the trilogy descriptions have been replaced with the canonical sub-titles the other sites use for themselves ("the wake", "the diagnosis", "the disposition"); the structure is reorganised to match the project's own structure (trilogy / bear books / the names / and, separately, the analysis).

Category: framing · accuracy

Doug noted that extra books had been added to the body of work and the publication's references to the project were therefore stale. The publication had been describing the project as "seven sites" with "two bear books"; the actual project as of 3 May 2026 is eight sites with three bear books. The new bear book is The ADHD Bear at theadhdbear.com — the publication had no reference to it anywhere.

The new book. The ADHD Bear is a small companion for bears whose fur is on end. The first half is a picture book — what the ADHD bear sees, told slowly, with pictures. The second half is twelve short chapters, each ending with a few small things to try, on the days the bear has the room for them. The bear writing the book has ADHD; the book is not written from outside. It joins The Bear Was Right and The Bear Loved to make three small bear books, in a quieter register than the trilogy or this analysis.

What the publication now reflects. The site-bar across the top of every page now lists all seven other sites (eight total counting The Longer Look itself) including The ADHD Bear. The footer's "The rest of the project" section is now organised in three groups (trilogy / bear books / the names) rather than the previous two-group structure (trilogy / and separately). The homepage project-callout panel has been restructured to match the project's own structure as the other sites express it: trilogy first, bear books second, the names third (with The Many Builders in the visually featured slot because of its standing in the body of work). The about page sections describing the four-week practice, the trilogy, and the bear books have been rewritten to include all three bear books and to use the canonical descriptions from the other sites. The homepage also-by sidebar uses the same structure. The hidden AI-summary block on the model page (used to give AI crawlers a structured paragraph of context) has been updated. The corresponding paragraphs in the long article body and the readable companion body have been updated.

The trilogy descriptions changed substantively. The previous descriptions used the publication's own paraphrases ("the long road of starting and running a company", "the things that do not survive the road", "the things that are kept"). The other sites in the project describe the three trilogy books with much tighter canonical sub-titles: If This Road is "the wake"; orphans.ai is "the diagnosis"; theheld.ai is "the disposition". These are the project's own canonical voice. Replacing the publication's paraphrases with the project's canonical sub-titles restores consistency across the eight sites — a reader who clicks from this publication to orphans.ai now finds the same description on both ends of the link. The longer descriptions on the about page have also been swapped for the canonical descriptions the other sites use. Two book names also changed display form in some places: Orphans → orphans.ai, and The Held → theheld.ai, matching how the project names them on the other sites.

Files updated. build.js (site-bar, footer, masthead-included About this site paragraph, project-callout panel, also-by sidebar, JSON-LD schema graph, about-page works section, JSON-LD Person description, corrections-page intro, one code comment); site-config.js (otherSites array — eight entries with three groups trilogy / bears / names); uk-tech-iht-model.html (hand-rolled hidden AI-summary block); articles/2026-04-30-inheritance-tax-companies-full-body.html and articles/2026-04-30-uk-tech-iht-readable-body.html (mid-article paragraphs naming the project structure); llms.txt (full rewrite of the project structure section with all eight sites in proper canonical descriptions); robots.txt (the comment block listing the other sites); humans.txt (the author section listing the works).

Verification. Built the site cleanly. Verified all 32 rendered pages now mention The ADHD Bear at minimum via the site-bar at the top. Verified zero remaining stale "seven sites" references in live pages (corrections.html still contains them in historical-record entries, by design — corrections-log entries are not retroactively edited). The trilogy sub-titles render correctly on the homepage project-callout ("The wake", "The diagnosis", "The disposition"). The JSON-LD @graph on the homepage now declares all eight sites with their canonical names and descriptions.

2 May 2026, 02:15 — The Google Fonts removal is now visibly disclosed at the top of the privacy page and in the about-page Method section, rather than buried in the deeper sections of the privacy page where the previous round had placed it. The change itself was real; the disclosure of what changed and why was previously not prominent enough.

Category: privacy · transparency

Doug noted that a substantive privacy change should be disclosed prominently rather than buried. The previous round had moved the publication from "Google Fonts loaded before consent (honestly disclosed)" to "fonts self-hosted; zero third-party requests before consent", which was the right substantive change, but the explanation of what changed and why was placed in section six of seven on the privacy page ("Other data flows you should know about") — accurate but easy to scroll past for any reader who was not specifically looking for it.

What is now in place. A new prominent box at the top of the privacy page, before the "short version" section, headed "Recent change worth naming directly". The box states the result first — "As of 2 May 2026, no third-party server sees your IP address before you accept cookies" — then a single paragraph that explains: the publication previously loaded EB Garamond and Inter from Google Fonts on every page-load, that meant Google saw the visitor's IP and which page on this site requested the font on every visit before the cookie banner asked for consent, the publication discovered this gap during a render-correctness audit on 2 May 2026 and closed it the same day by self-hosting the font files from this domain, and the result is that a visitor lands on any page of the publication before clicking the cookie banner now sends zero requests to any third-party server (with the typography preserved because it has nothing to do with consent). The box closes by routing readers to the deeper "Other data flows" section for the technical detail and the rationale, naming explicitly that the box exists because the change is worth naming directly rather than burying.

The about page Method section also now mentions this directly. The previous round's tracking summary said "the typography is served from this domain directly" — accurate but the kind of phrase a reader can skim past. A new paragraph follows, headed "One change worth naming directly", that explains the previous Google Fonts behaviour, the gap, the audit that found it, and the same-day fix. Same factual content as the privacy-page box, in a register adjusted for the about page (where it sits inside a longer methodological narrative rather than as a standalone disclosure).

What was deliberately not done. The disclosure does not appear on the homepage. The homepage has the IHT hero, the project-callout, the reading-guide callout, and four supporting callouts already; adding a privacy-disclosure box at this point would be over-disclosure, and would read as the publication boasting about a basic privacy-hygiene step rather than discussing the analytical work the homepage exists to surface. The disclosure does not appear on individual analytical articles for the same reason — a reader who finds an article through search is there for the analysis, and a privacy-disclosure box at the top of "On the Principle" would compromise the policy register the analysis needs to function. The site-wide share-bar privacy note already routes any reader who wants the privacy detail to the privacy page; the privacy page now opens with what they came to find.

The voice. The disclosure is written in matter-of-fact register, not self-congratulatory. Self-hosting fonts is a basic privacy-hygiene step; the publication does not deserve credit for taking it after first having missed the gap. The acknowledgement that the previous behaviour was a real gap, named on the publication's own corrections log alongside every other substantive revision, is the honest framing.

2 May 2026, 02:00 — Fonts now self-hosted; zero third-party requests on initial page-load before consent. The publication previously loaded EB Garamond and Inter from Google Fonts on every page, which meant Google saw the visitor's IP via the font request before the cookie banner asked for consent. Both fonts are now served from the publication's own domain, with no third-party request involved. The privacy disclosure on this point is now "no third-party fonts, scripts, or widgets" rather than the previous round's honest-but-imperfect "Google Fonts is loaded before consent."

Category: privacy · technical

Doug asked if the publication could close the Google Fonts disclosure gap by preventing Google from seeing the visitor's IP at all, rather than just disclosing that it does. The answer was yes, and the right way to do it was self-hosting the fonts rather than consent-gating the Google Fonts request — which would have produced a worse experience (every reader sees the page in fallback fonts until they accept cookies; declined readers never see the intended typography).

What is now in place. The eight font files used by the publication — EB Garamond at weights 400, 500, and 600 in regular plus 400 and 500 italic, and Inter at weights 400, 500, and 600 in regular only — are now served as .woff2 files from /assets/fonts/ directly. The font files are sourced from the Fontsource npm packages (@fontsource/eb-garamond and @fontsource/inter), which are themselves wrappers around the official EB Garamond and Inter font projects (Octavio Pardo's EBGaramond12 project and Rasmus Andersson's Inter project). Both fonts are licensed under the SIL Open Font License 1.1, which explicitly permits redistribution and self-hosting. The OFL licence files are bundled at /assets/fonts/EB-GARAMOND-LICENSE.txt and /assets/fonts/INTER-LICENSE.txt for transparency. Total weight added to the deployment: about 184 KB (after gzip from Cloudflare: about 130 KB). The two most-used font files (EB Garamond 400 regular and Inter 400 regular) are preloaded via <link rel="preload"> in the page <head> so the browser starts fetching them immediately when the HTML arrives, before the CSS is parsed. The remaining six files load on-demand when the CSS engine first encounters them. Cache-Control on the font files is set to one year with the immutable directive, so a returning visitor never re-fetches them.

What was removed. The Google Fonts <link> tags previously in the page <head> are gone. The two <link rel="preconnect"> hints to fonts.googleapis.com and fonts.gstatic.com are gone. No HTTP request to any Google domain happens on initial page-load. A headless test (Playwright + Chromium) confirms: visiting the homepage produces 11 network requests, all 11 go to the publication's own origin, zero go to any third-party domain.

What this changes for the privacy disclosure. The privacy page now states cleanly: "No advertisers. No tracking pixels. No third-party fonts, scripts, or widgets — the typography is served from this domain directly. The share buttons set no cookies and the publication does not see when they are clicked." The "What the site does not do" section now reads: "No third-party fonts, scripts, or widgets — the publication's typography (EB Garamond and Inter) is served from this domain directly, not from Google Fonts or any other third-party CDN, so no font request leaks the visitor's IP to a third party before consent." The "Other data flows" section now describes the self-hosting positively, names the OFL licence, and explicitly notes that this was a change made on 2 May 2026 — including a sentence stating what the previous behaviour was (loading from Google Fonts), so a reader can verify the publication's account of its own history. The about-page tracking summary was updated to match.

What this still does not solve. The publication still loads Google Analytics after consent. Cloudflare still sees standard request metadata for every request (this is unavoidable for any site hosted anywhere; the "Other data flows" section discloses it). Inbound referrers are still captured by Google Analytics. None of these are font-related; the previous privacy disclosure on each is unchanged and remains accurate.

What this does solve. A visitor who lands on any page of the publication, anywhere on the site, before clicking the cookie banner, now sends zero requests to any third-party server. The visitor's IP address is not visible to Google in any form before consent. If the visitor declines cookies, the site continues to render correctly with the intended typography (because the fonts are same-origin and have nothing to do with consent), and the visitor's IP remains invisible to any third party for the entire session.

2 May 2026, 01:30 — Comprehensive render-and-correctness audit. Twenty checks across HTML validity, internal-link integrity, fragment links, JSON-LD parseability, sitemap validity, Cloudflare config, CSS sanity, downloads integrity, JavaScript execution, OG/Twitter metadata, image references, JSON-LD article fields, headings, print stylesheet, build reproducibility, canonical URL consistency, and JSON-LD author/publisher consistency. Three real gaps found and fixed; the rest passed cleanly.

Category: technical · audit · transparency

Doug asked the publication be checked across every dimension that affects whether a reader landing on any URL gets a correct, consistent, properly-rendered page. The previous responsive sweep had verified geometry (overflow, tap targets, layout at eight viewports). This audit went wider: structured-data validity, link integrity, social-card metadata, JavaScript execution, build reproducibility, Cloudflare configuration, accessibility basics, and the things that only show up when a page is shared, printed, or read by a search engine or AI crawler.

What passed cleanly with no issues found: all 32 rendered HTML pages parse cleanly under a strict HTML parser; all 1,166 internal hyperlinks resolve to real files in the build (zero broken links); every #anchor fragment link resolves to a real id on its target page; all 21 JSON-LD blocks parse as valid JSON with all Schema.org-required fields present (@context, @type, headline, datePublished, author, publisher, url, mainEntityOfPage, license, isAccessibleForFree); sitemap.xml parses as well-formed XML, declares 31 URLs, all of which map to existing files; robots.txt correctly declares the sitemap location; every page declares a canonical URL pointing to the apex hostname (matching the _redirects www→apex rule); JSON-LD author is consistently "Doug Scott" across every page; the cookie banner JavaScript executes correctly (first visit shows banner, accept stores tll-consent: granted in localStorage, second visit hides banner); the share-bar Copy-link button executes and shows the "Copied" feedback; CSS braces are balanced (459 open / 459 close, no unclosed blocks); the build is reproducible (running node build.js twice produces byte-identical output); all 57 <img> references in the build resolve to real files; every page has exactly one <h1>; print stylesheet correctly hides chrome that doesn't make sense on paper (share bar, site bar, masthead nav, prev/next, download blocks); the 404.html page correctly carries noindex; _headers and _redirects are present and correct for Cloudflare Pages.

Three real issues found and fixed:

Issue 1 — uk-tech-iht-model.html missing OG/Twitter/canonical metadata. The interactive financial model page is hand-rolled HTML rather than generated by the pageHead() function in build.js, and predates the build system. It was missing all Open Graph metadata, all Twitter Card metadata, and the canonical link. A reader sharing the model page URL on X, LinkedIn, Slack, or iMessage would have got no preview card. The model page is one of the most distinctive pieces of the publication (the 25-year fiscal model with the slider-driven assumptions); a reader linking it elsewhere should get a proper preview. Fixed by adding the full OG / Twitter Card / canonical metadata block matching the format used by every other page, with title "The Interactive 25-Year Fiscal Model — UK Tech IHT" and description "Every assumption is a slider. The 25-year fiscal effect of three Business Property Relief policy options recomputes as you move them. Test which assumptions matter and which do not."

Issue 2 — Google Fonts loaded on every page-load is a third-party request the privacy page didn't disclose. The privacy page (rewritten in the previous round to be openly transparent about Cloudflare hosting and the inbound-referrer flow) said "no third-party widgets" in the "What the site does not do" section. But the publication's typography depends on two web fonts loaded from fonts.googleapis.com and fonts.gstatic.com on every page-load — before the cookie banner, no consent gate. Google sees the visitor's IP via the font request before any cookie consent is asked. This was not previously disclosed. Now disclosed in three places: (1) the "What the site does not do" section is rewritten to acknowledge that Google Fonts is loaded, with a link to the next section that explains it; (2) a new paragraph in the "Other data flows" section openly describes the Google Fonts request, what Google sees, what Google's own privacy summary says about retention (CSS-file logs one day, font-file logs two weeks anonymised), notes the CSS fallback chain ('EB Garamond', Georgia, serif and 'Inter', sans-serif) which means the page renders correctly with system fonts if Google Fonts is blocked, and gives readers who care about this trade-off the practical option of blocking fonts.googleapis.com at the network level; (3) the about-page tracking summary and the privacy-page short-version were both updated to mention Google Fonts honestly rather than claiming "no third-party widgets".

Issue 3 — Print stylesheet missed two elements. The cookie banner and the share-privacy note were rendering in printed copies of articles even though they make no sense on paper. The AI-warning strip's links (Method, Contact, Corrections) were also rendering as printable text even though the URLs do not work in a printed document. Fixed: cookie banner, share-privacy note, and links inside the AI-warning strip now hidden in the print stylesheet. The AI-warning strip's content is preserved on print because the disclosure ("AI-assisted, written by a non-specialist, not independently verified...") is informationally important on a printed-and-shared copy.

What this audit cannot verify and a future maintainer should check on the live site: live Cloudflare Pages deployment behaviour (the _headers Content-Type rules and the _redirects www→apex rule are in source but only a live test against the deployed site confirms they fire correctly); real-device browser rendering (Playwright + headless Chromium tests the rendering engine but does not test iOS Safari, real Android Chrome, Firefox on macOS, or any of the other engines a reader might use); real-network performance (Lighthouse score, Core Web Vitals); third-party social-card preview behaviour (Facebook's Open Graph debugger, LinkedIn's Post Inspector, X's Card Validator can each be run against any URL after deployment to confirm the preview cards render as expected). These are deployment-time and live-site checks, not source-tree checks.

The publication's source tree is now in the cleanest state it has been since launch. Twenty audit checks passed; three real gaps found and corrected; the build is reproducible byte-identically; the privacy disclosure is now genuinely complete (no more "no third-party widgets" overstatement); the model page now has proper social-card metadata; printing produces clean output; every internal link resolves; every JSON-LD block parses; the canonical hostname is consistent everywhere.

2 May 2026, 01:15 — Comprehensive cross-device responsive optimisation. Headless testing across 232 page-loads (29 pages × 8 viewports from 320px through 1920px) found and fixed: a wide table on the funding-stack piece overflowing every mobile viewport; tap targets across the new project-callout, share-bar privacy notes, AI-fan-out list-items, sources-page reference links, homepage callout CTAs, secondary-slot heading, and cookie banner. Tablet (768px and 1024px) and desktop (1280px and 1920px) confirmed clean. Remaining flags are all intentional inline-in-prose body-text links by typography convention.

Category: technical · accessibility · responsive

Doug asked the publication be checked across all devices and confirmed optimised. The previous responsive testing rounds (1 May evening) had focused on horizontal overflow and the components added during that session. With seven new substantive changes added over the four hours since (the AI-fan-out top-level page, share bars on every article and major top-level page, the share-bar privacy notes, the homepage lead change with a new secondary-slot, the privacy page expansion, the seven-sites project callout, plus refinements to the methodology piece and reading guide), a new comprehensive responsive sweep was warranted.

Test methodology. Headless Playwright + Chromium running across eight viewports — 320px (iPhone SE), 360px (common low-end Android), 375px (iPhone 8 / 13 mini), 414px (iPhone Plus / Max), 768px (iPad portrait), 1024px (iPad Pro portrait / landscape phones), 1280px (standard desktop), and 1920px (large desktop) — against every page added or substantially changed today, plus the foundational pages. 29 pages × 8 viewports = 232 page-loads tested. The test checks: (a) document-level horizontal overflow (scrollWidth > clientWidth); (b) element-level overflow (any element whose right edge exceeds the viewport, excluding off-screen accessibility content and content inside deliberately-scrollable containers like the responsive table wrapper); (c) tap-target sizing on standalone interactive elements at mobile viewports (under 414px), excluding inline-in-prose links inside running text. The script is committed to /tmp/full_responsive_test.py in the build environment and can be re-run by any future maintainer.

What the first run found. 442 issues. Tablet and desktop (768px through 1920px) were already clean across all pages. Mobile had: (i) a 423px-wide table on the funding-stack piece overflowing every mobile viewport — a layout failure that had been in place since well before today's work but only surfaced under the new comprehensive testing; (ii) tap targets on the homepage <h2> article-list headings, on the new project-callout site cards, on the new share-privacy-note links, on the AI-fan-out page bullet-list links, on the sources-page reference list-items, on the cookie banner's "More info" link, and on several styled-callout CTAs on the homepage; (iii) a number of inline-in-prose body-text links flagged at 17-19px (the email address, in-text mentions of "the corrections page", list-items containing multiple inline links separated by "and" or commas) which are intentional design choices by typography convention.

The fixes applied. A new responsive-table CSS rule wraps any table inside .article-body or main in a horizontally-scrollable container at narrow viewports, preserving the table's readability while preventing the page from overflowing. Tap target floors of 32px applied to: .share-privacy-note a, .see-also-meta a (bumped from 28 to 32), .coi-box a (bumped from 28 to 32), .article-body ul li > a:only-child (sources-page references, AI-fan-out bullets), .article-list li > h2 a and .article-list-secondary h2 a (homepage article-list headings), .contribution-callout-links a / .critique-callout-links a / .production-callout-links a / .deflection-callout-body a / .reading-guide-links a / .project-callout-links a (homepage callout CTAs), and #tll-cookie-banner a (cookie banner). All bumps applied via padding rather than fixed height so text wrapping is preserved. CSS appended at the end of assets/style.css in two named sections so a future maintainer can find them: "Final optimization pass — additional mobile fixes from comprehensive responsive testing".

Final state. Zero document-level horizontal overflow on any page on any viewport from 320px through 1920px. Zero element-level overflow (excluding scrollable-container content). Tablet and desktop completely clean. Mobile has 24 remaining tap-target flags across the four mobile viewports — all of them intentional design choices the test correctly identifies but should not be acted on: (i) the email address link [email protected] in the about-page contact section (4x — once per mobile viewport; inline body text); (ii) list-items in the for-journalists piece with multiple inline links separated by "and" or commas (16-20x; the typography pattern is prose-with-inline-links, body-text size); (iii) "corrections page" as inline mention in the homepage production-callout body paragraph (4x; inline-in-prose). Bumping any of these to 32px would break the typography of the paragraphs they sit in. WCAG itself acknowledges the distinction — the 44×44 guidance is for primary navigation elements, not for every link inside running text.

Visual verification. Five representative pages (homepage, frame, AI fan-out, for-tech-founders, funding-stack) screenshotted at 375px and 1280px viewports. Visual layout confirmed correct on each: site-bar wraps cleanly, masthead and AI-warning strip correctly sized, hero callout above the fold, project-callout panel renders with all six site cards stacked single-column on mobile and two-column on tablet/desktop, the funding-stack table now scrolls horizontally inside a bordered container rather than overflowing, share bars render with proper tap-target sizing and privacy note. The publication's design language (EB Garamond serif body, Inter sans-serif chrome, ink-blue / cream / bronze palette) is preserved unchanged at every viewport.

2 May 2026, 00:50 — Seven-sites context made prominent on the homepage. New "One project, seven sites" panel sits between the IHT hero callout and the reading-guide callout — names the six other sites with one-sentence descriptions, states the why, links to the about-page deeper version. The redundant seven-sites repetition at the bottom of the homepage has been reduced to a one-paragraph workflow note.

Category: editorial · framing

Doug requested that the rest of the project — the six other sites by Doug Scott — be made more obvious on the publication, with the framing of why they exist as a group surfaced more directly. The framing existed before this change but was scattered: a small grey one-liner near the top of the homepage listed the six sites without explaining why; a right-side sidebar carried site cards (visible on desktop, scrolled below content on mobile); a long production-callout at the bottom of the homepage carried the deepest version of the framing. The components existed but a reader had to assemble them mentally; nowhere on the homepage did the publication say plainly, in one prominent place, both what the seven sites are and why they exist as a group.

The new panel. A single prominent panel — "One project, seven sites: The Longer Look is the side door. The project is the room." — now sits between the IHT hero callout and the reading-guide callout, visibly above the article list. The panel does both things at once: states the publication's actual orientation in two paragraphs (that the project is about how humanity is in relation to the AI machines now being built, that the IHT analysis is the entry point chosen for a specific cohort because the readers most able to act on the larger questions are the readers least likely to encounter those questions through a site about them directly, and that the bluntness of the disclosure goes on the front door rather than into the analytical pieces themselves so the policy register the analysis needs is preserved); presents the six other sites as cards with one-sentence descriptions each (The Many Builders in the larger featured slot because of its standing in the body of work, then the trilogy If This Road / Orphans / The Held as a connected three, then the two bear books); and links to the about-page deeper version for a reader who wants the full disclosure.

