Review log: 2026-10-29

One style per role: the opening paragraph still taught "one style family per site", contradicting its own heading and the next paragraph. Rewritten so each role (cutout, comparison, cover, empty state, decoration) gets one style; the regenerate bullet now speaks of a role's style. Source: Principles, One picture style per role, and the Research log entry of 16 October 2026 that renamed the rule.

Picture text rule: "a stray generated hand is the most common reason a picture fails review" was an uncounted frequency claim. Now: generated hands often come out wrong (seen on one site, not counted). Why: every claim carries its evidence.

Start with the question: added a route to How to do UX and UI for the thinking before a screen and the persona panel rounds; the playbook had no link to the house method.

Writeback: Worked examples, Principles, Conflicts and Research log are now linked; "Sources" (no document has that name) now points to the matching Source document through the Source directory in AGENTS.md. The picture style docref now lands on the One picture style per role section instead of the Principles title.

Icon batch: "typically 48, 40 and 32 px" presented one site's icon sizes as a norm; now "every size the page shows it (48, 40 and 32 px on the site this came from)".

Renewals example: grammar, "a inspected" to "an inspected".

Links added: AGENTS.md, Research log, and WCAG 2.2 (to the W3C Recommendation, since the builder brief also goes to the condition without corpus access). Why: the pack named them with no link.

Disclosed checks for P2: "either researched validation strategy" pointed nowhere; it now names both strategies inline (on submission, or when the person leaves the field) so the packet stays neutral between conditions.

Review procedure: one reviewer-only line now maps each task to the worked example it exercises (P1 E1, P2 E2 and E3, P3 E4, P4 E5, P5 E6, P6 E7), kept out of the builder packets.

"This conversation" and "this session" referred to the authoring session of 13 September 2026, which a reader cannot see. Reworded to "a session that has already read this organization" and "the author of this pack did not run it".

Contamination boundary: now states that the six tasks were written on 13 September 2026 and do not exercise the site-wide rules, the house method or the Site coherence pass added since, so a result speaks only to what they cover.

E3 evidence: "GOV.UK defaults to progression or submission" understated the source. GOV.UK says not to validate when a person leaves a field unless research shows earlier checks help; Baymard recommends checking when the person leaves the field; candidate B follows Baymard and departs from GOV.UK. Sources: https://design-system.service.gov.uk/patterns/validation/ and https://baymard.com/blog/inline-form-validation, read 2026-10-29.

E4 check: the manual tabs example now carries a read date and the W3C statement that its code is not intended for production. Source: https://www.w3.org/WAI/ARIA/apg/patterns/tabs/examples/tabs-manual/, read 2026-10-29.

E2, E3 and E5 counterexamples now cite their sources: GOV.UK text input (https://design-system.service.gov.uk/components/text-input/), the GOV.UK validation pattern (character count) and the APG modal dialog pattern (https://www.w3.org/WAI/ARIA/apg/patterns/dialog-modal/), all read 2026-10-29. Our own readings are labelled as such.

E8 reasoning: added that within a site each kind keeps one card and one row anatomy, linking One presentation per kind of thing, so E8 is not read as permission to draw one kind three ways.

Extending this library: Research log is now linked.

Speed is part of the design: the rule said "within about 400 milliseconds feels instant" with no source and credited a "Doherty threshold", while this organization's own Myths, Conflicts and web.dev note reject 400 ms as a general rule. The rule now says to answer at once and show what is pending; the Why cites NN/g (about 0.1 second feels instant, about 1 second keeps flow; https://www.nngroup.com/articles/response-times-3-important-limits/, read 2026-10-29) and Core Web Vitals (INP at most 200 ms at the 75th percentile), and points to Myths for the 400 ms figure.

Links without underlines: the parenthesis read as part of NN/g heuristic 4, which does not mention underlines. It now cites the heuristics source note and labels the underline point as our reading. Source: https://www.nngroup.com/articles/ten-usability-heuristics/.

A motion budget for the first screen: SC 2.2.2 was paraphrased without its "shown in parallel with other content" condition and without a source. It now states the condition and links the new SC 2.2.2 block in the WCAG note.

The 50 ms and 400 ms figures had no source, and "this source owner is refreshed" was internal jargon. The block now links Tognazzini's 50 ms note and the 400 ms myth, and says why to recheck: stable Core Web Vitals change at most once a year, and INP replaced First Input Delay in 2024. Source: https://web.dev/articles/vitals, read 2026-10-29.