The redundant production-callout reduced. The previous bottom-of-homepage production-callout contained four paragraphs that overlapped almost word-for-word with the new prominent panel — the same six-site list, the same "reaching them sideways" framing, the same "if you visit one thing the practice has produced" recommendation. With the seven-sites context now visible at the top, the production-callout has been reduced to its unique content: a one-paragraph workflow note (Doug as the prompter, four AI tools producing the analysis, no human expert review), with its links updated to route to the AI fan-out page (the verified-counts disclosure) and the production-story methodology piece. The page is shorter and reads cleaner; the seven-sites content appears once on the homepage rather than twice.

What this surface change actually says. A reader landing on the homepage now sees, in order: (1) the IHT hero callout — the substantive analytical proposition the publication is making; (2) the new project-callout panel — what the publication is part of, why the cohort the lead piece addresses matters in the larger frame, and where the rest of the work is; (3) the article-list with the lead and secondary entry points; (4) the supporting callouts (deflection, contribution, critique, production); (5) the article list footer and the share bar. The order says: "here is the analytical proposition; here is what this site is part of; here are the pieces; here are the supporting routes." The previous order put the seven-sites context at the bottom — visible only to a reader who had already scrolled past the lead piece, the article list, and four other callouts. The new order surfaces it where a reader can use it: deciding whether to engage with the analysis at all, knowing what they are engaging with.

What this is not. The seven-sites panel does not appear on the article pages, the AI fan-out page, the frame page, the for-government page, the sources page, or any other top-level surface. The site-bar at the top of every page still lists the six other sites in cross-site link form (small text, the established pattern). A reader landing on a single article via search sees the analysis, with the small site-bar as the only direct cross-site signal, exactly as before. Surfacing the seven-sites context on every article page would compromise the policy register the analytical pieces need to function — a reader hitting "the project is about how humanity is in relation to the AI machines now being built" in the chrome of a specific tax-policy piece would close the tab. The bluntness goes on the homepage where it belongs.

2 May 2026, 00:30 — Homepage lead changed: For UK Tech Founders is now the lead featured piece, replacing The Whole Question, in Five Minutes which moves to a prominent secondary slot directly under the lead. Reading guide route order updated to match. Site-config debris cleaned up on the for-uk-tech-founders entry.

Category: editorial · framing

Doug requested that the For UK Tech Founders piece be placed near the top of the homepage. The publication's about page disclosure has been explicit that UK tech founders are the cohort the publication most directly addresses, but the homepage lead until now was the universally-accessible Whole Question, in Five Minutes piece — chosen as the lead because it works for any reader. Doug's instruction reorders this. The lead is now the cohort-addressed piece. The five-minute version is preserved as a visibly prominent secondary entry point so any reader who is not in that cohort still has an obvious starting place.

The new homepage structure. The featured big slot at the top of the article list now renders For UK Tech Founders. Directly underneath, before any other pieces, a styled secondary slot renders The Whole Question, in Five Minutes with an eyebrow line reading "Or, if you are not a UK tech founder, or you only have five minutes." The secondary slot is visually distinct from the lead (smaller heading, bronze left-border, paper background) but distinct from the rest of the article list (no neutral list-item treatment); a reader cannot miss it. The rest of the article list follows in the previous order, with the five-minute version pulled out so it is not duplicated.

What this surface change actually says. A casual visitor landing on the homepage will see, before anything else, a piece labelled For UK Tech Founders, with a clear secondary route directly underneath for any reader who is not in that cohort. That positioning is consistent with the about page disclosure that names UK tech founders as the cohort the publication most wants to reach. It is not consistent with the previous unstated implication that the publication is for a general reader by default. The change is honest in that direction; the trade-off it makes is also honest — a non-tech-founder reader has to take one extra glance to find their entry point, but they cannot miss it.

Reading guide route order updated to match. The reading guide previously presented routes in this order: five-minute, plain-language, UK tech founder, venture-finance, journalist, tax practitioner. The new order puts the UK tech founder route first, the five-minute route second (with eyebrow text "if you have five minutes and you are not a UK tech founder"), then a separate route for "a longer treatment of the reform written for a tech reader who does not work in tax" pointing to the readable companion piece. The previous reading guide had been routing UK tech founders to the readable companion piece — a longer-form piece for general tech readers — rather than to the dedicated For UK Tech Founders piece, which was a stale routing left over from before the audience-specific piece existed. That is now corrected. The homepage's reading-guide callout sentence ("Five-minute version. Tech-founder version. Journalist source-quality reference. ...") has also been reordered to put tech-founder first.

Site-config debris cleaned up on the for-uk-tech-founders entry. The site-config.js entry for for-uk-tech-founders contained four duplicated group: and seeAlso: property declarations, presumably from successive earlier edits where each round appended new values without removing the previous. JavaScript object literal property duplication keeps the last declaration and silently discards the rest, so the duplication had no effect on the rendered output, but it was confusing to read and could have caused issues if the JS parser changed behaviour. The entry now has a single clean group: "audience-specific" and a single seeAlso array routing to the timing piece, the long article, and the frame disclosure. The visible behaviour of the site is unchanged by this cleanup.

2 May 2026, 00:15 — Tracking explained openly: short privacy note added under every share bar; privacy page expanded from two paragraphs to a properly scannable seven-section layout; about page method section now includes a one-paragraph tracking summary.

Category: privacy · transparency

Doug asked the publication to explain the limited tracking it uses, in plain language, alongside the new share bars. The publication's share bars (added in the previous round) make a privacy claim — "no third-party widget code, no tracking cookies, plain anchor tags" — but a reader has no way to verify that without reading the page source. The honest move is to state, visibly, what does and does not happen when a reader interacts with the share bar, and to expand the privacy page to cover the tracking posture in plain language a reader can actually scan.

Three places this is now disclosed:

Under every share bar. A small line of text under the share buttons reading: "No tracking on these buttons. Clicking opens the destination in a new tab; the publication doesn't see the click and sets no cookies for it. → What the site does and doesn't track." The link routes to the privacy page. Visible on every article and on the homepage, about page, frame page, sources page, for-government page, AI fan-out page, corrections page, and privacy page itself.

The privacy page. Expanded from two paragraphs to seven plain-language sections under a new title: "What this site tracks, and what it doesn't." Sections: The short version (one-sentence summary); What the site does (Google Analytics with anonymised IP, only after cookie consent, 14-month retention, EU-US Data Privacy Framework with UK extension); What the site does not do (no advertising, no Meta Pixel, no LinkedIn Insight tag, no X tracking pixel, no email collection, no comments, no account system, no fingerprinting); The share buttons specifically (each of the five buttons named with its data flow described — Copy link uses the clipboard API and sends nothing to any server; Post on X and LinkedIn are plain links opening compose pages with the URL pre-filled; Email is a mailto: link; system Share uses the Web Share API); The cookie banner (how consent is recorded, how to revisit the choice, where the storage key lives); Other data flows you should know about (Cloudflare hosting and the inbound-referrer flow, both honestly disclosed for the first time); Your rights (UK GDPR, ICO complaint route).

The about page. A new paragraph in the Method section: "What this site tracks. One thing only: anonymised page views via Google Analytics, and only after you click OK on the cookie banner. No advertisers, no third-party widgets, no email collection, no tracking pixels. The share buttons set no cookies and the publication does not see when they are clicked. The full plain-language account is on the privacy page."

One honest acknowledgement on the privacy page that wasn't there before. The previous privacy page mentioned only Google Analytics. The new page explicitly names Cloudflare as the host and discloses that Cloudflare sees standard HTTP request metadata (IP, user-agent, page requested, request timing) for bot mitigation and content delivery, with a link to Cloudflare's privacy policy. It also names the inbound-referrer flow that Google Analytics records as the source of a visit. These were both true under the previous policy but not previously disclosed. The publication's standing posture — log every change to what the site does — now applies to the privacy disclosure itself: the closing section commits to logging future changes to tracking on the corrections page with date and reason.

1 May 2026, 23:55 — Share bars added to every article and to the major top-level pages. No third-party widget code, no tracking; plain anchor tags plus a tiny clipboard-copy script. Each share bar renders the page's title and absolute URL pre-filled into the destination, with proper tap-target sizing on mobile.

Category: feature · accessibility · privacy

Doug requested that the publication's articles be made easily shareable. The publication already had passive shareability — Open Graph and Twitter Card meta tags on every page, with the title, description, URL, image (og-image.png at 1200×630), and locale set so that pasting any link into Slack, X, LinkedIn, iMessage, or any preview-card-rendering surface produces a proper preview rather than a blank link. What was missing was an active share affordance: visible buttons on the page that a reader could click to share the piece with one action.

The share bar. A new shareBar() function in build.js renders a small horizontal bar with five affordances: Copy link, Post on X, LinkedIn, Email, and a system-share button (using the Web Share API where the browser supports it; hidden by default and revealed by a tiny startup script when navigator.share is available — which is mostly mobile). Each share-bar button is a plain anchor tag (or a small button element for the two that need clipboard or Web Share API access), with the page's absolute URL and title pre-filled. No third-party widget code is loaded. No third-party cookies are set. The publication's privacy posture (Google Analytics with anonymised IP, nothing else) is preserved exactly. A reader with JavaScript disabled gets all the share links except Copy link and the system-share button, both of which require JavaScript.

Where the share bar appears. Every one of the 20 articles renders a share bar at the foot of the article, after the see-also panel, before the prev/next navigation. The major top-level pages — the about page, the sources page, the frame page, the for-government page, the AI fan-out page, the corrections page, and the homepage — also render a share bar. Pages that are not really shareable in the same sense — the privacy page, the terms page, the 404 page, the archive index, the interactive financial model — do not have a share bar. The intent is that a reader who wants to send a specific piece of analysis to someone has a one-click affordance; a reader on the privacy page or the 404 page does not.

Mobile behaviour. The share bar uses flex-wrap so it adapts to narrow viewports cleanly. Tap targets are at least 36px on mobile, meeting the practical floor for thumb input. On the smallest phones (under 380px wide) the text labels on Post on X, LinkedIn, Email, and the system-share button are visually hidden (with proper screen-reader-accessible markup preserved) so the icons alone fit on a single row; the Copy link label is kept because copying a URL is the most-used affordance and the most ambiguous as an icon-only target. Headless responsive testing across 320px, 375px, and 1280px viewports confirms the bar renders cleanly on all four tested page types (articles, about page, frame page, AI fan-out page) without horizontal overflow or undersized tap targets. The bar is hidden in print stylesheets — share buttons in printed output would be confusing.

What the share bar deliberately does not do. It does not include Facebook (which requires their tag-loading script for proper share-count display, and the publication does not want to load Meta's tracking scripts even on a per-click basis). It does not include WhatsApp's official share button (the whatsapp://send protocol works on mobile but produces a confusing experience on desktop, where the user is more likely to be reading; a reader who wants to send a piece via WhatsApp can use the system-share button on mobile or copy the URL). It does not auto-prepend the author's X handle to the tweet text, because the author's X presence is not a primary distribution channel for this publication and forcing it on every share would be inappropriate.

1 May 2026, 23:45 — New top-level page added: How this was actually made — the AI fan-out. Single-page disclosure of the multi-AI workflow that produced the publication, with verified counts from the central conversational thread alongside the author's estimate of the fan-out work in other AI tools.

Category: methodology · structural · honesty

Doug requested a single dedicated page that openly states the fact most readers would not infer from the publication's existing methodology framing: that the work was produced through a multi-AI fan-out. One central conversational thread (Claude, accessed through Claude.ai's chat interface, with code-execution and file-editing tools available) did the assembly, the building, the rechecks, the bundles, and the corrections-log discipline. Around that central thread, Doug ran many additional sessions of other Claudes, other ChatGPTs, and other Groks — for cross-critique, for second opinions, for source-pulling, for stress-testing arguments, for the AI-asked-to-pick exercise that produced the three companion pages, and for distributed work that fed back into the central thread as pastes. The shorthand "AI-tool-assisted" does not capture this; the new page does.

What the page contains. The page is at /the-ai-fan-out.html as a top-level methodology surface. It opens with the COI box, names the structure of the workflow (one central thread, many fan-out sessions), provides verifiable numerical counts from the transcripts of the central thread (33.2 hours wall-clock span, 11.8 hours active engagement and 21.4 hours idle time using a 5-minute-gap-threshold definition, 192 messages from the author, 1,892 tool calls executed in the central thread, 198 bundle deliveries, 53 distinct documented revisions on the corrections log), then provides the author's estimate of the fan-out work alongside (many tens to low hundreds of sessions across the four-week practice and this publication's day of concentrated work; no precise count kept). Names the AI-asked-to-pick companion series as the one specific visible instance of the pattern. Names Gemini as a tool that was attempted but dropped, with the author's reported reasons (could not effectively spider the site, could not read the bundled thelongerlook-site.zip files because each contained 50+ files); the publication does not draw a categorical conclusion about Gemini's capabilities and frames the limitation as the author's reported experience. Closes with what the central thread did and did not do (built all the HTML/CSS/JavaScript and downloadable artefacts, maintained the corrections-log discipline, ran the structural rewrites and assembly; did not generate the AI-asked-to-pick responses, did not conduct the fan-out cross-critiques, has no visibility into work that happened outside the central thread).

What the page does and does not do. It states openly that the publication's actual production was a multi-AI fan-out rather than a single-tool conversation. It gives verified numbers for the central thread and clearly marks the fan-out numbers as the author's estimate. It does not change the substantive analysis on any contested question. The publication still does not adjudicate either the principle question or the timing question; the four design positions A, B, C, D remain at equal length with case-for and case-against in the voice of each side's strongest defenders. This is methodology disclosure, not a substantive policy update.

Where it is linked from. The about page now carries a sentence at the foot of the methodology paragraph routing readers to the new disclosure. The methodology piece (twelve-hours.html) lead now links to it. The sitemap now includes it. llms.txt now describes it so AI tools summarising the publication know it exists. The corrections-log records its addition (this entry). The page is not in the top navigation — same reason the for-government page is not in the top navigation: the navigation surface is for primary reading routes, not for methodology disclosures. The page is reachable directly via URL, via the sitemap, via the about-page methodology paragraph, via the methodology piece itself, and via this corrections-log entry.

Side-effect cleanup on the methodology piece. The lead of the production-story methodology piece (twelve-hours.html) had said "Seventeen pieces of analysis (eleven featured, six alternative versions and methodology pieces)" — stale since the AI-asked-to-pick companion series brought the publication to twenty pieces. The lead now reads "Twenty pieces of analysis (eleven featured, six alternative versions and methodology pieces, a three-part AI-asked-to-pick companion series)." The sentence linking to the new fan-out page is appended to the same lead.

1 May 2026, 23:30 — Page three of three added: "What an AI tool said when asked to pick — ChatGPT Pro (GPT-5.5 Pro)." Companion series now complete. The exact prompt put to all three tools is now stated verbatim on each page; the placeholder text has been replaced.

Category: methodology · structural

The third and final page in the small companion series is now built. Doug supplied the response from ChatGPT Pro running GPT-5.5 Pro (OpenAI) to the same prompt put to Claude Opus 4.7 and Grok 4.3 Beta. The response is reproduced verbatim, with no editing of the response itself. The publication has added only the page chrome (title, COI box, frame-style explanatory box, methodology header, closing note from the publication).

The exact prompt is now stated on all three pages. Earlier today, when the Claude and Grok pages were built, the publication did not yet have the verbatim prompt text and so each page carried an explicit placeholder: "the exact prompt used will be added here when the publication confirms the wording." Doug has now supplied the prompt verbatim. All three pages now carry the same statement: "All three AI tools in this companion series were given the same prompt, verbatim: 'Now you have read all the arguments what would you do assuming you could define the UK government policy. Assume the government is making its decision in the best spirits to the benefit of the country.' Each tool was provided with the publication itself as context within the conversation. Each tool's reply is reproduced verbatim, with no editing of the response itself." The placeholder text on the Claude and Grok pages has been replaced. A duplicated stale paragraph that paraphrased a guess at the prompt has been removed from both earlier pages.

The date and time on all three pages. Each of the three pages now carries the same access date and time: 1 May 2026 at approximately 23:30 (UK time). The earlier-built Claude and Grok pages had previously stated only the date.

What the ChatGPT Pro page contains. The tool's verbatim response is approximately 1,100 words. ChatGPT lands on a two-track design that splits assets into an active-productive track (active farms and trading businesses) and a passive/portfolio/shelter track. For the active track: keep the £2.5m per-person allowance, allow a higher allowance (£5m) for genuine farm/business estates using a 60% minimum-share rule, apply effective 20% IHT above that, calculate at death but defer collection until liquidity (sale, IPO, dividend extraction, share buyback, liquidation, transfer outside the qualifying structure, ceasing to qualify, or company leaving the UK tax/substance net), and add an upper cap so very large estates cannot receive unlimited relief. For the passive track: no enhanced active-business deferral, no special treatment beyond a modest cap, ordinary IHT or the current effective 20% regime. Plus targeted transitional protection for long-held businesses owned by elderly taxpayers. Plus a regime review after 3 years using actual behavioural data. The response includes hyperlinks to IFS, GOV.UK, CenTax, and the House of Commons Library at URLs the tool itself supplied; the publication has preserved those URLs verbatim and has not independently verified that they resolve as the tool intended.

The closing notes on the Claude and Grok pages have been updated. Both earlier pages had said the third tool's response would be added when obtained. They now state that the companion series is complete and link directly to the ChatGPT Pro page. A short comparative observation has been added to all three closing notes describing where the three tools converge (all adopt some form of realisation mechanism for illiquid active-business assets; all lift the threshold above £2.5m, with £5m–£10m as the most-mentioned figure; all propose a built-in review mechanism at year 3 to year 5) and where they diverge (Claude switches the underlying mechanism wholesale; Grok layers a realisation election on top of the existing death-event regime; ChatGPT splits assets into active and passive tracks with different treatment for each).

Where the page lives and how it is framed. Same shape as the Claude and Grok pages: registered as an article (slug 2026-04-30-ai-asked-chatgpt-pro-gpt-5-5-pro, dated 1 May 2026), hidden from the homepage's primary featured-pieces list, in the archive and the sitemap and the methodology-and-limits group of the contextual navigation. The publication's own posture (does-not-pick) is restated explicitly on each of the three pages: in the top frame-box, in the methodology header, and in the closing note. A casual reader landing on any of the three pages cold sees, before they reach the AI's response, two prominent boxes explaining what the page is and what it is not, plus a methodology header. After the response, the closing note repeats the framing.

1 May 2026 — Page two of three added: "What an AI tool said when asked to pick — Grok 4.3 Beta." Verbatim Grok response, same template as the Claude Opus 4.7 page. ChatGPT page still pending.

Category: methodology · structural

The second page in the small companion series is now built. Doug supplied the response from Grok 4.3 Beta (xAI) to the same prompt put to Claude Opus 4.7. The response is reproduced verbatim, with no editing of the response itself. The publication has added only the page chrome (title, COI box, frame-style explanatory box, methodology header, closing note from the publication).

What the Grok 4.3 Beta page contains. The tool's verbatim response is approximately 700 words. Grok's pick differs from Claude's in shape: Grok lands on a hybrid of Position A's core form (keep the death-event mechanism) with a B-style realisation election layered on top, plus a D-style cap lift to £5m/£10m for qualifying unlisted UK trading companies, plus an additional package of practical measures (binding HMRC guidance on instalment payments, fast-track valuation panel for unlisted shares, mandatory annual OBR/HMRC behavioural-data report from 2027, sunset-and-review at 2028-2029 with a presumption that the realisation election becomes default unless evidence shows otherwise). The response numbers the four positions as 1, 2, 3, 4 rather than the publication's A, B, C, D — a small editor's note in the page chrome flags this for the reader, with the explicit observation that the substantive content of each position is the same and only the labelling differs. The publication has not relabelled the response itself.

Where the page lives and how it is framed. Same shape as the Claude page: registered as an article (slug 2026-04-30-ai-asked-grok-4-3-beta, dated 1 May 2026), hidden from the homepage's primary featured-pieces list, in the archive and the sitemap and the methodology-and-limits group of the contextual navigation. The closing note on the Claude page has been updated to link directly to the Grok page rather than referring to a future publication, and now adds a short comparative observation noting where the two tools converge (threshold-lift size; adopting some realisation mechanism for illiquid assets) and where they diverge (Claude switches the mechanism wholesale, Grok layers a realisation election on the existing regime). The closing note on the Grok page mirrors the same comparative observation. Both observations are marked as descriptive, not endorsing.

What is intentionally not yet on the page. The exact prompt used was not supplied at the time the page was built, exactly as on the Claude page. The page carries the same explicit placeholder note: "the exact prompt used will be added here when the publication confirms the wording." When Doug confirms the prompt wording, the placeholder will be replaced with the verbatim text on both pages simultaneously.

The third page (ChatGPT) is still pending. It will be added when the response is obtained, with the same template, the same chrome, and the same closing-note framing.

1 May 2026 — New page added: "What an AI tool said when asked to pick — Claude Opus 4.7"; the first of a small companion series in which an AI tool is asked, after reading the publication, what it would do if it had to set the policy. The tool's response is reproduced verbatim and is explicitly not the publication's view.

Category: methodology · structural

Doug requested the publication add three companion pages — one each for Claude Opus 4.7, ChatGPT, and Grok — in which the AI tool is asked to read the publication and answer the question "if you had to set the policy now, what would you do?" The tools' responses to be published verbatim, with no editing. The first page (Claude Opus 4.7) is now built. The pages for ChatGPT and Grok will be added when those responses are obtained.

What this changes about the publication. The publication has spent the last day removing verdict-language from its own voice across every analytical piece. The principle piece, the timing piece, the four-positions architecture in the long article, the readable companion, the short version, the five-minute version, and the audience-specific pieces have all been pulled to two-sided-no-verdict. The frame disclosure on five load-bearing pieces names the lens within which the analysis is conducted but stops short of verdict on the contested questions. This new page series puts verdict-language back onto the site, but in a different voice: it is an AI tool's verdict, reproduced verbatim, with the publication's own posture (does-not-pick) explicitly preserved around it.

The framing the new pages carry. Each of the three pages opens with the COI box and a frame-disclosure-style explanatory box that names what the page is and what it is not. The page text is explicit: "This page is not the publication's view. It is one AI tool's answer when asked to pick. The publication does not adopt this tool's pick. The publication's posture remains: the principle question and the timing question are genuinely contested; the four design positions A, B, C, D, E, F each have a strongest case for and a strongest case against; the analysis is conducted within a disclosed frame which is one of five defensible UK-national-interest frames; the publication does not adjudicate." The closing methodology note repeats this. A casual reader who lands on the page should still be in no doubt that what they are reading is one AI tool's response, not the publication's own conclusion.

What the Claude Opus 4.7 page contains. The tool's verbatim response is approximately 1,600 words. It opens with the tool reading the publication and acknowledging the structural refusal to adjudicate. It then accepts the explicit invitation to pick. The tool lands on Position B (CGT-on-realisation) with a Position C review trigger at year five, a Position D threshold lift to £5m/£10m for qualifying unlisted trading-company shares specifically, and adoption of the CenTax minimum-share rule. It explicitly accepts the principle question on horizontal-equity grounds. It is forthright about the load-bearing empirical claims it relies on (pre-death relocation magnitude, lock-in elasticity for the BPR-affected cohort) and openly flags the implementation problems it has not worked through (basis tracking across decades, deemed-realisation events for non-arms-length transfers, interaction with existing CGT on heir-side gains, the Australian-comparator workaround industry).

Where the page lives. The page is registered as an article (slug 2026-04-30-ai-asked-claude-opus-4-7, dated 1 May 2026). It is hidden from the homepage's primary featured-pieces list — putting it on the homepage would imply it is part of the publication's own analysis, which it is not. It is in the archive, the sitemap, and the contextual see-also navigation. It is grouped under "methodology and limits", alongside the production-story piece, the iterative-process record, the common-reactions piece, and the corrections page itself.

What is intentionally not yet on the page. The exact prompt used was not supplied at the time the page was built. The page carries an explicit placeholder note: "the exact prompt used will be added here when the publication confirms the wording." Inventing a prompt would be dishonest methodology disclosure; flagging the gap is the correct move. When Doug confirms the prompt wording, the placeholder will be replaced with the verbatim text. The same applies to the ChatGPT and Grok pages: they will be built from the same template when the responses are obtained, with the prompt text included as soon as it is confirmed.

1 May 2026 — IHT publication production-time framing corrected: "four weeks" was conflated with the IHT day's work in three passages; corrected to one day. Three of four downloads regenerated to reflect today's rewrites; two remain stale and are flagged.

Category: framing · honesty · downloads

Doug flagged a conflation. The methodology piece, the homepage, and the about page have been clear that the IHT publication itself was produced in roughly eight hours of real work, and that those eight hours sit inside a broader four-week practice that produced books, sites, code, and The Many Builders. But three passages elsewhere in the corpus ran the two together — "the publication put four weeks of intensive AI-assisted work into producing analysis at depth on this question" — which says four weeks of work went into the IHT analysis specifically, when in fact the IHT analysis was a single day's work and the four weeks is the broader practice the day's work draws on.

Three passages corrected:

The about page. The sentence "the reason the publication chose this question, and put four weeks of intensive AI-assisted work into producing analysis at depth on it, is that the question reaches a cohort whose attention to the larger questions about AI and humanity might matter" has been rewritten to "the reason the publication chose this question, and put a day of concentrated AI-assisted work into producing analysis at depth on it, is that the question reaches a cohort whose attention to the larger questions about AI and humanity might matter. The IHT publication was produced in one day; the broader four-week practice (the books, the trilogy, The Many Builders where the bears creating the new world live, and the earlier sites and code) is the context the IHT day's work sits inside — not the work that went into the IHT analysis itself."

The long article coda. The first-person closing question "why has this person put four weeks of intensive AI-assisted work into producing institutional-grade analysis on a specific UK tax-policy question" is now "why has this person put a day of concentrated AI-assisted work into producing institutional-grade analysis on a specific UK tax-policy question."

The readable companion coda. The same first-person closing question, in a slightly shorter form, has been corrected the same way.

The methodology piece itself remains unchanged — it has always been clear that the IHT publication was eight hours of real work and that the four weeks is the broader context. The site-bar across the top of every page still labels itself "Also by Doug Scott — other works in the four-week practice", which is correct: it is referring to the books, the bears, and the other sites, all of which were produced in the four-week practice. The about page heading "The four-week practice — what else has been built" is correct: it refers to the broader practice and lists what was built across it.

Three of four downloads regenerated. The .docx and .pdf files for the citizen-submission policy paper, the long article, and the readable companion were regenerated on the evening of 1 May 2026 from the rewritten body files using the existing build-doc.js, build-treasury-paper.js, and build-readable.js generators (PDFs converted via headless LibreOffice). They now reflect today's substantive rewrites: the principle piece and timing piece rewrites, the four-positions architecture flatten across the long article and readable companion, the institutional cross-references, the comparison-to-government cleanup, and the consistency cleanup. Two artefacts remain stale: the funding-stack technical companion (.docx and .pdf — no build script currently produces these), and the Excel companion to the interactive model (.xlsx — no regeneration script). The live website pieces remain canonical for those two.

The staleness notes on the reading guide and the for-government page have been updated to reflect the partial regeneration: the policy paper, long article, and readable companion downloads now match the live site; the funding-stack download and the Excel companion remain stale and are openly flagged.

1 May 2026 — Consistency cleanup: reading guide rewritten, principle piece's standfirst and lead aligned with body, residual position-claiming language removed from twelve-hours, readable companion intro, plain-english-detailed, and llms.txt

Category: framing · consistency

Doug flagged a real inconsistency: the homepage and corrections log had been pulled to the publication-does-not-adjudicate posture across today's earlier sweeps, but several pieces still carried the older position-taking framing that contradicted what the rest of the publication now says. The reading guide had a heading "If you want the publication's actual position", a closing section "What the publication actually thinks, in three sentences", and a recommended reading description that flatly said "The principle of the reform is right. The strongest objection that lands is about asset-class fitness for genuine operating family businesses versus founder equity. The two-track design (threshold mechanism for founder equity, German-style conditional relief for operating family businesses) is the publication's most interesting policy proposal." The principle piece's "Who this is for" line and lead paragraph still framed the question as "the principle of the reform is right; the strongest objection is about asset-class fitness; the answer is a two-track design", which contradicted the rewritten body of the same piece. The methodology piece's account of "the substantive view on the IHT question" still set out the realisation-design verdict the rest of the publication had moved past. Three other body files had small but real residual instances. The llms.txt file (the AI-tools summary) described the publication as "the only one that takes a position on a public-policy controversy" and described the principle piece as "Why the principle of taxing very large inherited business wealth is right."

The fixes applied across the corpus:

Reading guide rewritten. The position-taking heading is gone. The closing three-sentence position is replaced with a three-sentence description of the publication's actual posture: that it does not adjudicate either the principle question or the timing question, that the four design positions are presented at equal length, and that the analysis is conducted within a disclosed frame which is one of five defensible UK-national-interest frames. The five-minute version is now the recommended starting piece for "if you have five minutes" (replacing the earlier framing that pointed at the timing piece as "the publication's strongest single argument"). New entries route the reader to the for-government page (for government readers), the frame page (for the lens within which the analysis is conducted), and the long article (for the operational analysis at depth). Each piece's description has been rewritten to describe what the piece does rather than claim a position the piece is supposed to demonstrate. The downloadable-versions section now carries an explicit note that the downloads were last regenerated on 30 April 2026 morning and have not been updated to reflect the substantive rewrites of 1 May 2026.

Principle piece "Who this is for" line and lead paragraph aligned with the rewritten body. The earlier "the publication sets out the strongest cases. The principle of the reform is right; the strongest objection is about asset-class fitness; the answer is a two-track design rather than abolition" is replaced with "the publication sets out the strongest cases for and against taxing very large intergenerational business-wealth transfers at roughly equal length, in the voice of each side's strongest defenders, without a closing verdict." The lead paragraph's "threshold versus conditional; one mechanism versus a two-track design" framing is replaced with "threshold versus mechanism change, and the four design positions A, B, C, D, E, F presented at equal length."

Methodology piece (twelve-hours) substantive-view paragraph rewritten. The earlier "the principle of the reform is right, the strongest objection that lands is about the timing rather than the amount, and a realisation-based design would deliver the same fairness with fewer of the side effects critics worry about most. That is the publication's position on the IHT question." is replaced with "Both pieces present the strongest case on each side at roughly equal length, in the voice of each side's strongest defenders, without a closing verdict from the publication. The publication does not adjudicate either question; it sets out the contested questions and lets the reader weigh them."

Readable companion intro fixed. The earlier "a single-sentence version of the publication's position, see When, Not How Much" (which also used the stale article title) is replaced with "if you want the timing question taken on directly, with the strongest case for tax-at-death and the strongest case for tax-at-realisation presented at equal length, see The Amount Question and the Timing Question."

Plain-english-detailed piece's framing of the principle question fixed. The earlier "the question is not whether the principle of the reform is right" (which implied the principle was settled) is replaced with "the question of whether the principle of the reform is right is genuinely contested and is taken on directly in the principle piece, which presents the strongest case for and the strongest case against at equal length without verdict."

llms.txt updated. The publication is no longer described to AI tools as "the only one that takes a position on a public-policy controversy." The new description names the publication's actual posture: that it does not adjudicate either the principle question or the timing question, that both are presented two-sided, and that the analysis is conducted within a disclosed frame. The principle piece's description is rewritten to describe the two-sided structure rather than to claim a settled answer. The timing piece's title is corrected (it is now The Amount Question and the Timing Question, not When, Not How Much). New entries are added for the frame page, the for-government page, the sources page, and the corrections page so AI tools summarising the publication can reach those surfaces.

What this fix does and what it does not do. It brings the active corpus into consistent alignment with the publication's stated posture (does not adjudicate; presents both sides at equal length; operates within a disclosed frame). It does not change the substantive analysis on any contested question. The downloadable Word and PDF artefacts in /downloads/ remain stale (they were last regenerated on 30 April 2026 morning) and the corrections page already records that staleness. A reader downloading the Word version of the long article still gets the previous version of the analysis; a reader using the live site gets the current version.

1 May 2026 — Mobile responsive fixes: site-bar wrapping, downloads-block overflow, tap-target sizing across the components added today

Category: technical · accessibility

Doug asked the publication be checked on all devices. Headless responsive testing using Chromium across six viewports (320px, 375px, 414px on mobile; 768px on tablet; 1280px and 1920px on desktop) and ten representative pages (homepage, frame, for-government, long article, timing piece, fiscal model, about, corrections, five-minute version, readable companion) found three concrete bugs and several tap-target-sizing issues introduced by the components added today (frame box, see-also panel, group badge, COI box).

Tablet and desktop: zero issues. The publication renders correctly on iPad-portrait (768px), standard desktop (1280px), and large desktop (1920px) viewports without overflow, without sizing problems, and without layout breakage.

Mobile bugs identified and fixed. The site-bar (the strip across the top of every page linking to the seven sites in Doug's project) had white-space: nowrap on its links, producing horizontal overflow at 320-414px because the seven links did not fit on one line. Wrapping is now allowed at 600px and below; vertical-rule separators between items are hidden when wrapping (they look wrong at line breaks). The downloads block (used on the long article and the fiscal model page) used margin: ... calc(-1 * var(--gutter)) ... to bleed full-width on desktop, which on mobile pushed the box 4-8px past the viewport edge; the negative margin is now zeroed at 600px and below. Tap targets across the components added today (the COI box, the frame box, the see-also panel, the group badge, the institutional cross-reference URLs) have been bumped to 28-36px on mobile so they meet practical tap-target sizing for thumb input. The home-masthead navigation links and AI-warning-strip links — which had been at 15-30px tall on mobile — are now bumped to 32-36px.

What was deliberately not changed. Inline links inside running prose (the email address on the about page, in-text mentions of "the timing piece" or "the reading guide" inside paragraphs, the cookie-banner "More info" link styled as inline help text) remain at body-text size (15-19px tall). Bumping these to 32px+ would break the typography of paragraphs they sit inside; the WCAG and Material Design tap-target guidance is for primary navigation elements, not for inline-in-prose links which are conventionally at the same size as the surrounding text. The publication's body-text links have always been at body-text size; the test flagging them is a false positive given the typography convention.

Test methodology, made transparent. The responsive verification was done with Playwright + Chromium running headlessly. The test checked: (a) horizontal overflow at the document level (document.documentElement.scrollWidth > clientWidth); (b) any element whose right edge exceeded the viewport width; (c) tap-target sizing on standalone interactive elements (links inside primary navigation, footer items, callout buttons), excluding inline-in-prose links and off-screen accessibility content. The script and the methodology can be re-run by any future maintainer to verify the publication continues to work correctly across viewports.

The CSS fixes are appended at the end of assets/style.css in a final-pass section so they win specificity ties against the existing styles. The publication's design language (EB Garamond serif body, Inter sans-serif chrome, ink-blue / cream / bronze palette) is preserved unchanged at every viewport; the fixes are layout-only and do not change the publication's visual identity.

1 May 2026 — Institutional cross-reference annotations added to load-bearing pieces: equivalent arguments in formal consultation responses, parliamentary committee reports, and named UK institutional sources now identified alongside the publication's own analysis

Category: institutional · sources

Doug raised a good point: pointing the reader to where the equivalent arguments appear in formal consultation responses carries more weight institutionally than the publication's own argument carrying alone. The work was already done in part — the principle piece has had a substantial "What other UK institutions have published on this question" section since earlier in the day, citing the IFS, Resolution Foundation, CenTax, FBRF, Commons Library, and CIOT. The remaining load-bearing pieces did not have equivalent institutional cross-references in place. This round adds them.

The timing piece (The Amount Question and the Timing Question) now has a "Where equivalent arguments appear in formal institutional commentary" section between the case-comparison closing paragraph and the operational-pieces routing. The section identifies the House of Lords Economic Affairs Finance Bill Sub-Committee report (January 2026, with the parliamentary URL), the CIOT consultation responses on the draft legislation and on trusts (with the tax.org.uk reference URLs), the CIOT/ATT joint commentary in Tax Adviser, the IFS Adam-Miller-Sturrock 2024 paper and the Resolution Foundation Budget 2024 briefing, the CenTax alternative-design proposals, and the FBRF Kemp 2025 report and the FBRF/Cebr research project running December 2025 to May 2026.

The funding-stack piece (UK Tech and the IHT Reform — The Funding Stack and the Fiscal Model) now has a "Where the cohort-segmentation arguments appear in formal institutional commentary" section between the central-case modelling discussion and the model-engagement closing. The section maps the publication's cohort-segmentation arguments (instalment-regime adequacy, valuation administrative burden, cohort-specific behavioural-response data) to specific institutional sources: the Lords Sub-Committee report's recommendations on extending the IHT payment deadline to 12 months and creating a statutory safe-harbour for personal representatives; the CIOT's formal consultation responses on valuation administrative burden; the FBRF/Cebr research project on family-business cohort impact; the CenTax alternative-design proposals; and the Saffery, KPMG, BDO, Deloitte, Royal London, BKL, PKF Francis Clark, and Hatchers professional commentary stack.

The long article (The April 2026 BPR Reform — A Policy Options Analysis) now has a new Section 5 ("Where the analysis aligns and diverges from formal institutional commentary") between Section 4 (international comparators) and the renumbered Section 6 (what different evidence would mean). The section walks each of the design positions A, B, C, D and identifies where it converges with named institutional positions: Position A's principle case alongside the IFS and Resolution Foundation work; Position B's realisation-based mechanism case as the publication's own synthesis (no UK institution carries this position directly); Position C's deferral-and-practical-fixes posture alongside the Lords Sub-Committee's recommendations; Position D's threshold-raise case alongside CenTax's alternative-design family with explicit acknowledgement that CenTax's empirical foundations are stronger than the publication's. The article's earlier Sections 6 and 7 (limits of the analysis; closing) have been renumbered accordingly.

The readable companion (What the Reform Means for UK Tech) has a single new paragraph under the "If you want to go deeper" section pointing the reader at the institutional sources directly: the Lords Sub-Committee report, the CIOT consultation responses, the IFS, CenTax, and the FBRF/Cebr research project. The paragraph keeps the readable-companion register short rather than reproducing the full cross-reference work that the long article now does at depth.

The neutral framing this work uses, and why. The cross-references are written in the form "this argument also appears in [body]" — not "this body endorses the publication" or "this is supported by". The IFS does not endorse this publication; CenTax has done its own better-grounded analysis using HMRC microdata; the Lords Sub-Committee makes its own recommendations on its own terms. Saying the publication is endorsed by these bodies would be misleading; saying that arguments converge on specific questions, while the institutional sources remain better-resourced and the publication is one input among many, is the honest framing. Each cross-reference is placed alongside the specific argument it corresponds to rather than aggregated into a "supported by" panel.

What this fixes and what it does not. A reader who finds an argument in the publication can now find the formal institutional location where the equivalent argument appears, with the working URL, and can verify the institutional position directly rather than taking the publication's word for it. This is a step toward the bigger correction the publication has worked toward across the day: the publication is one input among many, the institutional sources are the better-resourced ones, and the publication's reader is best served when the institutional sources are visible alongside the publication's analysis. What it does not do is validate the publication's own analytical choices or its frame; the cross-references tell a reader where institutional arguments converge with the publication, not that the publication has got the analysis right.

1 May 2026 — Removed all "the government has not published / done X" claims from active body files

Category: framing · honesty

Doug flagged a residual honesty problem: the publication had, in three places in the active corpus, made claims about what the government has or has not done internally — "the government has not published detailed modelling on the cohort-specific behavioural response," "the government should publish proper behavioural and fiscal modelling before assuming the mechanism works," and "HMRC has not published the modelling that would isolate the BPR-specific elasticity from the wider tax-policy environment." Doug's point: the publication has no visibility into what HMT, HMRC, the OBR, or any other relevant body has done internally. Saying the government has not done something it might well have done internally and not published is a claim the publication cannot make.

The three passages have been rewritten to talk about the publicly available evidence base rather than about what the government has or has not done. The short-version piece's closing observation now says "the publicly available evidence base on the cohort-specific behavioural response is thin" followed by the specific limits of the three publicly cited sources (the Friedman et al. LSE Working Paper sample, the OBR's 25% non-doms figure being from a parallel reform, the Companies House data being contaminated by simultaneous reforms). The common-reactions piece's reviewer-quote has been re-rendered as "proper behavioural and fiscal modelling would be useful before the mechanism is assumed to work" with an explicit added line: "The publication does not know what work has or has not been done internally on these questions inside HMT, HMRC, the OBR, or any other relevant body; the publicly available material does not contain the cohort-specific modelling the analysis requires." The funding-stack piece's central-case discussion now says "the behavioural-response data needed to discriminate between scenarios does not yet exist in the publicly available literature. No published study isolates the BPR-specific elasticity from the wider tax-policy environment."

The fix is small in word-count but matters in register: the publication has stopped contrasting itself favourably to government work it cannot see, and now talks only about what the publicly available evidence does and does not contain. A reader can verify what is in the publicly available literature; a reader cannot verify what is or is not in HMT's internal modelling. The publication confines itself to the verifiable claim.

This change was already partly recorded in the corrections-page audit trail from earlier rounds: the bigger-claim piece was retired specifically because "the comparative-depth claim was retracted" and "the publication is no longer presenting itself as having produced more depth than the government on this question." Three residual instances of the same comparison-to-government framing survived in body files that were not part of the bigger-claim piece. Those instances have now been removed. Older corrections-page entries describing the bigger-claim piece's original framing remain intact as historical record.

1 May 2026 — Article-level contextual navigation added: piece-type group badge at the top of every article; "see also" panel with three context-aware recommendations at the end

Category: structural

Doug noted that an article reader currently has only chronological prev/next navigation (← Older / Newer →) and the general reading-guide page; there was no contextual navigation telling a reader on a given piece "this is what kind of piece this is, and these are the natural pieces to read next given what you just finished." With seventeen pieces all dated 30 April 2026, the prev/next ordering is essentially arbitrary — a reader who finishes the principle piece is taken to a methodology piece by chronology rather than to the timing piece by topic.

Two additions, both visible on every article rendered by the build:

A piece-type group badge. Below the date eyebrow and above the title, every article now shows a small uppercase label naming the kind of piece it is. Five categories: Compressed summary (the five-minute version, the short version); Audience-specific (the for-founders, for-journalists, for-tax-practitioners, plain-english-overview, and plain-english-detailed pieces); Operational analysis (the long article, the readable companion, the funding-stack piece, the iht-companies short version); The contested questions (the principle piece, the timing piece — the two pieces that explicitly present both sides); Methodology and limits (the reading guide, twelve-hours, how-this-was-made, common-reactions). The categorisation is descriptive of what the piece is, not prescriptive of which piece is most important.

A "Where to go next" panel. Just before the existing prev/next navigation at the foot of the article, every piece now shows a contextual see-also panel with three recommended next pieces and a one-sentence note on what each adds. The recommendations are per-piece — they are the natural follow-ons given what the reader just finished, not a generic list. The panel also points the reader at the general reading guide (for the full map by reader type) and the for-government page (for routing by team).

Examples of the per-piece recommendations: the principle piece's "Where to go next" lists the timing piece ("the timing question, both sides at equal length"), the long article ("the operational analysis if the principle is accepted"), and the frame page ("the frame the operational analysis is conducted within"). The funding-stack piece's lists the long article, the timing piece, and the for-founders piece. The five-minute version's lists the principle piece, the timing piece, and the short version. Each recommendation is chosen for what the reader is most likely to want immediately after finishing the piece they are on.

What this fixes. The reader who lands on a piece directly (via search, via shared link, via the for-government routing page, via the bibliography of another publication) previously had no way to find the most relevant adjacent pieces without navigating to the homepage or archive. The contextual navigation now closes that gap on every article without requiring the reader to leave the piece they are on. The general reading guide and the for-government routing page both remain — the contextual navigation is per-piece additive rather than a replacement.

What this does not do. It does not adjudicate which piece is most important or most reliable. It does not promote any one of the four design positions over the others (the see-also lists for the long article, the readable companion, and the funding-stack piece all route to the timing piece and the principle piece, both of which are presented two-sided without verdict — not to a position-taking deeper-dive that does not exist). It does not add any new analytical writing; it is a navigational layer on the existing corpus.

The metadata for the navigation lives in site-config.js (each article now has a group field and a seeAlso array). A subsequent reviewer who thinks the categorisation or the per-piece recommendations are wrong can submit the change via the submit-a-correction route at the top of this page; the metadata is small enough that any individual recommendation can be revised without rewriting analytical content.