LCP, INP and CLS are now spelled out on first use.

The AA baseline sentence now links where this organization sets it (AGENTS.md, writing rules).

Large text is now also given in CSS pixels (about 24 px, or about 18.5 px bold), since websites are built in pixels. Source: https://www.w3.org/WAI/WCAG22/Understanding/contrast-minimum.html, read 2026-10-29.

New section "Criteria other documents rely on" with SC 2.2.2 Pause, Stop, Hide, SC 1.4.11 Non-text Contrast and SC 2.4.11 Focus Not Obscured (Minimum). Why: Principles, the Site coherence pass and the Research log relied on these criteria from memory. Sources: the three W3C Understanding pages, read 2026-10-29.

Hero worked example: the SC 2.2.2 mention now links the new WCAG block.

One rule had two names ("inclusive usability" and "universal usability"); it now uses the source's name, universal usability, in both places.

The dialog button aside now carries sourced platform conventions instead of unsourced claims. Apple places the default button "on the trailing side of a row"; Microsoft right-aligns a dialog's OK and Cancel buttons in a single row, with OK first. The text still says the rule does not come from a study. Source: https://developer.apple.com/design/human-interface-guidelines/alerts read 2026-10-29 (through Apple's JSON data endpoint); https://learn.microsoft.com/en-us/windows/win32/uxguide/win-dialog-box read 2026-10-29.

The forward action rule now covers cards, sections and dialogs, and treats a form's submit or Continue button as an exception that lines up with the form's left edge. Review now flags a lone left-edge action only in a card or section. Before, the rule told reviewers to flag the very form button that GOV.UK, the organization's forms source, prescribes. Source: https://design-system.service.gov.uk/components/button/ read 2026-10-29: "Align the primary action button to the left edge of your form."

The claim that empty states raise first use is replaced with what NN/g actually advises: show system status, teach what the area is for, and give a direct way into the task. The guidance reports no measured effect, so that effect is now marked as our hypothesis. Source: https://www.nngroup.com/articles/empty-state-interface-design/ (Kate Kaplan, 19 September 2021) read 2026-10-29.

The status line now names two layers: Principles 1 to 15 from the 13 September 2026 research pass, with evidence grades, and the working rules added from September to October 2026, from "Convey qualities through the experience" onward, mostly from one recipe site. Each later section counts as a working rule unless it names a stronger evidence class. The review date is now 29 October 2026. Before, the status line described only the September pass and hid the weaker evidence of the later layer. Source: the Principles outline, read 2026-10-29.

The routing paragraph now sends readers to How to do UX and UI for the thinking before a design, to the Design review playbook for a finished screen, and to the Site coherence pass for a whole site. Internal jargon is gone: the Evaluation pack "tests whether this organization improves an agent's design work", and "evidence-based writeback" became "what to record afterwards". Why: the paragraph skipped two documents that own parts of the work. Source: How to do UX and UI and Site coherence pass, read 2026-10-29.

A sentence about database descriptions in directory entries, which only made sense for one product, is deleted. The adoption line now reads "Adopted 17 September 2026" and names its evidence class: owner judgement from a design discussion on that day.

The claim that video services often pad thumbnails now says a service may pad them, with YouTube Shorts as a sourced example. The picture rule now requires text alternatives (WCAG 2.2 SC 1.1.1, Level A) with a quick test step, and Principle 10's checklist starts with text alternatives for images. Code jargon became plain words, and a stale status sentence became a dated one linked to the 16 October 2026 run record. Source: https://github.com/yt-dlp/yt-dlp/issues/9837 read 2026-10-29 (user reports that a Shorts thumbnail is a pillarboxed 1280x720 frame); https://www.w3.org/WAI/WCAG22/Understanding/non-text-content.html read 2026-10-29.

The three "screen coherence pass" links now read "site coherence pass over every screen" and point to the 16 October run record; the 13 and 14 October link points to its own run record. "One picture style per role" now says it is a candidate rule until a second site tests it, as the Research log requires, and links that log entry. Why: one method had two names, the links did not reach the run evidence, and the rule had been promoted without the second site the log asks for. Source: Site coherence pass run records; Research log.

Principle 9's basis now links Shneiderman's eight rules and the HAX library guidance, and keeps the old links as "our application". Principle 14 links the HEART summary for goal-based measurement, and its evidence label now covers persona panels and links the "modelled minds" label in How to do UX and UI. Why: the basis reached only our own application, not the source guidance, and one evidence label had three names. Source: the Shneiderman, HAX and Google Research source notes; How to do UX and UI.

The heatmap sentence now says a modelled heatmap can compare the two heroes only as a critique aid, never as evidence of where visitors look, and links the heatmap guidance in the Site coherence pass. Why: it contradicted that pass and Principle 14.

The sticky bar rule now says the bar must never fully hide the control that has keyboard focus, with top scroll padding as tall as the bar as the technique, and cites WCAG 2.2 SC 2.4.11 (AA). The quick test adds tabbing through every step. Source: https://www.w3.org/WAI/WCAG22/Understanding/focus-not-obscured-minimum.html read 2026-10-29; the note "Source: W3C WCAG 2.2".

Each "the same site coherence pass" Source line now links the 13 and 14 October run record and gives the dates. Lens jargon became plain words, such as "seen by a reviewer reading as a first visitor". Why: the back reference had no link, and the lens terms could not be decoded there.

An owner quote that named a private person became a short Source line with an evidence class and no name. "(only our images)" became "Use a real photo of that person that the site owns and may publish." Why: the page is public, and it named someone who is not a cited author and used internal shorthand.

The raw chat quote on invites became a Source line with an evidence class, and the rule now points to the persuasion rule in Copy and Messaging Principles. The Why now cites banner blindness and the Fogg model, links the Nielsen heuristic, and marks its outcome claim as unmeasured. "(claim only what is live)" became "(only what the person can use today)". Source: https://www.nngroup.com/articles/banner-blindness-old-and-new-findings/ and https://behaviormodel.org/, read 2026-10-29.

A long spoken transcript became one Source line: owner judgement, 23 October 2026, on a course shelf; a main action should feel like moving forward, so it sits at the right or centred.

The intro to the seven sections from 20 October no longer uses internal role names or quotes the approval. It says they were proposed during design work on a recipe site and approved by its owner on 20 October 2026, that only that site has used them, and that none is user-tested. Why: the claim that "the rules are general" had no evidence class.

One home per action now holds its own example, How to apply text and Source, which had sat under Speed. A raw owner quote that nothing cited is removed, and the section states the owner's ruling plainly. The Hick's law sentence is dropped, since Conflicts and open questions lists it as unverified, and the section links the NN/g heuristics and mental models notes. Source: the Principles outline and the note "Source: NN/g", read 2026-10-29.

Empty states, Every action answers, Undo and Show progress each gained a Source line with an evidence class. Undo now cites Nielsen on overused confirmations and marks the undo length as a craft choice; the goal-gradient claim is cited to Kivetz, Urminsky and Zheng 2006; the progress example says "next badge" instead of product terms. Sibling sections are now linked. Source: https://www.nngroup.com/articles/confirmation-dialog/ and https://doi.org/10.1509/jmkr.43.1.39 (abstract), read 2026-10-29.

Milestones now call themselves one of two recorded exceptions to Motion only to confirm, and the peak-end citation notes that the study measured painful episodes. The defaults rule now covers settings only: GOV.UK and NN/g disagree on pre-selecting form questions, so a consequential answer stays empty until testing settles it, and review flags such pre-selection. The Hick's law reasoning is replaced by Johnson and Goldstein 2003. Why: two sections each claimed to be the only motion exception, and "every choice" contradicted GOV.UK. Source: https://doi.org/10.1111/j.1467-9280.1993.tb00589.x, https://doi.org/10.1126/science.1091721 (abstracts), https://design-system.service.gov.uk/components/radios/, https://www.nngroup.com/articles/checkboxes-vs-radio-buttons/, all read 2026-10-29.

In Every page has one job, a raw quote became an owner ruling Source line. The examples are framed "on a recipe site", and "keep your kitchen" became "manage your account and who shares it". Each of the six sections from the 21 October check gained a Source line, and sibling sections cited by name are now links. Why: product jargon and raw chat meant nothing to a public reader.

The section on third-party content is now split into Rule, Why, Quick test, Example and Exception. The Why cites the ruling of LG München I, case 3 O 17493/20, and says it is not legal advice. Motion only to confirm now cites W3C SC 2.3.3 for its health claim, notes that criterion is AAA, and lists both exceptions. A status line that would go stale is deleted, and the print claim is softened and marked unmeasured. Source: https://www.gesetze-bayern.de/Content/Document/Y-300-Z-BECKRS-B-2022-N-612 and https://www.w3.org/WAI/WCAG22/Understanding/animation-from-interactions.html, read 2026-10-29.