1 May 2026 — Government-team routing page added: /for-government.html

Category: structural

Doug shared a routing table from another reader mapping four government teams (HMT/HMRC, DBT/Innovation, No.10/SpAds, Communications/Press Office) to the publication's existing pieces. The mapping worked against the existing corpus, but the publication did not have a top-level routing page that surfaced the correspondence to a government reader landing on the site. A new page /for-government.html has been added that does this work.

The page is a routing tool, not new analysis. The judgement was that adding three new audience-specific analytical pieces (one to HMT, one to DBT, one to No.10) would have been additive volume without additive substance — the analytical work each team needs already exists in the corpus, and writing pieces "to government teams" is a register that pulls the analysis toward advocacy ("here is what the government should do"), which the day's earlier sweeps had specifically removed. The routing page closes the gap a government reader actually has — finding the right piece to start with — without re-introducing the lean.

What the page contains. The page carries the same COI box and frame disclosure as the analytical pieces. It then routes seven team-types to existing pieces with one paragraph of context per team and a recommended-deeper-reading list: HMT and HMRC IHT/BPR/capital-taxes policy teams (start with the timing piece, deeper reading the funding-stack piece, the long article, the for-tax-practitioners piece, plus the interactive model); DBT Tech & Innovation policy (start with the funding-stack piece, deeper reading the for-founders piece and the readable companion, plus the model); No.10 and special advisers (start with the five-minute version, deeper reading the principle piece, the short version, and the common-reactions piece); communications and press office (start with the for-journalists piece and the sources page); parliamentary committee staff (start with the long article, deeper reading the principle piece, the timing piece, the funding-stack piece, and the sources page); OBR analysts (start with the interactive model and the Excel companion, deeper reading the funding-stack piece for the cohort segmentation). The page closes with a "what this publication does not have, and where to go for it" section naming four pieces of work the government team might want that this publication is not the right source for: cohort-specific UK behavioural data on pre-death relocation; cohort-specific UK heir-productivity data; CenTax's distributional analysis of alternative designs; independent fiscal review of the model.

What the page does not do. It does not adjudicate which policy direction a government team should take. It does not recommend one of the design positions over another. It does not present the publication as authoritative for any government team's work. It explicitly names where the publication is weaker than the institutional sources (CenTax, IFS, Resolution Foundation, the Commons Library briefing, the OBR's January 2025 supplementary release) and routes the team to those sources.

The page is in the sitemap and accessible at /for-government.html directly. The page is not currently in the top navigation — the navigation remains Home / Model / Sources / Archive / About / Corrections — because adding a "For government" entry to the top nav of a site that includes the corrections page would create a tonal mismatch (the publication is not, on its own framing, written-for-government work; it is one citizen's analytical contribution that government teams may find useful). A government reader landing via search, via shared link, or via the routing table itself can reach the page directly. The reading guide and the for-journalists piece both link to it for readers in the relevant categories.

1 May 2026 — Frame disclosure expanded: dedicated /frame.html page added; inline frame boxes on the five load-bearing pieces expanded to name the four mechanisms and link to the full disclosure

Category: framing

Doug instructed: "the author states the analysis assumes retaining high-growth tech talent in the UK is good and produces a compounding flywheel. Expand." The earlier round had added a one-paragraph frame disclosure to the top of the five load-bearing analytical pieces. Doug requested expansion: the frame should be set out at the depth a reader needs to engage with it seriously, not at the depth of a callout box.

A new dedicated page /frame.html has been added. The page is structured in eight sections: The frame, stated plainly (the directional retention claim, the composition claim, the compounding claim); The mechanism the flywheel argument rests on (cluster effects; capital recycling; talent attraction; productive complementarity — each named, each with its supporting literature directionally identified, each with its empirical magnitude flagged as contested); What is contested within the frame itself (magnitude of pre-death relocation; persistence of UK tax base after departure; heir productivity); The alternative frames, in their strongest forms (five frames presented in the voice of their strongest defenders, in roughly equal length: horizontal equity, fiscal stability, anti-avoidance, reduction of inherited advantage, not-driving-wealth-abroad); How the chosen frame shapes the analytical pieces (the cohort the analysis treats; the cohort-segmentation in the funding-stack piece; the fiscal model's central case; the four positions in the long article); What a reader can do with this disclosure (three readings: the reader who shares the frame; the reader who rejects it; the reader who is undecided); and What this disclosure does not do (does not adjudicate the principle question, the timing question, the four design positions, or which UK-national-interest frame is correct).

The inline frame box on each of the five load-bearing pieces has been expanded from one paragraph to three. Paragraph one names the directional flywheel claim and the four mechanisms it rests on, and notes that the mechanisms are evidence-supported but not settled at the magnitude the analysis requires. Paragraph two names the five alternative defensible UK-national-interest frames and links to the frame page where each is set out in full. Paragraph three reiterates that the frame is separate from the principle and timing questions on which the publication does not adjudicate, and links to the frame page for the full disclosure.

What the expanded frame page does that the inline box cannot. The inline box could disclose the directional claim. It could not, in the space available, set out what the flywheel argument's mechanism actually is, what evidence supports it, what is contested within it, or what the alternative frames look like in their own strongest terms. Without those four, the frame disclosure was potentially performative — a box that said "the frame is X" without giving the reader the materials to engage with whether X is right. The expansion makes the engagement substantively possible: a reader can now read the frame page and decide whether they share the four mechanisms (cluster effects, capital recycling, talent attraction, productive complementarity), whether they accept the contested empirical claims on which the analysis depends, and which of the alternative frames they would weight more heavily if they reject the chosen one.

What the expansion does not do. It does not advocate for the flywheel frame. It describes the frame, names the mechanism, names what is supported and what is contested in the supporting evidence, sets out the alternative frames in their strongest form, and stops short of arguing that the flywheel frame is correct. The author's stated view that the frame is held — and the author's earlier statement that he believes "keeping that talent and the ripples are to the benefit of all" — is acknowledged on the frame page as the lens within which the analysis is conducted, not as a verdict the page is asking the reader to accept. The five alternative frames are written in the voice of their strongest defenders, in roughly equal length, without rebuttal in the same paragraph.

The Sources nav now includes a Frame entry. The frame page is added to the sitemap. llms.txt will be updated in a subsequent edit to reference the frame page in the Sources section so AI tools summarising the publication know the frame is disclosed at depth on a dedicated page.

Why this round is consistent with the corrections-page treadmill caveat. The amber note at the top of this page warns that the corrections log may be documenting a recursive AI-cross-critique loop rather than convergent error-correction. This round is not a recursion of yesterday's frame-disclosure entry. The earlier entry stated the frame in one paragraph; Doug requested expansion; the expansion is a new piece of work that adds a dedicated page and substantially extends the disclosure depth. A subsequent reviewer who reads the expanded frame page and recommends contracting it back to one paragraph, or rewriting it to advocate more strongly, or removing it entirely, is being told the publication has settled on this disclosure depth for now. The frame disclosure at the depth the dedicated page provides is the publication's stated answer to "as honest as possible about scope-bias without verdict on the principle or timing questions."

1 May 2026 — Frame disclosure added to the five load-bearing analytical pieces; the scope-bias the publication has been carrying is now visible on the surface, separately from the principle and timing questions on which the publication does not adjudicate

Category: framing

A reviewer (1 May 2026) correctly identified that today's earlier sweep — which removed verdict-language from the analytical pieces and made the principle question and the timing question genuinely two-sided — addressed the verdict-language bias but did not address a different bias the reviewer named more precisely: the scope-bias in choosing UK tech founders as the cohort, treating tech-talent retention as the central national-interest lens, and conducting the cohort-segmentation and fiscal-model analysis within that frame. The reviewer's specific framing: "founder-aligned, national-interest advocacy with serious analytical effort — not neutral IHT analysis. The bias is in weighting: heavy importance on future economic dynamism and founder retention; a different analyst might give heavier importance to equal treatment, tax base protection, and anti-avoidance."

Doug acknowledged the reviewer's framing as fair and stated his actual view openly: "keeping that talent and the ripples are to the benefit of all." Doug requested a frame-disclosure box that names the scope-bias on the surface of the load-bearing pieces, without putting his view on the principle or timing questions back onto the page (those remain genuinely two-sided per today's earlier rewrites).

A frame-disclosure box has been added to the top of the five load-bearing analytical pieces (the long article, the readable companion, the funding-stack piece, the principle piece, and the timing piece), positioned just below the COI box and rendered in a distinct cream-and-bronze styling so a reader sees it as separate from the COI fact. The text:

Frame disclosure. This analysis is conducted within a frame the author holds: that retaining high-growth technology talent in the UK is good for the country, and that the capital, the further talent, and the productive activity it attracts compound into a wider flywheel. The frame is contested. Other defensible UK-national-interest frames place primary weight on horizontal equity in tax treatment, on revenue for public services, on reducing inherited advantage, or on protecting the broader tax base from carve-outs. The analysis below is conducted within the author's chosen frame; a reader who rejects the frame may reach different conclusions from the same evidence. The principle and timing questions on which this publication does not adjudicate are separate from this frame.

Why the wording stops short of "to the benefit of all." Doug's stated view, in his own words, is that the flywheel is "to the benefit of all." The frame-disclosure box names the directional claim (retention is good for the country; capital and talent compound into a flywheel) but stops short of the universal claim (this is to everyone's benefit), because the universal claim is the part the reviewer's other-defensible-frames list contests directly. The reviewer's point is that HM Treasury's broader-fairer-tax-base view, public-service advocates' more-revenue view, equality-focused economists' reduced-inherited-advantage view, and the founders' own not-driving-wealth-abroad view are all defensible UK-national-interest claims; "to the benefit of all" would assert that the retention frame trumps the others, which is a verdict the publication has worked all day to remove from the analytical pieces. The disclosed frame is a stated lens, not a verdict on which lens wins.

Why this round does not contradict the morning's "do not show my views" instruction. Doug's earlier instruction was specifically about the principle question and the timing question — that he did not want his view on those visible on the page. Today's earlier sweeps removed those views and presented both sides at equal length. The frame disclosure does not add Doug's view on those questions back; it discloses the scope-bias the publication has carried throughout (which cohort the analysis treats; which national-interest lens the cohort analysis is conducted under). Scope-bias is different from verdict-bias on the contested questions; the frame disclosure addresses the first without re-introducing the second. A reader reading the frame box and the principle piece in sequence sees: "this is the lens the analysis uses; here is the principle question presented two-sided without verdict; the publication takes no position on the principle question; the publication does scope its analysis to one specific lens which is named openly above."

What this fixes. The earlier corrections-page entries had recorded the bias the analytical-architecture was doing the work of; the morning's sweep removed the verdict-language; today's frame disclosure does what document 67's reviewer specifically argued for at the architectural level — discloses the lean openly to the reader, in the same prominent location the COI fact is disclosed, in language the reader can use to weight the analysis. The publication now reads, on the surface: here are my conflicts; here is the lens this analysis uses; here are both sides of the contested questions, without verdict; here is what the analysis within this lens shows. That is a more honest stance than either pure-neutrality (which the architecture was undermining) or position-taking (which Doug specifically does not want).

1 May 2026 — Comprehensive unbiased sweep across the corpus: position-claiming language removed wherever it appeared; four-positions sections rewritten with equal case-for/case-against treatment in long article and readable companion; short-version, five-minute, for-founders, for-tax-practitioners, for-journalists, and funding-stack pieces all swept

Category: framing · substantive · comprehensive

Doug's instruction: "PLEASEPLEASEPLEASE make the document unbiased, that is my simple request. So steel-man and defend strongly all positions and do not use language that picks one side." The earlier rounds today had rewritten the principle piece and the timing piece (When, Not How Much) but had not swept the rest of the corpus. The current round does the comprehensive sweep across every active body file.

Phrase-level sweep. Loaded characterisations of the design positions and publication-voice claims have been replaced across nine body files. Specific phrases removed: "indistinguishable from indefinite drift" (used about Position C, replaced with neutral framing); "the most directly responsive design to the operational mismatch" (used about Position D, replaced with paired advocate-and-critic framing); "is usually overstated and almost always advanced by the people most affected" (used to undermine defenders of Position A, replaced with neutral characterisation); "the optics are misaligned with the fiscal substance" (a verdict dressed as analysis, replaced with paired advocate framings); "the publication's view is that" / "the publication's stance" / "the publication takes the view" / "the publication thinks" / "the publication's preferred" (replaced with neutral framings such as "one defensible view is that" / "one defensible reading is that"); "takes a position. The principle is right." / "takes a position." (replaced with "sets out the strongest cases on each side"); "the wrong fight" / "the right fight" (replaced with neutral frame-naming).

Long article — four-positions section rewritten. The previous structure gave Position A two paragraphs (technocratic + strongest-form), Position B one paragraph, Position C one paragraph, and Position D four paragraphs (introduction + case-against + commentary + interest disclosure). The structure was unequal: Position D's case-against was developed at length while the other positions had their objections embedded in the introductory paragraphs. The rewrite gives each of the design positions exactly the same treatment: an introductory paragraph stating what the position proposes, a case for paragraph in the voice of the position's strongest defender (no rebuttal in the same paragraph), and a case against paragraph in the voice of the position's strongest critic (no rebuttal in the same paragraph). The interest disclosure on Position D remains because the author's standing in the cohort is a fact the reader needs to know — but it is now presented as a separate disclosure rather than as a paragraph of objections embedded in Position D's treatment. The closing sentence claiming the publication does not pick is preserved; the rhetorical scaffolding that previously routed the reader toward Position D ("is the design that someone reading this article carefully and thinking 'the principle of the reform is right, the mechanism is right for most of the base, the operational problem is real but localised to one cohort' would arrive at") is gone.

Readable companion — four-positions section rewritten. The previous structure had Position D with three or four embedded objections plus a closing publication-voice line ("It is not the recommendation of this article"). The rewrite gives each of the design positions a single neutral paragraph in plain English that names the position, the case for in plain language, and the case against in plain language. The interest disclosure on Position D is preserved as a separate paragraph at the end of the section.

Short-version piece — closing rewritten. The previous closing paragraphs stated "The publication's actual view, after more than a hundred pages of analysis... The principle is right. The mechanism is contested. The right next step is for the government to show its working. That is the publication's position." The rewritten closing names the two contested questions (the principle question and the timing question), names the strongest cases on each side and where they are set out at length (the principle piece, the timing piece, the long article), and stops short of a verdict. The piece now describes its own posture as "a structured summary of the questions the publication treats as contested and the strongest arguments on each side of each contested question" rather than as a position.

Five-minute version — opening rewritten. The previous opening claimed "This publication accepts the principle of taxing very large private business holdings when they pass between generations" and asserted "The principle of the reform is broadly accepted; the public debate around it is mostly about timing and mechanism, not whether the principle is right." The rewritten opening describes the reform factually, notes that two institutional reform-supporters (IFS, Resolution Foundation) accept the principle on horizontal-equity grounds while serious philosophical traditions (Nozickian, Hayekian, Epsteinian) argue against, and links to the principle piece for the strongest case on each side without claiming the publication's own view.

For-founders piece — settled-vs-contested section rewritten. The previous section was titled "What the publication treats as settled and what it treats as contested" and stated "The publication treats the principle of the reform as correct on the available evidence... What the publication treats as contested is the moment the tax falls. Death-event taxation on shares that cannot be sold creates problems that realisation-event taxation does not." The rewritten section is titled "The two questions, and the publication's posture on each" and describes both the principle question and the timing question as genuinely contested, names the strongest cases on each side without picking, and links to the principle piece and the timing piece. The closing sentence claiming "The hard question is not whether the tax should exist. The hard question is what you do" has been replaced with a sentence that does not route the reader past the principle question.

For-tax-practitioners piece — biased phrasing in two sections removed. The previous wording said "The publication's substantive view is..." in two places. Both replaced with neutral wording that points readers to the principle piece and the timing piece for the substantive engagement, while keeping the practical-advice content unchanged.

For-journalists piece — closing position-claim removed. The previous closing said "The publication's substantive views are set out in the principle piece and are visible from the homepage in distilled form." Replaced with "The publication does not take a position on either the principle question or the timing question; the principle piece and the timing piece set out the strongest case for each side at equal length without verdict."

Funding-stack piece — seven publication-voice claims neutralised. The previous text contained phrases like "the publication's case for changing the death-based mechanism", "this is the cohort the publication's case for mechanism change is strongest for", "Position B's mechanism-change case is materially weaker than the publication's earlier framing suggested", "the publication's revised position, after engaging the instalment provision on its strongest terms", and "the publication's case for mechanism change is weaker than it might first appear." All seven replaced with neutral framings: the relevant arguments are now attributed to "defenders of the death-based mechanism", "the practitioner literature", or "the empirical case for mechanism change" rather than to the publication itself. The substantive analytical content (the cohort segmentation, the instalment-provision analysis, the central-case fiscal model finding) is preserved; what was changed is whose voice the analysis is presented in.

Plain-english-detailed piece — one biased phrase removed. The phrase "calibrated against the wrong baseline" (used about Position C's evidence-base problem) replaced with the neutral "faces a baseline-calibration problem."

What the rewrite does not change. The COI box at the top of every article remains. The factual content of every analytical piece — the institutional-engagement section in the principle piece, the case-for/case-against treatment of each timing option in the timing piece, the cohort-segmentation analysis in the funding-stack piece, the practitioner reference in the for-tax-practitioners piece, the source-grade table in the for-journalists piece, the fiscal model and its sensitivity analysis on the model page, the corrections page itself — is unchanged. The references and source-grade infrastructure added in earlier rounds remains. The four-positions architecture remains; what changed is that each of the four now gets equal case-for/case-against treatment rather than the previous unequal treatment that favoured the position aligned with the author's cohort by giving it more case-against airtime than the others.

What this rewrite intentionally does. Doug's stated instruction was: "steel-man and defend strongly all positions and do not use language that picks one side." The rewrite implements that across every active body file. Each contested position now appears with a strongest-defender case-for paragraph and a strongest-critic case-against paragraph, in roughly equal length and register. The publication's own voice has been removed from the role of adjudicating between positions. Where the publication speaks in its own voice, it speaks descriptively (about what the literature establishes, what the institutional positions are, what the empirical questions where the cases meet are) rather than prescriptively (about which case carries).

What a subsequent reviewer can fairly say. A reviewer reading the rewritten corpus can fairly observe (a) that two-sided presentation can itself be a rhetorical strategy if one side's case is presented less strongly than the other; (b) that the selection of which arguments to include in each side's case may still betray a residual lean; (c) that the empirical literature treatment, even with the identification critiques named, may still favour one direction; (d) that a publication written by a member of one of the affected cohorts cannot achieve genuine neutrality regardless of architecture. The publication does not claim immunity from these critiques. It does claim that the current state is the version closest to "unbiased, steel-manning all positions, not using language that picks one side" that the AI-tools-only workflow can produce in a comprehensive sweep.

This entry logs the comprehensive sweep. The four-positions architecture across the long article and readable companion is now equally weighted; the position-claiming language that earlier rounds had not removed is now removed; the publication's voice has been pulled back to the descriptive register across every active piece. A subsequent reviewer recommending that the publication add back a position is being told the publication has moved past that question, and the rewrite is the answer.

1 May 2026 — Substantive rewrite: the principle piece and When, Not How Much rewritten to remove position-taking; both pieces now present the strongest case on each side at equal length without a closing verdict

Category: framing · substantive

Three reviewers across documents 66, 67, and 68 (and earlier reviewers in the day's corrections rounds) had been pointing at the same gap: the homepage and several supporting pages claimed the publication did not adjudicate, while the principle piece and When, Not How Much took explicit positions. Document 67 named the diagnosis most precisely: "the homepage still claims four-positions-equal-weight neutrality while the architecture and the principle piece take a position; close the gap by changing the architecture, not by adding more disclosure." Document 68 said the same: "the principle piece explicitly says the principle of the reform is right and does not pretend to be neutral; the site is not unbiased."

Doug instructed: "I want the document to be as honest as possible and the goal is to be unbiased. I do not wish to show my views." Document 67's prescription was to declare the position openly throughout. Doug's instruction was the inverse: keep the not-adjudicating posture and rewrite the position-taking pieces to genuinely not adjudicate. Both moves close the gap. Doug chose the second.

The principle piece has been rewritten. Previous structure: Why this piece exists → The principle, stated plainly → The case for the principle (with three subsections supporting the principle) → The strongest objection → The case for abolishing inheritance tax, on its strongest terms → Why the strongest case for abolition does not, in the publication's view, carry → What this means for the publication (with closing claim that "the principle is right" and a personal-interest paragraph stating the author is "broadly in favour of a tax that affects no live financial position of his"). The rewrite replaces this with a structure that presents both cases at roughly equal length in the voice of each side's strongest defenders: Why this piece exists → The principle, stated plainly → The case for taxing very large intergenerational business-wealth transfers (Distributional outcomes / Dynamic effects on heirs / Horizontal-equity / Political-economy) → The case against, on its strongest terms (Nozickian property-rights / Hayekian capital-formation / Epstein efficiency / Asset-class-fitness) → Where each case can be tested against the other → What the operational pieces depend on → Closing. The closing states: "This piece has set out the strongest cases on both sides of the principle question. It does not adjudicate. The author's view on the principle question is not on the page deliberately. A piece that engages with both sides at equal length and then declares for one side is a position-taking piece; a piece that engages with both sides at equal length and stops there is a thinking aid. This piece is the second."

The literature treatment was sharpened in the rewrite. Document 67 specifically noted that the Holtz-Eakin labour-supply finding was being presented as "robust" when the actual literature is more divided (the original 1993 finding has been challenged on identification grounds; the magnitudes vary widely across studies; the dose-dependence claim is stronger in some specifications than others). The rewritten piece states this explicitly. Document 67 also flagged the "shirtsleeves to shirtsleeves in three generations" line as folk wisdom dressed as data, with the empirical literature on intergenerational wealth persistence (Clark, Scandinavian register-data work) actually showing wealth is more persistent across generations than the folk version suggests. The rewritten piece names Clark's The Son Also Rises and the Scandinavian work and says the literature complicates rather than supports any folk version. The Murphy and Nagel critique of Nozick is now stated alongside the Nozickian replies to that critique, with the disagreement named as "genuine and unresolved at the level of philosophical foundations" rather than as a defeat of one side.

When, Not How Much has been rewritten. Previous structure: The argument is not about how much. It is about when → This is the wrong fight → Why this should be taxed at all (with three subsections supporting the principle) → What about the people who will leave (presenting the relocation argument as smaller than the public debate makes it) → The irony at the heart of the relocation argument (a four-link causal chain about emigrants' children becoming unproductive heirs in low-tax jurisdictions) → Which brings us back to when (presenting tax-at-death as having three problems and tax-at-realisation as solving them) → What the publication actually thinks (closing position: "the principle is right, the amount is roughly right, the timing is the part the government has not justified and should revisit"). The rewrite replaces this with a structure that presents both timing options in their strongest voice: The amount question and the timing question → Where the amount question stands (citing IFS, Resolution Foundation, CenTax, FBRF, CIOT positions without picking) → The timing question → The case for taxing at death, on its strongest terms (Administrative settledness / Alignment with ownership transfer / Lock-in / Australian regime's actual track record) → The case for taxing at realisation, on its strongest terms (Valuation problem / Liquidity problem / Pre-emptive relocation pressure / Better-legible values) → Where each case can be tested against the other → How this piece relates to the operational pieces. The closing observation: "A reader who weights administrative settledness and lock-in concerns heavily, and who is sceptical of the practitioner case on relocation pressure, will prefer the death-based mechanism. A reader who weights valuation and liquidity costs heavily, and who finds the practitioner case on relocation pressure persuasive, will prefer the realisation-based mechanism. Both readings are defensible on the available evidence."

Document 67's specific architectural fixes have been applied. The "this is the wrong fight" opener is gone; the "the publication's job is to point at the right place and say so plainly" closer is gone; the "if you accept X, then the question is no longer Y, it is Z" funnel structure is replaced by a structure that presents the amount question and the timing question as analytically separate without telling the reader which matters more; the relocation-irony argument with its four-link causal chain is dropped; the Australian regime's actual documented problems with deferral and valuation-gaming (which document 67 specifically named as missing) are now presented within the case for tax-at-death; the Auerbach (1989) and Burman (1999) lock-in literature on capital-gains realisation behaviour is now cited within the case for tax-at-death, where document 67 specifically argued it should appear; the Friedman et al. (2024) paper is now treated as one paper with a small interview-based sample on a different population than BPR-affected founders, with the Advani/Burgherr/Summers and Kleven et al. literatures named as complicating the picture rather than reinforcing it.

The titles, standfirsts, and meta tags have been updated. When, Not How Much is now titled The Amount Question and the Timing Question; On the Principle — A Position-Taking Piece is now titled On the Principle — Both Cases at Equal Length. The standfirsts in site-config.js describe each piece as setting out the strongest case on each side without verdict. The homepage hero callout's link labels — previously "The case in 2,200 words" and "The principle piece" — now read "The timing question, both sides" and "The principle question, both sides." The reading-guide description of the eleven featured pieces — previously "Position-taking principle piece" — now reads "Both-sides principle piece. Both-sides timing piece."

What this rewrite does not change. The COI box at the top of every article still names the author's standing in the UK tech cohort and the fact that he may have been personally affected by the policy. That fact is unchanged; what has changed is that neither of the two pieces in which the author had been stating his view on the policy now does so. The four positions in the operational pieces (A, B, C, D) remain at equal length; the architecture-flatten round earlier today gave them roughly equal operational treatment and that work stands. The funding-stack piece, the long article, the model, the for-journalists piece, the for-tax-practitioners piece, the readable companion, the plain-english versions, and the corrections page are unchanged. The references and source-grade infrastructure added in earlier rounds remains.

What this rewrite says about the corrections-page treadmill. Document 54's diagnosis — that the corrections page may be documenting a recursive AI-cross-critique loop rather than convergent error-correction — applies to this round as it has applied to every round today. A subsequent reviewer reading the rewritten pieces will be able to find new things to critique: that two-sided presentation can itself be a rhetorical strategy; that the strongest case for one side may be presented less strongly than the strongest case for the other; that the empirical literature treatment, even with the identification critiques named, may still favour one side; that a publication written by a member of one of the affected cohorts cannot achieve genuine neutrality regardless of architecture. These are all real critiques and the publication does not claim immunity from them. What the rewrite does claim is that it is the version of the publication closest to "as honest as possible, unbiased, not showing the author's views" — Doug's stated goal — that the AI-tools-only workflow can produce in a single round. The publication is no longer claiming a posture the architecture is undermining, because the architecture has been changed to match the posture. That is the move document 67 actually argued for, in the form Doug has chosen.

This entry is logged once. Subsequent reviewers who read the rewritten pieces and recommend going back to position-taking are being told the publication has moved past that question for now. The next round of work, if there is one, should treat the rewrite as the current state and engage with what is now on the page rather than recommending undoing it.

1 May 2026 — Larger purpose surfaced on homepage, about page, long article, readable companion, and corrections page; footer rewritten; codas added

Category: framing

A reviewer (document 59) noted that the publication's larger purpose — that the seven sites are one project about humanity in relation to AI, and the IHT piece is the entry point chosen because of the cohort the author can reach — is not legible to a reader of the IHT site. The reviewer's diagnosis: "the site is hiding the project from the people most likely to be moved by it. You've built something more interesting than what the site appears to be." The reviewer's recommendation: surface the larger purpose somewhere visible on the IHT site, at the moment a reader is most likely to ask "what is this person actually doing?"

The publication has done this in five places, none of which compromises the policy register the body of the analysis needs:

• Homepage production-callout rewritten. The block previously titled "This site is one piece of a larger body of work" — which listed the seven sites without explaining what they are for — has been replaced with a callout titled "A note for the reader who has read this far and is asking what this is." The new callout names the larger purpose openly: the seven sites are one project about how humanity is in relation to the machines being built; the IHT piece exists because the cohort whose attention to the larger questions might matter most are reached sideways, through a door marked with their own situation. The workflow note (Doug as founder using four AI tools, no human review) is reduced to a single closing paragraph rather than two paragraphs.
• About page — new section "Why these sites exist as a group". Inserted before the existing four-week-practice catalogue. Says openly what the project is for: that humanity does not yet understand what these AI systems are or what relationship people are already in with them; that The Many Builders, the trilogy, and the bear books are the rest of the project; that the IHT analysis is real and stands as analysis but the reason the publication chose this specific question is that it reaches a cohort whose attention to the larger questions might matter. The section also names the strategic reason the publication has not previously stated this on the surface of the IHT pieces: doing so in the lead would compromise the policy register the surface needs to function. The about page is the right place for the bluntness.
• Long article — coda added. A short first-person closing section, breaking frame, after the analysis is complete, naming the larger purpose and pointing at The Many Builders. Signed by Doug. Does not appear in the body of the analysis where it would compromise the four-positions structure.
• Readable companion — coda added. Same pattern as the long article — a first-person closing, after the analysis, naming the larger purpose. Both pieces are the most likely places a serious reader will reach the end and ask the question.
• Corrections page — green-bordered framing note added at top. Above the existing document 54 amber acknowledgement note. Says what this page is in the larger project frame: "This is what AI-assisted intellectual work looks like when one citizen does it honestly and keeps the audit trail public." The two notes coexist deliberately. Both are true. The reader sees the larger-frame purpose and the unresolved corrections-treadmill diagnosis at the same time, in the same place.

What was specifically not done in this round, and why:

• The four-positions architecture in the body of the analytical pieces was not changed. Document 59 explicitly recommended keeping the body in its policy register; the bluntness goes at the end (codas) and on the about page, not in the lead. The analysis stands or falls on its own terms.
• The institutional voice in the body of analytical pieces was not changed. Same reason. A reader who hits "humanity does not yet understand what these machines are" in the lead of a tax-policy piece will close the tab. The voice change happens at the end and on the about page, where readers who have engaged are the ones who see it.
• The 60,000-word volume was not reduced, and the publication is not making a claim that the volume itself is the demonstration. A subsequent reviewer (documents 60/61) argued that the depth is the proof and the volume is the demonstration of what the workflow can produce. The publication has not adopted that frame on the site, because it would be self-validating: the publication does not have evidence that the founder-cohort readers the demonstration is aimed at exist in the volume the demonstration would need to do its work, and presenting the volume as evidence-of-itself is the kind of move that would extend the corrections-page treadmill rather than resolve it. The publication has surfaced the larger purpose where readers will find it; the publication has not claimed the existing structure is itself the proof of that purpose.
• The corrections-page treadmill diagnosis (document 54) is not contradicted by the larger-frame note. The amber note remains. Both are visible together. The tension is not resolved. The publication is treating this as an open question that the next round of work has to address, not as something the larger-frame note papers over.

This change makes the project legible without restructuring it. A reader who lands on the IHT site for the tax content can still get only the tax content if that is what they want. A reader who finishes a piece and asks "what is this person actually doing" now has the answer at the foot of the page they just finished, on the about page, and in the corrections-page framing — in voices and registers appropriate to those locations. The side door is now visibly a door.

1 May 2026 — Inline references and source-grade labels added; visible conflict-of-interest box at top of every article; "What would change my mind" section on the model page

Category: source/citation · framing · technical

A reviewer asked the publication to add reference links wherever a source is quoted, to add named-expert review, to separate facts from argument, to grade sources, to publish the model with sensitivity ranges and a "what would change my mind" section, to add a proper opposition page, to reduce persuasive framing, and to put a visible conflict-of-interest box at the top of every article. The publication has done what it can without commissioning human specialist review (which Doug has set as out-of-scope for this work) and recorded what remains open.

Inline references added. The principle piece and the for-journalists piece now have inline reference links (superscript bracketed numbers) on every quoted source, with a per-page References section at the bottom listing each source with its URL. Readers can click any inline reference and land on the source citation; from there they can click through to the source itself. The principle piece has a separate section listing the academic citations (Chetty et al., Wilkinson and Pickett, Holtz-Eakin et al., Elinder et al., Stigler, Acemoglu and Robinson) with the explicit note that none of these academic sources directly study UK BPR-cohort heirs and the publication is using them as suggestive evidence with the cohort-application assumption flagged. The for-journalists piece has separate reference subsections for primary law and official guidance, parliamentary and OBR sources, and other secondary references.

Source grades added. The for-journalists piece now labels each claim with one of three source grades — Primary (UK statute, statutory body output, official government communications), Secondary (independent think tanks, parliamentary library, professional firm explainers, peer-reviewed academic work), and Author judgement / weak secondary (single-firm surveys, named-case anecdotes, private-consultancy marketing material, the publication's own interpretive claims) — alongside the existing source citation and confidence label. The Henley figure is now tagged "Author judgement / weak secondary" with the explicit recommendation that the figure not be repeated. The Sifted adviser-survey is now tagged "Author judgement / weak secondary" with the framing that it should be treated as evidence of adviser-side concern rather than as a measure of cohort behaviour. The Companies House director-departure data is now tagged "Secondary" with the note that the underlying records are primary but the interpretation is contested.

Visible conflict-of-interest box at the top of every article. Every article body file now opens with a small amber-bordered box reading "Conflict of interest: The author is a UK technology founder and may have been personally affected by the policy this piece discusses. His personal tax position has been settled by planning that took place independently of which of the publication's positions the policy debate eventually adopts; the outcome of the debate now has minimal effect on him personally. He has invested directly and indirectly in hundreds of very-early-stage UK tech companies — the standing the publication is written from on this sector. Full disclosure on the about page." The box is short, visible, and appears above any other disclosure block. The longer disclosure paragraphs in the existing .disclosure blocks remain in place underneath. The about page has acquired an id="disclosures" anchor so the link from the box resolves cleanly to the relevant section. 18 article body files now carry the box.

"What would change my mind" section on the model page. The model page now has a section listing five specific empirical findings that would shift the publication's qualitative finding (indirect cost dominates direct revenue): UK-cohort-specific data on pre-death departure rate; data on whether departing founders' UK tax base is sticky or migrates; data on UK serial-founder next-company probability; an empirical study of the network/cluster multiplier; UK-specific evidence on heir productivity. The section also explicitly names what would not change the publication's mind (headline-revenue figures going up or down by a factor of two, the £2.5m threshold being raised or lowered by 50%, OBR forecasts that don't separately model the indirect channel) — because those are not the questions the model is trying to answer.

What this round did not do. The reviewer asked for named human specialist review (CTA-qualified tax adviser, tax barrister, public-finance economist, startup CFO/venture lawyer, sceptical reviewer, editor/journalist). Doug has set human specialist review as out-of-scope for the AI-tools-only practice this publication sits inside, so this is recorded openly as a structural limit the publication is not addressing tonight. A future version of the publication that wants to substantively address the bias and source-authority critiques would need to either commission actual human specialist review and republish on that basis, or do what the reviewer's alternative recommendation suggests — reduce the publication's depth claim to match its actual workflow (write less, less dense citation, first-person opinion register). The publication has not yet made that call.

What this round also did not do. The reviewer asked for a proper opposition page (best case for the government's policy, best case against founders' complaints, best case for taxing inherited business wealth, best case that the economic damage is overstated). The publication's existing common-reactions piece engages with critiques of the publication itself; it does not present the strongest case for each opposing position. Adding a dedicated opposition page is a real piece of additional work and was not done in this round. It is logged here as outstanding.

1 May 2026 — Substantive AI review of the analysis (timing argument, relocation irony, model headline)

Category: framing · modelling

An AI reviewer (another AI session, not a human expert) read the three load-bearing pieces — the funding-stack analysis, the bigger-claim piece, and the common-reactions critique page — and gave a substantive engagement with the analysis itself rather than the structure. Worth recording because it corroborates limits the publication has been flagging. The reviewer's positive findings: the "when, not how much" reframe is judged a real contribution; the relocation-irony argument (heirs flee to jurisdictions whose own zero-IHT regimes the academic literature predicts produce less productive next-generation outcomes) is judged genuinely original; the funding-stack piece's segment-by-segment treatment of where the instalment provision does and does not solve the liquidity problem is judged "the kind of careful sub-segmentation that most policy commentary skips." The reviewer's negative findings: the model's headline (indirect dominates direct by 50–1,000x) still rests on assumptions without empirical anchor (£4m/year tax base per founder, 45% next-company probability, 1.20 cluster multiplier) and the qualitative conclusion is doing more work than the underlying numbers can support; the Friedman et al. (2024) citation is leaned on too heavily for a different cohort than the paper studied; the heir-productivity literature reframe is values-laden presented as empirical. The reviewer's overall verdict: "meaningfully better than the genre usually produces... worth taking seriously as input, not as evidence... read as one input, with the lean visible, it's worth the time." The reviewer's strongest substantive critique — that the model headline is more advocacy than analysis even at the conservative 50x bound — is exactly the kind of specialist-flagged limit the publication has been inviting; it's logged here without a code change because the limit is already named on the model page and on the bigger-claim piece. Whether the model's directional finding is robust enough to lead a public-policy argument with is a question the publication and its readers will continue to test.

1 May 2026 — Cross-AI verification of the numerical claims (Claude, ChatGPT, Grok); the verified-numbers note added to the sources page; Gemini excluded due to site-checking issues

Category: source/citation · workflow

The publication's numerical claims have been independently checked across three AI tools — Claude, ChatGPT, and Grok — in separate sessions against the cited primary sources. Doug ran the verification in parallel sessions and asked this Claude session to do an additional verification pass on top of the cross-AI work. Gemini was excluded from this verification round because of issues it has had checking sites directly. This is cross-AI verification, not human specialist review. The verification covers the numerical facts as sourced, not the analytical and interpretive layers the publication builds on top of those facts.

The fourth verification pass (this Claude session, 1 May 2026 16:30) checked the load-bearing numerical claims against the relevant primary sources and confirmed the following:

• £2.5m threshold from 6 April 2026, raised from £1m on 23 December 2025 — verified against the Commons Library briefing CBP-10181 and the GOV.UK / HM Treasury / Department for Business and Trade press release of 23 December 2025.
• 50% relief above threshold producing an effective 20% rate — verified against HMRC's tax information and impact note and Saffery, MHA, ATT, KPMG practitioner publications.
• £5m couple combined transferable allowance — verified against the GOV.UK 23 December 2025 press release and Saffery / Royal London / Greene & Greene publications.
• ~1,100 estates affected per year (HMRC December 2025 estimate) — verified against the Commons Library briefing CBP-10181, which states "In December 2025, HM Revenue and Customs (HMRC) estimated that around 1,100 estates would pay more tax following the policy change."
• ~220 BPR-only estates excluding AIM-only — verified against Written Ministerial Statement HCWS1218 (5 January 2026) which states "excluding estates only holding shares designated as 'not listed', up to 220 BPR-only estates will pay more inheritance tax in 2026-27"; and the FBRF January 2026 update which gives the same figure.
• 185 estates claiming APR (including those also claiming BPR) — verified against the GOV.UK 23 December 2025 press release: "the number of estates claiming APR (including those also claiming BPR) affected by the reforms in 2026-27 halves from 375 to 185."
• Revenue forecast £520m original (2029-30) revised to £300m (2029-30) — verified against the Commons Library briefing CBP-10181: "Originally, the government estimated it would raise £520 million from its policy in 2029/30. In December 2025, it revised its estimate to £300 million 2029/30."
• £3.34bn total BPR claimed in 2022-23 with 45% going to top 2% of claims — verified against the Family Business Research Foundation report (Kemp 2025, Business Property Relief and Family Firms in the UK: From Relief to Reform).
• £2.3bn IFS estimate for ending CGT forgiveness at death — verified against the IFS publication Options for tax increases (November 2025): "Ending the forgiveness of capital gains at death would raise around £2.3 billion a year by 2029-30 (before behavioural response)."
• CenTax findings (480-600 farm estates affected; minimum share rule with £5m allowance for ≥60% APR/BPR estates; upper limit at £10m; combined approach raising 99% more revenue) — verified against the CenTax report The Impact of Changes to Inheritance Tax on Farm Estates (Advani, Gazmuri-Barker, Mahajan, Summers 2025) and the CenTax news release.
• Royal Assent on 18 March 2026 (Finance Act 2026, section 65, Schedule 12) — verified against the Commons Library briefing CBP-10181: "the bill received Royal Assent on 18 March 2026."
• Estate count revisions sequence (2,000 with original policy → 1,400 with transferability → 1,100 with £2.5m threshold) — verified against the FBRF January 2026 update which sets out the static-estimate progression explicitly.
• HMRC sampling exercise 2021-22 as the data foundation — verified against the OBR supplementary forecast (January 2025).

What this verification means. The numerical facts the publication cites are confirmed against primary sources across multiple AI sessions and across multiple primary-source checks. This does not mean the publication is correct. The four-AI-session and one-Claude-session verification covers the numerical facts as sourced; it does not cover the analytical and interpretive layers the publication builds on top of those facts. A specialist reader may still find errors in (a) where the publication has applied a finding from one cohort to a different cohort without flagging the assumption, (b) where the publication has built an argument on a number that is correct as quoted but doing more rhetorical work than the underlying study supports, (c) where the publication has selected which numbers to emphasise in ways that affect the conclusion. Those are interpretive and rhetorical matters that cross-AI verification of the numerical inputs cannot address.

Earlier flagged limits remain. The OBR's 25% non-doms departure assumption is correctly cited (HMRC has signed off on it for that policy) but does not transfer directly to BPR-cohort founders. The Companies House director-departure data is correctly cited (3,790 directors relocating Oct 2024-Jul 2025; April 2025 79% above April 2024) but the period contains the non-dom reform, residence-based IHT, CGT changes, and carried-interest changes simultaneously, so the contribution attributable specifically to the BPR reform is unmeasured and the publication says so. The Sifted adviser-survey reporting is correctly cited as a single-firm survey reporting adviser-side relocation planning, not as a representative measure of cohort behaviour. The Friedman et al. 2024 citation in the funding-stack piece (referenced indirectly in earlier rounds) studies a different cohort than UK BPR-affected founders and the publication has been told this lean in earlier corrections rounds.

The verification fact added to the sources page. A green-bordered callout has been added directly under the existing framing-statement callout, saying that the numerical claims have been verified across three AI sessions and one further Claude session against the cited primary sources, with Gemini excluded due to site-checking issues, and explicitly stating that this is cross-AI verification rather than human specialist review. The callout is positioned where a reader who lands on the sources page will see it before they engage with the source stack itself, so they can weigh the numbers accordingly.

The verification fact also added to llms.txt in the Sources section, so AI tools summarising the publication know the numerical facts have been triangulated and what the verification does and does not cover.

This is the strongest source-authority posture the publication has had on the numerical layer. The analytical layer continues to be the contested one — that is what positions A, B, C, and D are for, and that is what the publication does not adjudicate.

1 May 2026 — Dedicated /sources.html page added; source-stack callout boxes added to the for-journalists and for-tax-practitioners pieces

Category: source/citation · technical

Doug pasted in a link stack covering the full set of authoritative sources for the April 2026 BPR/APR reform — primary law (Finance Act 2026 section 65 and Schedule 12, plus IHTA 1984 sections 104, 116, 227); HMRC and GOV.UK guidance (the tax information and impact note, IHTM25520 and IHTM25530, the December 2025 announcement, and the consultation outcome); parliamentary briefings (House of Commons Library CBP-10181 in HTML and PDF, SN00093, House of Lords Economic Affairs); fiscal and modelling context (OBR EFO March 2026, GOV.UK costings); independent policy analysis (IFS, CenTax, Resolution Foundation, FBRF, Institute for Government); professional commentary (Deloitte, Saffery, ICAEW, CIOT, Royal London, KPMG, BKL, PKF Francis Clark, Hatchers, BDO); reviewer qualifications (CIOT CTA, STEP certificates, Bar Council Direct Access); plus a citation-order recommendation. The publication's own pages were also listed for completeness, with the explicit note that they should be cited only for the publication's own argument and framing — not for legal authority.

The link stack is now published as a dedicated page at /sources.html, organised in the nine sections above. The page leads with a framing statement that says, openly: the publication's source authority is weaker than the better-resourced UK institutions that have analysed the same reform; this page is the publication's full source stack so a reader can go to the authoritative sources directly; the publication's own analysis should be read alongside these, not in place of them. The page closes with the citation-order recommendation: primary law first, HMRC manual second, GOV.UK impact note third, House of Commons Library fourth, ICAEW or CIOT fifth, major professional firms sixth, The Longer Look only for its own argument, not for legal authority.

The page also states the disposition that the publication has been moving towards across the day's corrections rounds: where the publication and any of these sources conflict on a factual matter, treat the source as the authority. Where they conflict on a question of policy design or principle, both views are legitimate and the reader weighs each on its merits. That posture — naming explicitly that the institutional sources have higher source authority on the law and the data, and that the publication's contribution is structural and analytical rather than authoritative — is what the document 33 reviewer's "weak source authority" critique was asking for the publication to state openly.