The copy heading became "Copy rules learned on one recipe site". The intro dates the rules, labels their evidence class as simulated reviewers and owner judgement on one site, not user-tested, and names Copy and Messaging Principles as their owner. Why: "They hold for any product" overstated a panel of two modelled reviewers on one site. Source: How to do UX and UI; AGENTS.md.

The brief now allows only the two capture widths, 390 by 844 and 1440 by 900, and the designer row measures at both, not three. Why: the capture step holds two widths, but the brief and that row pointed reviewers at a third. Source: the Capture once step, read 2026-10-29.

Capture once now names three views: visitor (signed out), editor (signed in, draft mode off) and draft (signed in, draft mode on), matching the brief and the draft mode lesson. The auditor names its two views as visitor and draft, as the Skill does. Why: the view names changed from block to block.

The auditor step now links the Skill, and the purpose adds: "To run a pass step by step, follow the Skill: run a site coherence pass." Why: the method named its Skill only as plain text.

The status line, two weeks out of date, now says all five changes were built into the site's code on 15 October 2026 and a browser test holds the header at 73 pixels at 1440 by 900; the effect on visitors is not measured. Source: the site's code, checked 2026-10-29.

The comparison with outside sites now links NYT Cooking (https://cooking.nytimes.com) and Ottolenghi (https://ottolenghi.co.uk), both measured on 14 October 2026. Why: the organization's evidence rule requires a URL. Read 2026-10-29.

The first run now says "not recorded" instead of "not yet recorded", and the purpose adds that neither whole-site pass recorded before and after audit numbers. A broader claim about reviewer judgement was left out, because the fit pass did record measured numbers.

The purpose's rule set now includes Principle 10: WCAG 2.2 A and AA are requirements, not calls. A new Accessibility item in the brief says every reviewer reports a failure it sees, and every change to colour, size, focus or motion is checked against the criteria before it goes to the owner. Why: no step checked accessibility. Source: Principles, Principle 10.

The validation section now cites the paper and dataset with read dates and licence, says UEyes recorded people looking at static screenshots on a lab monitor with no task, and explains the NSS score. The limits now say the prompt asks about two seconds while the score uses one, and that the vision model, prompt and per-page scores are not recorded, so a new model or prompt needs a new validation. Source: https://doi.org/10.1145/3544548.3581096, https://zenodo.org/records/8010312 (licence CC BY 4.0) and https://github.com/YueJiang-nj/UEyes-CHI2023, all read 2026-10-29.

The route for copy work now links "Principles: evidence-based website, product and agent-facing copy" in Copy and Messaging Principles. Why: a public reader could not follow an unlinked organization name.

The undefined terms "passer", "sentence approval" and "quick-glance list" are gone: the text now says whoever runs the pass, the old-to-new wording table, and a separate list for the owner to check at a glance.

The delivery paragraph now says that in the agent setup these runs used (October 2026), reviewer agents did not write report files and long messages were cut off, so the workaround applies only where that holds. The hard limits now read "change no code or content and start no servers of your own". Why: a property of one setup was stated as a rule of the method.

In the 16 October run, an internal code name became "sign-in", "the part picker" became "a shared recipe link that opens cook mode on choosing which part to cook", and "non-production stage" became "test deployment". Labels that visitors see on the site stay. Source: the site's code, checked 2026-10-29.

The claim that the method ran once now says it has run several times between 13 and 16 October 2026, all on one site, links the Runs section, and notes that no second site has tried it. Giving no number keeps the sentence true after the next run. Source: Site coherence pass, Runs section.

Step 1 now says to read Principles from "One focal path per screen" onward, instead of stopping at "Motion only to confirm". Why: Principles now continues past that section, and the method's purpose already says "onward".

The heatmap limits now name the UEyes eye-tracking dataset and link the validation in the Site coherence pass. The model and exact prompt were not added, because nothing in the organization records them. Source: Site coherence pass, Validation (14 October 2026).

"the usual cause of a crop through a person" now reads "one known cause". Why: there is only one worked example, so nothing backs a claim about frequency.

The heading is now "The toolkit". The body says the toolkit is being built and is not published as of 29 October 2026, so the audit and renderer should be built from the two tool sections; it links axe-core and replaces "keyed headless mode" with "a later version may run them without an agent". Why: the heading and body dates disagreed and the status was ten days old. Source: https://github.com/dequelabs/axe-core read 2026-10-29; a check on 2026-10-29 found the toolkit not yet published.