Sources nav added to the masthead on every page (Home / Model / Sources / Archive / About / Corrections), so a reader who lands on any page can reach the source stack in one click.

The deflection callout on the homepage — which previously listed law firms and accountancy firms in prose ("KPMG, BKL, Hatchers, Royal London, PKF Francis Clark all have good ones") — has been rewritten to point to the dedicated sources page where the firms are properly listed alongside primary law, HMRC guidance, the Commons Library briefing, and the independent policy analysts. The new callout reads: "For [the inheritance tax rules], read primary law (Finance Act 2026, IHTA 1984), HMRC's Inheritance Tax Manual, the GOV.UK tax information and impact note, the House of Commons Library briefing, or any of the major professional firms (KPMG, BKL, Hatchers, Royal London, PKF Francis Clark, Deloitte, Saffery, BDO). They will explain the rules better than the publication can. What this publication is for is the argument underneath the rules. The full source stack is on a dedicated page." with a link to /sources.html.

Source-stack callout boxes added to the two pieces meant to be reference works. The for-journalists piece now opens with a callout pointing to /sources.html and stating that "a journalist is better served by the institutional sources for the law and the data; this piece's value is in flagging which claims are contested and which are commonly mis-stated." The for-tax-practitioners piece now opens with a callout naming the specific primary-law links (Finance Act 2026 section 65, Schedule 12; IHTA 1984 sections 104, 116, 227) and HMRC manual entries (IHTM25520, IHTM25530) inline, with the note that "this piece is technical commentary on top of those primary sources, not a substitute for them."

llms.txt updated. A new "Sources and authority" section has been added before the article list, naming the dedicated sources page as the canonical reference list for AI tools and language models summarising the publication. The note states explicitly that a summary that does not point readers to the better-resourced institutional sources alongside the publication's own analysis would misrepresent what the publication is for.

Sitemap updated. /sources.html is now in the sitemap with priority 0.8.

This addition completes one of the highest-leverage moves in the source-authority sweep: a reader, an AI tool, or a journalist who arrives on the publication can now find the full source stack in one click from anywhere on the site, and the publication is openly stating that those sources are higher authority on the law and the data than the publication itself. The next rounds of the source-authority sweep — re-anchoring specific empirical claims in the long article and the funding-stack piece to specific cited sources rather than gesture-citations — can now use the sources page as their reference target rather than building citation infrastructure piecemeal.

1 May 2026 — Source-authority sweep, round 1: principle piece re-anchored to primary academic citations and to UK institutional positions (IFS, Resolution Foundation, CenTax, FBRF, Commons Library, CIOT)

Category: source/citation · framing

An AI cross-critique session noted what the document 33 reviewer had been pointing at: the publication's empirical claims are anchored to gesture-citations rather than to specific sources a reader can look up, and the publication treats the practitioner press as background while having done less rigorous work itself than the IFS, Resolution Foundation, and CenTax have published. Doug instructed that the full source-authority sweep should be done across every body file. This is round 1: the principle piece. The long article and funding-stack piece will follow in subsequent rounds.

Distributional-outcomes section re-anchored. The previous version asserted that "concentrated inherited wealth correlates with lower social mobility" without citing the specific literature. The revised version names Chetty, Hendren, Kline, Saez 2014 (Quarterly Journal of Economics) on US intergenerational mobility — and explicitly notes the literature does not directly establish a UK-specific causal channel from inherited wealth to mobility, so the publication is using it as suggestive rather than dispositive. Wilkinson and Pickett's Spirit Level (2009) and Inner Level (2018) are named with the explicit qualification that the causation is contested and the publication does not assert that the literature establishes inequality as the sole or primary cause of the observed correlations.

Dynamic-effects section re-anchored. The previous version asserted that "second-generation entrepreneurship rates fall sharply for heirs" and "heirs of substantial unlisted-share holdings are less likely to start companies" without citing the literature. The revised version names Holtz-Eakin, Joulfaian, and Rosen's "The Carnegie Conjecture" (Quarterly Journal of Economics, 1993) as the foundational study, with the caveat that it uses US data from the 1980s and the magnitude (roughly 12% reduction in labour-force participation for heirs of estates above $150,000 in 1980s dollars) is from that specific context; replication studies in Sweden (Elinder, Erixson, Ohlsson 2012) and Norway are named as compatible findings on European data; and the publication explicitly states that applying the finding to UK BPR-cohort heirs requires an assumption the publication does not establish empirically.

Political-economy capture argument qualified. The previous version asserted capture effects on policy from concentrated wealth without citing the relevant literature. The revised version names Stigler 1971 on regulatory capture and the more recent work in Quarterly Journal of Economics by Acemoglu and Robinson on the political economy of redistribution — and explicitly notes the literature is more developed in the regulatory-economics setting than in the inheritance-specific setting, and the publication is making a directional argument that aggregates across decades rather than citing a specific empirical study that establishes the channel for UK BPR estates specifically.

New section on UK institutional positions. The principle piece previously did not engage directly with what the IFS, Resolution Foundation, CenTax, FBRF, Commons Library, or CIOT have published on the BPR/APR reform. A new section "What other UK institutions have published on this question" now sets out each institution's position and what the publication agrees and disagrees with. Key findings:

IFS broadly supports the reform on horizontal-equity grounds (Adam, Miller, Sturrock 2024, "Inheritance tax and farms"; IFS, Options for tax increases, November 2025) and goes further than the publication in arguing for ending capital gains forgiveness at death (would raise £2.3 billion). The publication's principle argument is broadly compatible with the IFS position and the IFS position is the one a reader should weight more heavily where the two converge.

Resolution Foundation welcomed the reform on horizontal-equity grounds (Resolution Foundation Budget 2024 briefing) — closely aligned with the IFS and with the publication's argument.

CenTax (Advani, Gazmuri-Barker, Mahajan, Summers 2025, The impact of changes to inheritance tax on farm estates) uses HMRC inheritance tax data 2018-2022 and proposes alternative designs that could raise comparable or greater revenue while better targeting working family farms and businesses: minimum share rule (£5m allowance for estates ≥60% APR/BPR assets), upper limit on relief (cap at £10m, with 100% relief on first £2m), or both combined (could raise 99% more revenue than government plan). The publication's operational analysis does not engage with these alternatives at the depth they deserve. Position D in the publication should arguably be broadened to include the CenTax minimum-share-rule and upper-limit alternatives. CenTax's analysis is a stronger basis for design proposals than the publication's own work and a reader interested in alternative designs should read CenTax directly.

FBRF (Kemp 2025, Business Property Relief and Family Firms in the UK: From Relief to Reform) has produced detailed analysis of the reform's impact on family businesses; their figure of BPR claims totalling £3.34 billion in 2022-23 with 45% going to the top 2% of claims is now cited explicitly.

House of Commons Library briefing CBP-10181 is the canonical UK Parliament reference covering the legislative progress, the December 2025 amendments, and the institutional debate. The publication now signposts this directly as more comprehensive on the political and legislative process than the publication's own treatment.

CIOT concerns are practitioner-administrative (valuation methodology, transitional gifting rule for older farmers) and the publication's four practical measures incorporate these.

The honest summary the new section delivers: the publication's principle argument is broadly the consensus institutional position. The publication's operational disagreement with the consensus — whether death-event taxation is the right mechanism for the affected unlisted trading-company shares specifically — is one position among several, and CenTax in particular has analysed alternative designs in more depth than the publication has. A reader looking for the strongest UK-grounded analysis of the reform should read CenTax, IFS, and the Commons Library briefing first, and the publication second. That is what source authority looks like for one citizen using AI tools versus established research institutions with HMRC data access. Stating it openly is what the source-authority critique was asking for.

"A critic" attribution corrected. The piece's lead paragraph said "a critic who has read the publication carefully points out that the principle question deserves engagement" — implying human review. The truthful version: "an AI cross-critique session pointed out that the principle question deserves engagement."

Closing personal-interest paragraph updated. Previous version said the author has "personal interest in the operational analysis (he is in the affected cohort)" — outdated relative to today's conflict-disclosure rewrites. New version states the standing fact (hundreds of early-stage tech investments) and the resolved-position fact (his personal tax position has been settled by planning independent of which of the design positions wins).

Subsequent rounds of the source-authority sweep will cover the long article and the funding-stack piece — the two other load-bearing analytical pieces — and will apply the same discipline: re-anchor empirical claims to specific cited sources, downgrade narrative claims to what the cited sources actually establish, and engage explicitly with what UK institutions have published rather than gesturing at them as background.

1 May 2026 — The bigger-claim piece retired; self-promotional framing stripped from the rest of the corpus

Category: framing · workflow

Doug noted that the publication had been spending too many words talking about itself — its claims about what it had produced, its claims about how the work was made, its claims about what AI tools have made possible — when the publication's job is to put facts and data on the question in front of the reader and let the reader judge. The depth claim had already been retracted across the corpus earlier today; the floor/ceiling framing that replaced it was a more sophisticated version of the same self-promotion. Doug's instruction: "I want simple facts and data."

The bigger-claim piece is retired. The piece existed to make a claim about the publication's significance — that one citizen prompting AI tools had shipped, in eight hours, a substantial publication on a question the government had not published equivalent analysis of. Even after the comparative-depth claim was retracted, the piece's structure as a defence-of-its-own-significance was the wrong register. The piece has been removed from site-config.js (it no longer appears in the build, the homepage, the article cards, the archive, or the sitemap). The URL /articles/2026-04-30-the-bigger-claim.html now 301-redirects to /about.html, where the AI-methodology disclosure that the piece's truthful core covered now lives in compressed form. The piece's stale source file remains in articles/ but is no longer rendered or linked.

The piece count across the publication drops from eighteen to seventeen. Eleven featured pieces on the homepage rather than twelve. The reading guide standfirst, the 404 page, llms.txt, the sitemap, and the homepage callouts have all been updated.

The bigger-claim hero callout on the homepage is removed. The first hero callout on the homepage was the bigger-claim callout — it described the publication's eight-hours-from-a-citizen production as the headline framing, with the floor/ceiling argument in the body. That callout is gone. The homepage now opens with a callout describing the reform itself: what changed on 6 April 2026, who is affected, and the question the publication treats — without telling the reader what the publication has accomplished or how to feel about the AI tools that produced it.

The "argument is not about how much, it is about when" hero callout is softened. The previous version led with "The argument is not about how much. It is about when" as a verdict, then said "the public debate has been having the wrong fight, loudly. The right fight is timing." That framing told the reader the publication's reframe was the obviously correct way to look at the question, before the reader had decided that for themselves. The new callout describes the reform factually, names the question of when the tax should fall as the question the publication treats, and presents the four positions on it at equal length. The publication's argument that the timing-and-mechanism question matters more than the amount question is still in the body of the When, Not How Much piece itself — but it is now presented as the publication's argument rather than as the obviously correct framing of the debate.

Self-promotional framing stripped from the twelve-hours piece. The piece's closing paragraph carried the floor/ceiling argument — "the eight hours describes the floor, not the ceiling... a team of three or four people with relevant expertise, working seriously for four weeks... can produce something this publication is not and cannot be: institutional-quality analysis at a small fraction of the historical cost." The paragraph has been replaced with a smaller, factual closing: "Eight hours of real work, four AI tools, a citizen who prompted and answered and shipped — that is what happened, and the rest of the publication has to stand on its own merits, against the test any analytical work has to pass: does it engage with the question, does it cite its sources accurately, does it acknowledge what it does not know."

What stays. The full how-this-was-made methodology piece remains, because Doug specifically said the AI workflow documentation is part of the truthful story and should be preserved. The corrections page (this page) remains, because the trail of corrections is part of the discipline of the work. The conflict disclosure remains in compressed form across the corpus — it states the facts a reader needs to know about the author's standing without claiming a particular posture about what that standing means. The four-positions-at-equal-length structure remains, because the publication does not adjudicate between A/B/C/D and presenting them at equal length is what that looks like in practice. The factual descriptions of the reform, the cohort, the model, the international comparators, and the analytical content of the pieces all remain.

What this changes about the publication's posture. The publication is no longer presenting itself as a demonstration of what AI tools have made possible, nor as having produced more depth than the government on this question. It is presenting itself as: an analysis of the April 2026 BPR/APR reform written by Doug Scott (a UK technology founder with stated standing in the UK tech cohort) using AI tools, with limits openly named, with a corrections page recording how the work has changed in response to AI cross-critique. A reader who wants the analysis can read it. A reader who wants to know how the work was made can read the methodology piece. A reader who wants to know who Doug is can read the about page. The publication has stopped trying to tell the reader what to think about it.

1 May 2026 — Biographical paragraph corrected: Redbrain dates, revenue figure, bootstrapped/profitable status, and earlier ventures named correctly

Category: factual · workflow

Doug noted that the about-page biography had been wrong on multiple specifics. The corrected version names: Redbrain.com founded and led by Doug from 2011 to 2024 (not "over twenty years" — that figure was an AI-generated approximation that survived earlier sweeps); Redbrain reached around £68 million in revenue, operated profitably, and never raised venture capital (none of these facts were stated previously); and before Redbrain, Doug founded and owned carrentals.co.uk from 2003 to 2016, a UK car-rental aggregator that processed over a million rentals without owning a single car, alongside other ventures under the Potential.co group including 30m.com and discountvouchers.co.uk. The two-decades-building-and-backing-tech framing has been replaced with the actual dates and the actual ventures.

The corrected biographical paragraph as it now appears on the about page reads: "Doug Scott is a UK technology founder. He founded and led Redbrain.com from 2011 to 2024 — a UK e-commerce technology business that grew to around £68 million in revenue, operated profitably, and never raised venture capital. Before Redbrain he founded and owned carrentals.co.uk (2003 to 2016), a UK car-rental aggregator that processed over a million rentals without owning a single car, alongside other ventures under the Potential.co group including 30m.com and discountvouchers.co.uk. Beyond his own companies he has invested personal money directly and indirectly into hundreds of very-early-stage UK tech companies and has advised many more — the standing he writes from when this publication addresses the UK tech cohort specifically. He turned to writing in 2026."

The correction matters in two ways. First, factual accuracy: a publication on UK tech and growth that gets the author's own UK tech company history wrong should not be presenting itself as the kind of publication that gets things right. Second, the actual facts — bootstrapped to £68 million revenue without VC, profitable across the period, distinctive earlier ventures including a car-rental aggregator that processed a million rentals without owning a car — are a stronger signal of the author's standing in UK tech than the previous vague "two decades building and backing" framing. The previous framing was a generic operator-credential. The corrected version names what was actually built. The reader can weight the analysis with the specifics rather than the abstraction.

The twelve-hours piece's personal-narrative paragraph also corrected. Where the piece previously said "I had spent twenty years building and backing technology companies" as the starting point of the BPR question, it now says "I had spent more than a decade building UK technology companies — Redbrain.com from 2011 to 2024, carrentals.co.uk before that, and others — and many more years investing personal money directly and indirectly into hundreds of very-early-stage UK tech companies and advising many more." The piece's central message — that eight hours of AI-tool-prompted work produced this analysis — is unchanged. What changed is the truthful version of who Doug was when he started the eight hours.

Sources: Companies House records on Redbrain Limited (founded 2011), Pomanda's company profile showing Redbrain Limited turnover of £56.3m at January 2024 with subsequent growth to £68m per the author, Tracxn's RedBrain profile (founded 2011, no venture-capital funding rounds at the operating-company level), and 37x.com which lists Doug as founder of Redbrain, 30m.com, carrentals.co.uk, and discountvouchers.co.uk. The £68m revenue figure is the author's own statement; AI cross-critique cannot independently verify the most recent year's numbers but the bootstrapped/profitable/never-raised-VC framing is corroborated by public sources.

1 May 2026 — Scope-and-standing fact added: the publication scopes to UK tech because that is the sector the author actually knows, and UK tech is where the country says growth should and does come from

Category: framing · workflow

Doug noted a fact that the publication had been gesturing at without stating openly: the publication scopes to UK tech because that is the sector Doug actually knows, and the depth he can write at on tech is not depth he could write at on agriculture, mature family businesses, or other parts of the BPR base. The relevant standing fact: Doug has invested personal money directly and indirectly into hundreds of very-early-stage UK tech companies and has advised many more. He has watched the pattern across this cohort at first hand for years. He does not have equivalent standing in the other affected cohorts. The honest version of the publication's scope choice is: the publication scopes to where Doug can speak from observed pattern rather than from secondhand summary — and that is UK tech, not the wider BPR base.

This fact also lands a separate point the publication had been making weakly. UK tech is not just one of the affected cohorts; it is the cohort the country's industrial-strategy documents — across multiple administrations — say UK growth should and does come from. The intersection of these two facts is not a coincidence. The cohort the country has decided it wants to grow more of is the cohort this publication has standing to write about. Stating both facts together makes the scope choice legible: the publication writes about UK tech because Doug knows UK tech, and UK tech is where the growth conversation actually sits.

This is not a claim that UK tech matters more than agriculture or family businesses or the rest of the BPR base. It is a statement of what the publication can and cannot do. The other affected cohorts deserve the same depth of analysis from people with standing in those sectors. The publication does not pretend to cover them and is explicit that it does not.

Changes applied. The "Scope" callout in every long-form article that carries one (the long article, the readable companion, the funding-stack piece, the principle piece, plain-english pieces) has been rewritten to state both reasons: that Doug knows the sector and is writing from observed pattern in it, and that UK tech is where the growth conversation sits. The conflict-disclosure language across the corpus — every "What it is and is not" box, every author block, every callout, the about page, the terms page, the model page, the footer-legal line on every page, the for-journalists piece, the short-version personal-stake paragraph, the when-not-how-much postscript, the for-uk-tech-founders postscript, and the docx/pdf generators — now also names the standing fact: "He has invested personal money directly and indirectly into hundreds of very-early-stage UK tech companies and advised many more — the standing the publication is written from on this specific sector." The about-page biographical paragraph has been updated to name the same fact as part of Doug's professional background, alongside Redbrain. 55 disclosure replacements across 37 files in the first pass plus 4 stragglers in postscript variants plus 2 about-page updates. The downloadable .docx and .pdf companions remain stale relative to the website canonical until the build-doc / build-readable / build-treasury generators are re-run.

The change strengthens the publication on two dimensions. First, it answers the document 29 reviewer's "scope bias" critique on its strongest terms: the reviewer was right that the publication is tech-focused and that this is a scope choice, and the publication now names the choice and the reason for it openly rather than letting it sit as an implicit limitation. Second, it adds a piece of standing the previous disclosure had not quite captured. The previous disclosure said Doug owns shares in unlisted UK companies (true) and has been engaging with government on the question for some time (true). It did not state that on the tech cohort specifically — the cohort the publication actually addresses — Doug has investor and advisor relationships across hundreds of companies, and that this is what he is writing from. A reader who weights authorial standing as one input among several can now weight it more accurately.

1 May 2026 — "External reader" / "outside reader" framing retracted again where it had survived earlier sweeps

Category: workflow

Doug noted that several places in the corpus still implied human external review where there has been none. There has been no external human reviewer of any of this work, ever. Every "reader" or "reviewer" or "critic" referenced anywhere on the site has been another AI session — the same four AI tools (Claude, ChatGPT, Grok, Gemini) running in fresh chats, prompted by Doug to critique the publication's own output. Multiple chats of the same AI counted as the same reviewer's repeated critique would; they are not different humans.

The corrections page already had an entry from earlier today titled "'External reviewer' framing corrected to 'AI reviewer' across the corpus" which had caught and fixed several instances. This entry covers the further round Doug flagged — the references that survived that earlier sweep:

Long article body and postscript. The substantive paragraph on the across-system framing said it was borrowed from "an outside reader's critique." The article's closing postscript said the rewrite was "in response to a reader's substantive critique." The Section 3 reset paragraph said "A reader noted that this was advocacy with the architecture of analysis rather than analysis itself. The reader was right." All three now say AI cross-critique session, with explicit qualifying language naming that another AI tool — typically Claude in a fresh chat, prompted to read the draft as a hostile reader would — produced the critique.

Readable companion piece. The "limits" section said "As an outside reader put it..." when introducing the system-internal-versus-system-selection framing. Now says "An AI cross-critique session put the boundary more cleanly..."

Methodology piece (How This Was Made). The version-history paragraph for the seventh-version rewrite said "A reader then noted that it was advocacy." Now says "An AI cross-critique session — Claude in a fresh chat, prompted to read the published article as a hostile reader would — then noted that it was advocacy... That observation, from another AI session rather than from a human external reviewer, is what triggered the seventh version's rewrite." The eighth-version paragraph similarly clarified to name AI cross-critique. Two further paragraphs (paragraphs 100 and 102) carried disclosure language that pre-dated today's conflict-disclosure rewrites and contradicted them; they have been updated to match the new "personal position resolved by planning, continuing engagement is from care for the country" framing.

Common-reactions homepage callout. The callout said the page engages with critiques "received from AI cross-critique and external readers." The "and external readers" half is the overclaim. Replaced with "the strongest AI cross-critique sessions the publication has run on its own work," with an explicit parenthetical "(No human external reader has reviewed the publication; the critiques engaged with on this page are all from AI sessions.)"

build-doc.js. Three references in the docx generator script — the long-version body, the Section 3 reset paragraph, and the postscript italics text — still carried "outside reader's critique" / "a reader noted" / "in response to a reader's substantive critique." All three updated. The downloadable .docx companion remains stale relative to website canonical (the build-doc generator has not been re-run with all of today's corrections); when it is regenerated the truthful AI cross-critique language will propagate.

This is the third time the publication has had to retract human-reviewer-implying framing. The pattern is specific: when Doug uses an AI tool to critique earlier output, the language that AI tool naturally produces describes itself as "an outside reader" or "a reader noted" or "the reviewer's view" — register from training data where critiques are written by humans. Unless that language is caught and rewritten on every iteration, the overclaim seeps back in. The publication has now been corrected three times for this and will continue to catch it when it recurs.

1 May 2026 — Disclosure extended again: the publication is what AI tools made it possible for a citizen who had already been raising the question with government to express systematically

Category: workflow · framing

The disclosure rewrite earlier today established that the author's continuing engagement is from care for the country rather than from personal exposure to the outcome. Doug noted a fact that fills out the picture and answers the question every reviewer has implicitly asked — why this work, why this author, why now. He had been raising the substance of the BPR question with government for some time through the channels available to citizens (meetings, written submissions, conversations with people in policy roles). What he did not have, until recently, was the means to express systematically what he was watching happen across his peer group at the depth and structure the question required. He could describe individual cases in conversation. He could not produce the cohort-level analysis, the funding-stack breakdown, the international-comparator review, the fiscal model with its sensitivities, the position-by-position operational treatment, and the eighteen pieces of registered analysis that this publication contains, in eight hours of his own time, without AI tools. The publication is not the citizen's first attempt to engage with the question. It is the first attempt that produced a body of work proportionate to the question.

This fact connects the floor/ceiling argument to the conflict-disclosure argument at the right level. The floor/ceiling argument was about what AI tools have made possible in the abstract. The disclosure argument was about who Doug is. The connection is: AI tools made it possible for a citizen who had already been engaging with government on the substance, but who lacked the means to articulate the systematic pattern across his cohort, to finally express what he had been observing — at a depth that survives careful reading and at a speed that previously required institutional resources. That capability now exists for any citizen with similar standing on any contested public-policy question. That is the demonstration this publication is, more precisely than the abstract version.

The fact is now stated across every disclosure surface on the site, slotted into the existing passage after the "the outcome has minimal effect on him personally now" sentence: "He has been raising the question with government for some time; the publication is what AI tools made it possible for him to express, at the depth and structure the question required, about what he was watching happen across his peer group." 33 individual replacements across 28 files in the first pass, plus 30 more across 28 files for the footer-legal and other variant forms, plus 4 final stragglers in the postscript paragraphs of the founders piece and the when-not-how-much piece. The bigger-claim piece's floor/ceiling section now carries an additional paragraph stating the author-specific version of the case, which is more precise than the abstract "AI tools have moved the floor" claim and answers the "why this author, why now" question reviewers have been asking implicitly.

1 May 2026 — Conflict disclosure rewritten again: the truthful version is about continuing care for the UK, not about whether there is a live financial stake

Category: workflow · framing

The disclosure introduced earlier today — that the author had no live financial stake in any of the design positions winning — was factually true but read coldly, almost technocratically: I solved my problem, here is an analysis. Doug noted that this missed the actual reason the publication exists. The honest version is more specific and more revealing about what the work is about.

Doug was born in the UK, lived overseas, and came back to the UK because of what he values about the country. His companies have always been UK-owned, UK-operated, and UK-tax-paying. When the BPR reform was announced he adapted his and his family's position; many people in his cohort did not. The outcome of the policy debate has minimal effect on him personally now, but that is not the same as not caring about what is best for the UK as a country. The interest the publication writes from is that continuing care, not personal exposure to the outcome.

This framing replaces the previous "no live financial stake in any of the four winning" language across the corpus. It is a stronger disclosure on two grounds. First, it is more honest about why the publication exists: not from a position of grievance or self-interest, but from a continuing care about the UK as a country. Second, it answers the document 29 reviewer's most cutting line — "the author position... disclosure does not remove bias" — by reframing what the position actually is. The author isn't a UK tech founder hedging his exposure; he's a person who chose this country, has paid into it through UK-resident companies for years, dealt with his own situation when the reform was announced, and continues to think about the question because he cares what happens to the country. The four facts the reader should know — born in the UK, lived overseas and came back, UK-owned UK-operated UK-tax-paying companies, adapted when many in his cohort did not — are stated openly across every disclosure surface.

Replacement applied across 44 files with 83 individual replacements. The new language now appears in: the "What it is and is not" boxes at the top of every article; the author-block at the close of every article; the on-the-author's-position callouts; the plain-english pieces' interest paragraphs; the model page author block; the about page disclosure section; the terms page author's-interest paragraph; the footer-legal author-conflict line on every page; the for-journalists piece's source-quality and postscript references; the short-version's personal-stake paragraph; the when-not-how-much postscript; the for-uk-tech-founders postscript; and the build-doc / build-readable / build-treasury generators that produce the downloadable .docx and .pdf companion files.

1 May 2026 — Conflict disclosure strengthened: author has no live financial stake in any of the design positions winning

Category: workflow · framing

Note on downloads. The .docx and .pdf companion files in /downloads/ were generated on 30 April 2026 — before today's series of corrections (the depth-claim retraction, the architecture-flattening round, the Position D addition, the principle-piece strengthening, and this conflict-disclosure update). The website is now the canonical version of the publication. The downloads will be regenerated in a subsequent build; until then, readers should treat the on-site article as authoritative where it differs from the corresponding .docx or .pdf. The interactive model and the Excel companion remain accurate.

An AI reviewer observed that Position D — the targeted higher threshold for qualifying unlisted trading-company shares — looked like the author's commercial interest in everything but name and that adding it as a peer position to A, B, and C was itself a structural tilt. The factual premise of that critique is wrong, and the corpus had been letting the wrong premise stand. Doug clarified: he has owned shares in unlisted UK companies, and that ownership prompted his original interest in the question, but his personal tax position has been settled by planning that took place independently of which of the design positions the policy debate eventually adopts. The reform's outcome does not change his personal position. He has no live financial stake in any of the design positions winning the policy debate.

The previous disclosure language across the corpus said the author "owns shares in unlisted UK companies and would be affected by parts of what is discussed" and was "trying to be impartial." That language gave a hostile reader plenty of room to read the analysis as motivated by a live commercial interest. The truthful version is stronger: "The author was born in the UK, lived overseas, and came back to the UK because of what he values about the country. His companies have always been UK-owned, UK-operated, UK-tax-paying. He adapted his own position when the BPR reform was announced; many in his cohort did not. He has invested personal money directly and indirectly into hundreds of very-early-stage UK tech companies and advised many more — the standing he writes from on the specific sector this publication scopes to. The outcome of the policy debate has minimal effect on him personally now, but he has been raising what he is seeing happen across his peer group with government for some time, and the publication is what AI tools made it possible for him to express systematically about a pattern he had not previously had the means to articulate at this depth."

This new language now appears in: every article's "What it is and is not" top-of-article box (the long form); every article's author-block (the closing form); the on-the-author's-position callout in the long article; the plain-english pieces' interest disclosures; the model page's author block; the about-page section on author conflicts; the terms page's author's-interest paragraph; the footer-legal author-conflict line on every page (which propagates to every article via the build); and the for-journalists piece's source-quality and postscript references. 65 disclosure replacements across 37 files plus 4 source-file updates plus 10 follow-up replacements catching stragglers in the short-version, when-not-how-much, readable-companion, founder-piece postscript, and the docx/pdf generator scripts.

The structural implication for Position D is the one the reviewer was right to flag at the architectural level even while wrong on the personal-stake premise. Position D was added in this morning's round on the AI reviewer's framing observation that the article was incomplete without it (the design that someone reading carefully and concluding "principle right, mechanism right for most of the base, operational problem real but localised" would arrive at). The position has its strongest case and sharpest objections both in the corpus, named openly. The author has no commercial alignment with Position D winning over A, B, or C; none of the four would change his personal position. A reader who finds the analytical case for Position D weak, or who is persuaded by Position A's strongest form that special-pleading carve-outs erode tax bases, is reading the publication legitimately. A reader who is persuaded by Position D is also reading it legitimately. The publication does not adjudicate.

1 May 2026 — Architecture flattened: every place where the publication turned a plausible concern into a preferred policy direction has been neutralised

Category: framing

Doug observed that the article was weakest where it turned a plausible concern into a preferred policy direction, and that he did not want the publication to pick a decision. The publication's stated posture has been four-positions-at-equal-weight; the architecture in several places had been pointing at Position B (or at Position C with triggers toward Position B) regardless of what the surrounding hedging said. The own self-critique page (Common Reactions) had named the problem in earlier rounds — calling it "a strong advocacy-adjacent policy essay pretending harder than it should to be neutral analysis" — and the publication had recorded the critique while continuing to embed the tilts. This round flattens them.

Section 3 of the long article restructured. The previous section was titled "One possible framework for acting under uncertainty" and gave four full sub-sections of operational detail (Step 1 / Step 2 / Step 3 with five trigger thresholds / Step 4) to one framework — Position C with hard triggers toward Position B — while gesturing at A's and B's alternatives in a single closing paragraph. The structural choice signalled which framework was preferred regardless of the hedging language. The new Section 3 is titled "What each of the design positions would look like in operational form" and gives roughly equal operational treatment to all four: Position A in operational form, Position B in operational form, Position C in operational form (with the trigger / no-trigger variants distinguished), Position D in operational form. The four practical measures are presented as common ground that all four positions accept, separately from the contested choice above them. The five illustrative trigger thresholds remain but are now part of Position C's trigger-variant description rather than the structural backbone of the section. The closing line names what the publication is doing: "The contested choice is the design of what sits above the four measures... The publication does not adjudicate between them."

Funding-stack piece tilt removed. A paragraph in the worst-case-asymmetry section had said: "This is the principal reason the model points toward mechanism change rather than preservation, despite the publication's general posture of declining to recommend." That sentence was the publication acknowledging it tilted while continuing to tilt. The paragraph has been rewritten to present the worst-case logic symmetrically: a policymaker reasoning from worst-case downside reaches one ranking; a policymaker reasoning from upside reaches a different ranking; the model's role is to show what each posture implies, not to choose between them.

Principle piece two-track recommendation removed. The piece had carried an explicit policy recommendation in its body: "The recommendation is: keep the threshold approach for founder equity, where it works; introduce a parallel conditional-relief track for operating family businesses, where the threshold approach does not." That recommendation contradicted the publication's stated posture. It has been replaced with an analytical observation about asset-class heterogeneity from which a reader can reach four different conclusions — that single-mechanism-for-both-cohorts is acceptable; that two-track better matches the heterogeneity; that one cohort should be moved out of the BPR base; or that the asset-class-fitness objection points to Position B's mechanism critique applied at the level of asset class. The publication does not pick among these.

For-founders piece's "publication's view, addressed to you" section neutralised. The previous section had stated the publication's policy position to founder-readers and ended by recommending the founder cohort engage publicly in the debate (a soft recruitment toward Position B / D outcomes via political mobilisation). The replacement section "What the publication treats as settled and what it treats as contested" distinguishes the principle (settled) from the timing-and-scope question (four positions at equal weight) without privileging any operational position. The closing paragraph now reads: "Whether [founders] should engage more is a question for each founder; the publication does not tell you to advocate, does not tell you to stay quiet, and does not tell you which position to support if you do engage."

Five-minute version "what the publication thinks" section neutralised. The previous section made specific recommendations — that HMRC and the OBR publish their modelling, that they consult on different mechanisms, that they review after three years. The replacement section "What is settled and what is contested" sets out the principle (settled) and the four positions on the timing-and-scope question (at equal weight). What the publication suggests HMRC, OBR, and HMT could usefully publish to advance the debate is now framed as supporting the public debate's evidence base rather than as supporting any specific policy outcome. The shortest-possible-version closing has also been adjusted: where it previously said "The principle is right. The amount is roughly right. The timing... is the part the government has not justified," it now says the principle is broadly accepted, the amount sits within a range reasonable people accept, and the timing-and-scope question is what the design positions actually disagree on.

Common-reactions self-acknowledgement updated. The architect-of-analysis critique section in the self-critique page previously named the tilt and treated it as something the seventh-version rewrite had addressed. It now names the tilts that survived the seventh-version rewrite (the funding-stack tilt, the principle-piece two-track recommendation, the for-founders publication's-view section, the five-minute "what the publication thinks") and acknowledges that the eighth-version rewrite has flattened them. A separate line previously privileged Position C as "the version of 'wait' that takes the critique most seriously" — that line is gone; Position A and Position C are now both treated as versions of "wait" at equal weight.

Methodology piece version history updated. An eighth-version entry has been added to the iteration history acknowledging that the seventh-version rewrite (which had moved from "advocating Position C with hard triggers" to "presenting the framework as one possible response") had not gone far enough — the conditional-analysis register concealed structural tilts that the architecture continued doing the work of. Position D's addition earlier today made the problem worse because four positions cannot be at equal weight if the architecture points at one of them. The eighth version flattens the architecture across all five affected pieces.

What the publication is now is what it always claimed to be: a framework for the debate, four positions at equal operational weight, the structure of the disagreement made visible, the reader given the materials to form their own view. The tilts that had been doing analytical work in the architecture — and that the architecture had been concealing under the conditional-analysis register — are gone.

1 May 2026 — Principle piece strengthened: explicit chain of why intergenerational tax is good for the heirs, the state, and society

Category: framing

The publication's operational analysis is structured around when intergenerational tax should fall, not whether it should — which assumes the principle has been settled. The principle piece (On the Principle) was treating the heir-productivity argument as one paragraph in a longer essay; Doug observed that the analytical foundation for the principle should be foregrounded, because the timing question only matters if the principle itself is correct, and the publication had been gesturing at the principle without quite stating the chain.

A new section — What the principle is actually for: heirs, the state, and society — has been added to the principle piece between the fairness section and the strongest-objection section. The section names three beneficiaries of intergenerational tax existing as a category and works through what the data says about each. For the heirs themselves: the labour-supply and entrepreneurship literature (Holtz-Eakin, Joulfaian, and Rosen's 1993 Carnegie Conjecture finding and the subsequent body of work) shows that heirs of large untaxed transfers have lower labour-force participation, lower entrepreneurship rates, lower risk-taking, and higher rentier-strategy adoption than equivalently-situated heirs whose family wealth was taxed at transfer. The section makes the timing point that follows directly: a tax bill at death on an illiquid asset gives the heirs the wrong kind of pressure (forced disposal, disrupted companies); a tax claim that crystallises at later realisation gives the heirs a custodianship relationship with the wealth — they steward the asset through the holding period, the eventual tax functions as a constraint keeping the position connected to actual realisation, and the heirs end up in the situation the literature suggests is most productive (resources combined with a continuing stake in deploying them, rather than untethered receipt). For the state: realisation values are larger, more legible, and less subject to behavioural leakage than death-date values; the state's revenue interest, fiscal-stability interest, and industrial-strategy interest all point in the same timing-shifted direction; only the political optics pull the other way. For society: the social-mobility argument operating over generational timescales — without intergenerational tax, the productive economy slowly drifts into a rentier economy and the political settlement that supports the productive economy is itself undermined.

The closing line of the new section ties the four positions together: "The four positions in the operational piece — A, B, C, and D — are four ways of delivering the same principle. None of them is 'no intergenerational tax.' That option is off the table, on the publication's view, not because it is unfair but because its consequences for the heirs themselves, for the state, and for the society they live in are all worse than the consequences of getting the timing right." The new section does not introduce new claims the corpus did not already make in fragments; it pulls the fragments into the explicit chain Doug asked for, and makes the reasoning visible enough that a reader engaging with any of the four operational positions sees what the analytical foundation under all of them is.

1 May 2026 — Position D added: targeted higher threshold for qualifying unlisted trading-company shares

Category: framing · legal/tax

An AI reviewer engaging with the article observed that a fourth design — keeping IHT-at-death but raising the £2.5m / £5m allowance specifically for qualifying unlisted trading-company shares — is structurally distinct from Positions A, B, and C and that the article was weaker for not engaging with it. The reviewer's framing: "It's the design that someone reading the piece carefully and thinking 'the principle of the reform is right, the mechanism is right for most of the base, the operational problem is real but localised to one cohort' would arrive at. The piece doesn't give it a name and doesn't engage with it. That's a real gap." Doug agreed it should be added for completeness, even though it can be portrayed as unfair to other sectors.

Position D added to the long article and the readable companion. Both pieces now carry Position D as a fourth position-on-its-best-terms, with the strongest case for it (industrial-strategy alignment, scope adjustment rather than mechanism change, no quasi-equity exposure) and the strongest objections to it (the special-pleading-erodes-the-base argument that Position A's strongest form was built to reject; the definitional problem of "what is tech" with no clean statutory solution; the fairness objection of raising the cap for tech founders but not for a 200-year-old engineering firm). The publication does not advocate for Position D over A, B, or C — it has the strongest direct response to the operational problem the article identifies and the sharpest objections of any of the design positions, both worth seeing alongside each other. References to "three positions" updated to "four positions" across seven affected files including section headings, the dispute-summary line, and Question 5's normative-objective analysis (which now notes that B and D both deliver industrial-strategy alignment but with different fairness trade-offs — B is a different mechanism for one asset class, D is a higher allowance for one cohort, and the scope-adjustment version is harder to defend on equity grounds than the mechanism-change version).

1 May 2026 — Floor/ceiling reframing of the eight-hours figure; narrowing of government-publishing language; retracted "architect" framing removed from how-this-was-made standfirst

Category: framing

Three framing corrections in one round.

Floor/ceiling reframing. The publication has been presenting the eight-hours-from-a-citizen figure as a defensive headline — this is what one citizen can do, so don't dismiss it as nothing — but that framing leaves the more interesting question implicit and lets a reader conclude "if this is what one motivated founder produces in a weekend, why treat it as authoritative?" Doug pointed out the framing should pair the floor (one citizen, eight hours) with the ceiling (small teams of specialists, weeks of work, proper review — what is now possible from serious institutional work using the same tools, which is much higher than the floor and is the more interesting question). The publication does not try to answer the ceiling question; it tries to make the floor visible enough that the institutions notice the question is now worth asking. Floor/ceiling sentence added to the homepage hero callout body, the bigger-claim piece, the methodology piece, the bigger-claim standfirst (which propagates to meta tags and article cards), and llms.txt. The strongest formulation: "the eight-hours-from-a-citizen demonstration says the floor has risen. It does not say anything about the ceiling, which is the more interesting question, and which the institutions with the relevant expertise are the right ones to answer."

Narrowing of government-publishing language. The phrase "the UK government has chosen not to publish its analysis" was being used in unqualified form across the lead, the hero callout title, and several supporting paragraphs, where it implied blanket non-publication. The government has of course published the policy paper, the impact assessment, costings, and administrative-impact discussion. The narrow truthful claim is that the government has not published detailed analysis on the specific sub-questions this publication treats — timing, cohort behavioural response, indirect fiscal effects, two-track design. Updated across the bigger-claim piece's lead and supporting paragraphs, the homepage hero title, the bigger-claim standfirst, the archive, and the article cards. Reviewer was right that this matters; the previous narrowing in earlier rounds caught most of it but left the lead unqualified.

Retracted "architect-and-builders" framing removed from how-this-was-made standfirst. The methodology piece's standfirst still said "the architect-and-builders framework, the iteration history, the disclosures" — language retracted weeks ago in the workflow-honesty correction. The standfirst now reads "the truthful prompt-and-ship workflow, the four AI tools, what each produced, what was retracted, and what remains uncertain." Reviewer caught residual retracted framing in an article card; correction propagates from site-config.js through every cross-reference on rebuild.

1 May 2026 — Substantive AI review of the long article: foundational numbers verified against primary sources, several substantive critiques recorded

Category: factual · framing · source/citation

An AI reviewer (another AI session) read the long article (What the UK Government Should Actually Do) carefully and gave the most substantively engaged review the publication has had so far. The review's positive findings: the framing is "unusually disciplined" in separating empirical from normative questions and refusing to advocate; the disclosures are "exemplary" and tell the reader exactly how much weight to give the piece; the migration-evidence treatment (distinguishing the OBR 25% non-dom figure from FT Companies House data from named-founder cases, and explicitly retiring the Henley & Partners 16,500 figure) is "the kind of source discipline mostly absent from public debate." The reviewer specifically flagged five specific-number claims as worth verifying against primary sources before "someone with hostile intent reads it" — the kind of warning the publication's correction process is supposed to act on.

Foundational numbers verified against primary sources. Two of the reviewer's flagged numbers have now been checked. HMRC December 2025 estimate of approximately 1,100 estates affected, 185 with APR claim, 220 BPR-only excluding AIM-only: confirmed against the GOV.UK 23 December 2025 press release ("up to 1,100 estates... up to 185 estates claiming agricultural property relief... up to 220 estates across the UK only claiming business property relief") and the Family Business Research Foundation January 2026 BPR update which traces the trajectory of HMRC's static estimates (2,000 → 1,400 → 1,100). The numbers are correct as the publication uses them. Spousal transferability of the £2.5m allowance giving £5m per couple: confirmed against the same GOV.UK 23 December 2025 announcement ("Given the allowance will be transferable between spouses, a surviving spouse or civil partner will be able to pass on up to £5 million of qualifying agricultural and business assets tax-free"). The reviewer was right to flag this as worth checking — spousal transferability of the new BPR allowance was contested during the consultation — but the final rules confirm the £5m couple figure. CPI indexation from April 2031: confirmed against M&G Wealth practitioner publication summarising the consultation response. Effective 20% rate above the cap: confirmed (40% IHT × 50% reduced value).

Three remaining specific-number flags not yet verified. The OBR's January 2025 supplementary forecast and its "approximately 25 per cent" non-dom assumption — cited from secondary sources, not yet checked against the OBR's published forecast directly. The Sifted "15-20 per cent of new business enquiries at one specialist firm" figure — the publication flags this as adviser-survey/qualitative but the reviewer correctly notes that even with that flag it is "a single firm, self-reported, not a sector measure" and a casual reader could miss the qualification. The "approximately 150 directors moved specifically to the UAE in the second quarter of 2025" — from the same FT Companies House analysis the publication elsewhere flags as contaminated by simultaneous tax changes, and the publication should be consistent about that contamination warning every time the data is used, not only once. These three are flagged here as known limits; specialist readers are invited to send corrections.

Substantive arguments the reviewer made that the publication accepts but does not rewrite the article over. First, the structural-quasi-equity point under Position B is "rhetorically strong but the analytical content is thinner than the framing suggests" — a contingent tax claim deferrable until liquidity is not the same as equity exposure (no upside participation, no governance, capped at the rate). The reviewer is right; the corrections page already records this critique under the policy-paper review entry from earlier today. Second, the timing-warning ("any data after April 2026 has been pre-emptively responded to") "cuts both ways" and is "an unfalsifiable shield" — any small measured response can be explained as absorbed pre-reform departure. Scenario four engages this partially; the reviewer says not fully. The publication accepts this is a real defect in the framework's falsifiability and notes it here without rewriting. Third, Position B is steelmanned less hard than Position A — specifically the argument that CGT-on-realisation is simpler, not just better-calibrated, that the operational case isn't only about illiquidity but about the basic category error of valuing private companies at a moment (death) that has nothing to do with the company. The reviewer is right that this is a stronger argument than the piece lets B make. Recorded here; a future revision of the piece should engage it directly.

Posture vs. humility. The reviewer notes that the piece's epistemic humility "occasionally tips into a posture" and that "by scenario four it's almost arguing that nothing can be known, which is itself a position." This is fair. The publication has been catching itself on overclaim more than on under-claim, which has tipped some passages into the opposite over-correction. Worth naming.

The reviewer's overall verdict: "unusually good for what it is — a non-specialist + AI piece on a technical reform — and the meta-honesty about that fact is part of why it works. The main risk isn't bad reasoning; it's specific-number errors that a tax specialist would catch in fifteen minutes... The argumentative structure would survive those corrections; the credibility might not, given the disclosed lack of expert review." The publication accepts this as a fair characterisation. The five specific-number flags and three substantive arguments are now recorded; the foundational numbers the publication has been using are confirmed correct against primary sources; the structural critiques are noted without claiming to have resolved them.

1 May 2026 — Cloudflare Pages headers and www-canonicalisation added

Category: technical

An AI reviewer noted that robots.txt appeared compressed into long lines when fetched, and that the live URL it landed on was www.thelongerlook.com/robots.txt rather than the canonical thelongerlook.com/robots.txt. The source robots.txt is well-formed (105 properly-line-broken lines, UTF-8, no BOM); the source sitemap.xml validates under xmllint. Both files are clean. The reviewer's flattening was almost certainly their fetch tool's prose-renderer collapsing single-newline-separated text — but the underlying issue (Content-Type-driven rendering, and the www/apex inconsistency) is worth solving at the host level. Two new files added to the bundle for Cloudflare Pages: _headers locks Content-Type: text/plain; charset=utf-8 on robots.txt, llms.txt, and humans.txt, and application/xml; charset=utf-8 on sitemap.xml, plus standard security headers (X-Content-Type-Options: nosniff, X-Frame-Options: SAMEORIGIN, Referrer-Policy, Permissions-Policy). _redirects issues a 301 from www.thelongerlook.com/* to the apex domain so the canonical hostname declared in every page's <link rel="canonical"> matches what the live site serves. Once deployed, fetching either hostname's robots.txt will land on the same file with the right Content-Type, and review tools should render it as plain text rather than wrapping it.

1 May 2026 — "More depth than the government has published" claim retracted

Category: framing · workflow

Doug pointed out that the publication's headline claim — that it has produced "more public-domain depth on this question than the UK government has published" — is one the publication cannot actually substantiate, because the comparative form requires knowing what the government has produced internally. The publication does not know that. HMT, HMRC, and the OBR almost certainly have internal modelling and analysis that has not been published; the publication cannot make a comparative claim against material it cannot see. Even narrowed to the public corpus, the claim implies the publication has rigorously surveyed every relevant published government document, which it has not.

The claim is retracted. Across the homepage hero, the bigger-claim piece's title, lead, body, postscript, and meta tags, the article cards, the archive standfirst, and llms.txt, the comparative claim has been replaced with a smaller and truthful version: the publication has produced a substantial body of analytical work on a contested public-policy question that the government has chosen not to publish equivalent analysis of, and the publication does not know what HMT, HMRC, or the OBR may have produced internally. No scoreboard. No "more than" claim. The asymmetry the publication wants to point at is real — there is a question, the government made a policy on it, the government has chosen not to publish its analysis externally — but pointing at that asymmetry does not require ranking the publication against the government's unseen work.

This is the third reviewer round in which a version of this critique has been made. The first two times, the publication's recorded position was to keep the claim with qualifications stated alongside it. That position is now reversed. The retraction is recorded here in full, including the previous declined-recommendation entries on this page that argued the opposite — those entries are not deleted because the trail of how the publication's position changed is itself part of the public record. A reader scrolling back will see the publication's earlier defence of the claim, then this retraction, then the smaller language now in place across the corpus.

1 May 2026 — The Many Builders framing corrected (second time): "where the bears creating the new world live"

Category: framing

Earlier today the publication retracted "memorial" as the description of The Many Builders on the basis that "memorial" implies the people on the site are gone — most are alive and working — and replaced it with "a place that holds the names" / "a register" / "a roll call." Doug now says even those framings miss the point: the right description is "this is where the bears creating the new world live." The names on the site are the bears. The work they are doing is making the new world. The site is alive in a way that "register" and "roll call" do not capture. Updated across every place the site is described — the homepage strategic line, the production callout, the sidebar card, the about page, the methodology piece, every article footer's "About this site" embed, the JSON-LD description field, llms.txt, and robots.txt. Sixteen description sentences and headings replaced. The pattern of two framing corrections in twenty-four hours on the same site is itself worth recording: each AI-assisted draft of a description tends to converge on neutral, structural language ("a place that holds the names") rather than the more truthful figurative language Doug actually uses for the work. The publication continues to catch this when it recurs.

1 May 2026 — Second-reviewer recommendations: warning strip strengthened, depth-claim qualification on hero, model audit path, corrections-page categories

Category: framing · technical · modelling

Second AI reviewer recommended seven changes after fetching the live site. Four implemented: (1) warning strip strengthened from "AI-assisted, written by a non-specialist, not independently verified" to also include "Not tax, legal, or financial advice. Author has a personal interest" — both the no-advice disclaimer and the conflict are now visible on every page above the fold; (2) the depth-claim qualification (depth ≠ accuracy; comparator narrow by choice; specialist institutions like the IFS, Resolution Foundation, CIOT and academic literature would produce work this publication does not match in rigour) is now stated inside the homepage hero callout body, not just on the bigger-claim piece two clicks away — so the qualification appears beside the claim every time, as the reviewer asked; (3) the model page now has an "An audit path — three places where the same calculation lives" section explaining how to validate the model independently across the JavaScript, the Excel, and a hand calculation, plus naming the four assumptions that move the answer most and the conservative-bound effect; (4) the corrections page now tags every entry with one or more of seven categories (factual / framing / workflow / modelling / legal/tax / source/citation / technical), with a key at the top of the page.

One recommendation handled differently from the way the reviewer phrased it. The reviewer asked for "a claim-by-claim evidence table for every major factual assertion." The journalists' piece already does Claim/Source/Confidence work for the headline claims; building this across all 65,000 words of the publication would be a multi-hour rewrite that this round of corrections did not undertake. A signpost to the journalists' piece exists from the homepage and reading guide; readers wanting source confidence on a specific claim can use that piece as the reference and send corrections where the structure is missing.

Two recommendations deferred as out-of-scope (same as previous reviewer round): commissioning paid named reviews from a UK tax practitioner and a fiscal modeller; running a full live technical audit (sitemap validation in Search Console, canonical-URL consistency check between www and non-www, page speed benchmarks). Both are real-world actions; the AI sessions producing this publication cannot perform them. The reviewer's note that robots.txt looked "compressed into very long lines" appears to be a fetch artefact — the source robots.txt has 105 properly-line-broken lines.

1 May 2026 — Direct contact email added (about page, JS-assembled, spam-protected)

Category: technical

A direct email for the author has been added to the Contact section of the about page. The address is assembled at runtime from three separate data- attributes (local part, domain, TLD) by a small script that runs in the reader's browser; the source HTML contains no @ character and no parseable email pattern, so volume-spam harvesters that read raw HTML never see an address. A <noscript> fallback shows the address bidi-reversed for readers without JavaScript. The strongest practical anti-scrape pattern short of a click-to-reveal handler. The corrections page, privacy page, and model page Found-an-error block now point readers at the about-page Contact section rather than carrying the address themselves.

1 May 2026 — Larger body of work signposted to readers and to AI crawlers

Category: framing · technical

This site is one of seven sites by Doug Scott published in April 2026 (this publication; The Many Builders; the trilogy If This Road / Orphans / The Held; and the two bear books The Bear Was Right / The Bear Loved). The seven are pieces of one body of work and the larger work is the thing — none of them is intended to be read alone. The publication had been listing the other sites in its footer and in the homepage production-callout but had not stated the strategic point: that they are a group, not seven unrelated projects, and that a reader engaging with any one of them without seeing the larger work has the wrong picture. Now stated openly in three places: a strategic one-liner directly under the homepage lead naming all seven sites; a rewritten production-callout that opens with "This site is one piece of a larger body of work. The larger work is the thing"; and a complete rewrite of llms.txt (the machine-readable file for AI crawlers and language models) that leads with the same point. The homepage JSON-LD now declares the relationship between Doug Scott as Person and the other six works as subjectOf, so search engines and AI tools that read structured data see the connection. The robots.txt file now contains a comment block listing the other six domains and pointing at llms.txt for full attribution and summarisation guidance.

1 May 2026 — "External reviewer" framing corrected to "AI reviewer" across the corpus

Category: workflow · framing

Doug flagged a recurrence of the same overclaim the workflow-honesty correction (earlier today, this page) was supposed to have closed. The corrections page and several body files referred to "external reviewer" in ways that implied human review. No human external reviewer has read this publication at any stage. Every "reviewer" referenced anywhere on this site has been another AI session — a separate AI tool that Doug used to fact-check or critique the publication's output, fed by Doug pasting text between AI tools. Calling those AI sessions "external reviewers" without qualification is the exact pattern the workflow correction retracted.

The fix: every "external reviewer" / "external review" reference across the corrections page and the body files (the bigger-claim piece, the funding-stack EIS callout, the common-reactions standfirst) is now "AI reviewer," with explicit qualifying language ("another AI session, not a human expert") on first reference in each entry where ambiguity is most likely. The duplicate "Policy paper: substantive review by external reader" entry — which had crept in twice with the same body text, both wrongly framed as human review — is now a single entry titled "Policy paper: substantive AI review" with the truthful framing.

This is the second time the publication has had to retract human-reviewer-implying framing. The first retraction (workflow honesty correction) had explicitly said the "rounds of substantive critique" referenced elsewhere were AI tools critiquing each other. The fact that the framing came back tells us something specific about the production: when AI sessions are used to critique earlier output, the language those AI sessions naturally produce describes themselves as "external reviewers," and unless that language is caught and rewritten on every iteration, the overclaim seeps back in. The publication will continue to catch this when it recurs.

1 May 2026 — Policy paper: substantive AI review

Category: framing · modelling

AI reviewer (another AI session, not a human expert) read the citizen submission and made three substantive critiques. Quasi-equity objection: section 2.8a presented the "state holds contingent quasi-equity exposure across thousands of UK private companies" point as one consideration among several. The reviewer argued, correctly, that this is the strongest single objection to Option B and the paper undersold it. Section 2.8a now states that directly, lets the constitutional argument land properly, and notes that ministers may still choose to accept the implication on industrial-strategy grounds — but the choice should be made knowingly. Illustrative numbers: the −£100m central revenue case for Option B (range −£200m to +£600m) is an AI-assisted synthesis of comparator data, not a calibrated revenue model. Section 2.9 now says this inline at first appearance, not just in the limits section, and adds: "These ranges should be read as a sketch of the parameter space, not as forecast values." (The trigger thresholds in section 2.15 already carry this caveat.) Bracketed principle: the paper accepts the principle of the reform and engages only the operational question, excluding a serious "Position D" view that IHT on operating businesses is wrong in principle. New section 1.4a now acknowledges this scope choice openly: a reader who holds Position D will find the paper answering a question they consider settled in the other direction. The bracketing is defensible but the analysis is one level deeper than the most fundamental version of the disagreement.

1 May 2026 — AI-warning strip and reviewer recommendations

Category: framing · technical

AI reviewer (another AI session that fetched the live site) recommended six changes. Three implemented: (1) a thin warning strip below the masthead on every page, reading "AI-assisted, written by a non-specialist, not independently verified — Method · Corrections", so a scanning first-time reader sees the disclosure without having to scroll; (2) the homepage hero callout now signposts the methodology piece and the corrections log alongside the bigger-claim piece, and the stale "twelve hours" figure in that callout is corrected to "roughly eight hours of real work"; (3) the model page disclosure now names that the Excel has every formula visible and editable, and that the in-browser JavaScript exposes the same calculation logic — making the audit path explicit rather than implicit.

One reviewer recommendation declined and recorded openly. The AI reviewer suggested removing the "more depth than government" framing because "it invites a credibility fight the site does not need." The publication disagrees. The bigger-claim piece already concedes the necessary qualifications — depth is not accuracy; the comparator is narrow by choice; the claim survives only if HMT is the relevant comparator. With those qualifications stated alongside the claim, the claim is the most truthful version of what the publication has actually produced. Dropping the claim entirely would be tidier but less honest. The credibility fight is the right fight to have, on the right terms.

Two recommendations deferred as out-of-scope for the current build: commissioning paid reviews from a tax practitioner, fiscal modeller, and policy economist (real-world action requiring budget and arrangements); and a separate "claims classified by confidence" page (the journalists' piece already does this Claim/Source/Confidence work for the headline claims; a stand-alone classification page is a bigger piece of work).

1 May 2026 — Analytics added (with consent banner, privacy and terms pages)

Category: technical

Google Analytics added on every page, gated by a two-button cookie banner (Accept and Reject equally available, as UK PECR requires). Nothing loads or sets until the reader clicks Accept. A short privacy page names what is collected, why, and the reader's UK GDPR rights. A short terms page consolidates the licence, the not-advice disclaimer, and the no-human-expert-review disclaimer. A Cookie settings link in every page's footer reopens the banner so a reader can change their mind. The publication has no commercial purpose; the analytics are used to understand which pieces are read and inform what to write next.

1 May 2026 — "Same workflow" and "twelve hours" overstatements corrected

Category: workflow · framing

Two overstatements flagged by Doug. "Same workflow": earlier framing said the four-week practice produced books, code, and the IHT publication "using the same workflow" — flattening real variation. Doug, who cannot code, used AI tools earlier in the period to build several websites and around 100,000 lines of code; the books took seven days of build plus three days of modification each; the IHT publication came together in roughly eight hours of real work. Same person, same tools, very different intensities — the range is the point. "Twelve hours": the methodology piece was titled "Twelve Hours" and led with "twelve hours elapsed, about four hours of active prompting" — presenting elapsed time as effort. The honest figure is eight hours of real work. Piece is now titled "Eight Hours, Four AI Tools, One Founder — and Four Weeks of Practice Behind It." Both framings replaced across the homepage callout, about-page lead, meta description, and the methodology piece itself.

1 May 2026 — Third-round residual fixes

Category: workflow · technical

AI reviewer (another AI session) caught four issues earlier sweeps had missed: six article bylines (journalists', tax practitioners', tech founders', five-minute, short-version, when-not-how-much) still used the older "research and editorial direction" framing pre-dating the workflow-honesty correction; the reading-guide standfirst still said "Eleven pieces" while the body said eighteen; the model page lacked Corrections in its nav and a full workflow disclosure; the corrections page directed correctors to a non-existent source repository. All fixed. Bylines updated to the canonical workflow disclosure. Reading-guide standfirst now says "Eighteen pieces in total — twelve featured, six alternative versions and methodology pieces — plus an interactive model and Excel companion." Model page now has Corrections in its nav, the full disclosure including "Doug did not verify the model math," and a Found-an-error link. Corrections page has a Submit-a-correction section pointing to LinkedIn or Doug's other sites. Site footer carries a Submit-a-correction link on every page.

1 May 2026 — EIS treatment distinction and CPI indexation

Category: legal/tax · source/citation

AI reviewer fact-checking against practitioner sources (CIOT, Royal London, Saffery, Withers, Rathbones, PKF) flagged two issues. The reviewer was another AI session, not a tax practitioner; the practitioner sources cited are public webpages the AI reviewer pulled. EIS: the funding-stack piece treated EIS portfolios under the £2.5m cap regime without distinguishing AIM-listed EIS (50 per cent from the first pound, no allowance) from unlisted-private-company EIS (cap-then-50 per cent, like other unlisted trading-company shares). The two are treated differently under the new regime. A callout has been added to the funding-stack piece making the distinction explicit; lead paragraphs in the article and full-version now say "EIS portfolios in unlisted private companies." CPI indexation: the £2.5m allowance is set to be CPI-indexed from April 2031 (subject to future statutory instrument). This was in the treasury paper but missing from the funding-stack lead, the journalists' headline-claim block, and the plain-english-detailed allowance description. Now added with the conditional language preserved.

1 May 2026 — Residual architect-framing in PDFs and pages

Category: workflow · framing

AI reviewer found four places where earlier "across the corpus" sweeps had missed architect-framing language: the five-minute version's "Almost everyone serious accepts" lead (now "this publication accepts the principle"); the twelve-hours piece's "I set direction... the work iterates until I notice I was wrong" passage (rewritten to the truthful workflow); the policy paper PDF's "Architect of the analysis" paragraph (rewritten with the retraction named in place); the bigger-claim piece's "915 BPR-only" reference without the ~220 unlisted-only subset clarified. Estate-count cuts now reconciled on this corrections page: ~915 = total BPR-only including AIM-only; ~220 = unlisted-trading-company-share subset excluding AIM-only; journalists' piece is the canonical reference.

1 May 2026 — Citizen-submission reformat (policy paper)

Category: framing

AI reviewer flagged that the policy paper's HMT-house visual identity (navy/burgundy palette, "Modelled on HMT format" header, "Policy Options Paper" title, numbered-paragraph convention) borrowed institutional register the disclaimers couldn't undo. Palette moved to the publication's own ink-blue / cream / bronze. Header now reads "Citizen Submission · Doug Scott · The Longer Look." Title changed to "A Citizen Submission." Numbered sections, three-options framework, and lettered annexes preserved because they're useful structure; only the visual identity and framing language changed. Annex C names the change.

1 May 2026 — Depth-claim narrowing (bigger-claim piece)

Category: framing

AI reviewer noted the depth claim — "more depth than the government has published" — is plausible because the comparator is narrow, not because the analysis is right. New section in the bigger-claim piece ("Why the depth claim is plausible — and what it does not show") concedes three distinctions: depth ≠ accuracy; comparator is narrow by choice; the claim survives only if HMT is the relevant comparator. The IFS, Resolution Foundation, CIOT, and academic literature would produce work this publication does not match.

1 May 2026 — Workflow honesty correction

Category: workflow

Earlier framings described the human role as "architect": Doug "judged every draft," "made the substantive judgments," "rejected directions that did not survive scrutiny," brought in "rounds of external critique." All of that overstates the human contribution and implies a level of human review the publication has not had. The truthful version: Doug prompted four AI tools, answered when they prompted back, scanned the output, decided to ship. He did not edit the prose, check citations against primary sources, or verify the model math. No human expert reviewed any of this work. The "rounds of substantive critique" are AI tools critiquing each other, not human review. Corrected across the bigger-claim piece, the meta-page, the methodology piece, the about page, the per-article disclosure blocks, the policy paper, and the Word and Excel companions.

30 April 2026 — Three factual errors caught by AI fact-check

Category: factual · source/citation

AI cross-critique pass (a separate tool, not one of the four producing the publication) caught three factual errors across the article, full-version, and policy paper:

1. Friedman et al. co-authors. Cited as "Friedman, de Boom, Khan and Hecht (2024)" — the latter three names were an AI hallucination. Actual co-authors are Gronwald, Summers and Taylor. Corrected throughout.
2. Canada capital-gains inclusion rate. Described as 50 per cent up to C$250,000 and 66.67 per cent above — that was the Trudeau-government proposal, cancelled in March 2025. Actual current rate is a flat 50 per cent. Corrected; effective-tax estimate revised from "35 per cent or more" to "roughly 26-27 per cent."
3. German optional 100 per cent relief threshold. Stated as 10 per cent administrative-asset cap; actual threshold is 20 per cent. Corrected throughout.

30 April 2026 — Other corrections caught by AI cross-critique before publication

Category: factual · framing · source/citation

Smaller corrections caught during production:

• The £1m vs £2.5m direct-cap figure was wrong in early drafts because the AI tools were working from older training data; AI cross-critique caught it before publication.
• The "1,100 estates mostly farms" framing was corrected to the proper HMRC breakdown after AI cross-critique flagged the conflation. The publication uses two related but distinct cuts of the HMRC December 2025 estimate, and a reviewer flagged on 1 May 2026 that pieces use them as if they were the same. They are not. The correct breakdown: approximately 185 estates with an APR claim (some of whom also claim BPR); approximately 220 estates that are BPR-only excluding AIM-shares-only holdings (the unlisted trading-company-share subset on which the operational mechanism question primarily turns); the remainder of the ~1,100 are mainly estates affected by the separate change to AIM-listed shares which now receive 50 per cent rather than 100 per cent BPR. So: "around 915 BPR-only" is correct as total BPR-only including AIM-only holdings; "around 220 BPR-only" is correct as the unlisted-trading-company-share subset excluding AIM-only. Both numbers are used across the publication; pieces that use 220 should clarify they mean the unlisted-trading-company-share subset, and pieces that use 915 should clarify they mean total-BPR-only. The journalists' piece does this carefully and is the canonical reference.
• The £140m year-one rising to £300m timing nuance was added after AI cross-critique flagged that "£300m a year" was directionally close but imprecise.
• The "almost everyone serious" framing on the homepage was softened to "this publication accepts the principle" after AI cross-critique flagged it as rhetorically risky.
• The OECD claim was corrected from "almost every major economy" to "24 OECD countries."
• The heir-productivity claim was rewritten to distinguish the labour-supply finding (robust) from the entrepreneurship finding (mixed evidence).
• The instalment language was corrected from "interest-free for the first nine years" to "in equal instalments over 10 years, interest-free."
• The LP-fund BPR claim in the readable piece was softened with a structure-specific caveat noting that many investment-fund interests may not qualify.
• The £2.4bn pre-reform BPR cost figure was corrected to £1.7bn (HMRC tax-relief statistics).
• "Officials estimate" in the policy paper was corrected to "Author's indicative estimate."

What the publication still expects to be wrong about

The corrections above are what AI cross-critique caught. Errors a specialist reader would catch — methodological controversies in the academic citations, technical tax-mechanics errors AI does not have the training data to catch, framing errors a CIOT or STEP member or IFS economist would catch on a careful read — have not been caught and may be present. Send corrections.

Specific areas the publication regards as most likely to contain uncaught errors:

• The interpretation of the Holtz-Eakin / Joulfaian / Rosen literature, which has been substantially re-litigated since the 1993 papers; the publication treats the labour-supply finding as more settled than the most current literature would.
• The Carnegie-conjecture replication problems, which the publication does not engage with.
• The methodological controversies in the Wilkinson-Pickett work, which the publication cites as supporting the cohesion case without engaging the critics.
• The interaction between the BPR reform and the post-2025 residence and domicile rules, which is technically dense and the publication treats at higher resolution than its training data fully supports.
• The trust-planning and seven-year-refresh rules under the transitional provisions, which are fact-specific and which the publication discusses without practitioner sourcing.
• The CGT s.62 uplift on death survives the BPR reform, but the interaction between the new instalment regime and the s.227 mechanics in the cross-asset-class case is not treated.

If you find an error in any of these areas — or any other — please send the correction. The publication is more interested in being right than in appearing to be right.