The cited post now carries its date and source, and "which the product needs to deliver its value" now reads "which he had heard the product needs to deliver its value", matching the post's own hedged wording. The direct post link was not found, so the text cites the account and says so. Source: post by @zeeg dated 25 September 2026, read 2026-10-29 through a public mirror; account https://x.com/zeeg.

The owner now approves every new sentence before it ships, and changes to the voice or a promise are listed separately so the owner can see them at a glance. Why: the internal process names this replaced were never defined.

"The founder judges." now reads "The owner decides.", matching the role name used everywhere else. An unattributed quote is now framed as an example ("for example "something that runs all my ads"").

The Steve Jobs line now quotes "a lot of times, people don't know what they want until you show it to them" and cites BusinessWeek, 25 May 1998. The Bloomberg page is paywalled, so the wording was confirmed on WIST. Source: https://www.bloomberg.com/news/articles/1998-05-25/steve-jobs-theres-sanity-returning and https://wist.info/jobs-steve/19780/, both read 2026-10-29.

The link text "Principle 4, keep essential actions available" now reads "keep frequent actions available". Why: the principle says "frequent", and "essential" changed its meaning. Source: Principles, Principle 4, read 2026-10-29.

A new entry, "Learning from use: owner rulings and a site-wide check, 19 to 23 October 2026", links each principle added on those dates and records the copy rules of 16 October, which no entry covered. It says the provider and model are not recorded. Why: the log stopped at 16 October while Principles kept growing, and every pass needs an entry. Source: the Principles version history, read 2026-10-29.

A verbatim owner chat quote is replaced with a paraphrase: "The first screen of a recipe site's landing page, which the owner judged good but overwhelming". Why: raw chat does not belong on a public page.

A new entry, "Review and corrections, 29 October 2026", records the method, provider and model, the changes per document and the documents changed elsewhere in the review. It states that three of the ten source notes (web.dev, WCAG 2.2, Shneiderman) were corrected and that the queued deepening pass is still owed for all ten. Source: the organization's document list with update dates, read 2026-10-29.

The 7 October dark mode pass is renamed "Pass 04: 7 October 2026, dark mode", since two headings both said Pass 02. A correction line also says the organization was private when passes 01 and 02 were written and is public now, so their "remains private" lines describe that time.

The rule at the top now asks for the agent, provider and model name. A correction says no earlier entry recorded provider and model and they cannot be recovered; the new entry names its own (Anthropic Claude Opus 5.5). Past entries stay as written, since the log is append-only. Source: AGENTS.md.

The Research log route now says "Record date, agent, provider and model name", so cross-provider review can be checked.

The Principles route now says "Keep each principle together under its own heading: its rule, evidence class, sources, application, and limits." instead of "one principle per block", which no longer matches the document. Source: Principles outline, read 2026-10-29.

The routing lead now begins: "Read the relevant Principles first. Before designing, follow How to do UX and UI; use Worked examples for a concrete decision, the Design review playbook to review a screen, and the Site coherence pass with its Skill for a whole site." Why: the reading order skipped the house method and the whole-site pass.

"The original setup describes improved agent work after reading those guides." now reads "The author of those private guides reported that agents worked better after reading them." Why: a public reader cannot find "the original setup".

The purpose now defines three verdicts (contradicted, unsupported, conditional), and every verdict starts with one of them. The three-click verdict is now contradicted, citing NN/g. References to a session and private guides became plain wording. The review date stays at 13 September 2026, because only one entry was rechecked at its source. Source: https://www.nngroup.com/articles/3-click-rule/ read 2026-10-29.

A dark mode entry under Source tensions covers the four disagreements from the 7 October pass: base darkness, contrast target, elevation and theme switch. It marks Apple as a native-platform convention and labels the synthesis as ours. Under Coverage gaps, a line says the dark mode principle and the Material, Apple and Radix source notes are still owed, after a recheck pass. Principles got no dark mode section. Source: https://developer.apple.com/design/human-interface-guidelines/dark-mode, https://www.nngroup.com/articles/dark-mode/, https://www.radix-ui.com/colors/docs/palette-composition/understanding-the-scale, read 2026-10-29; Material 2 and 3 and WCAG 2.2 as recorded in the Research log, read 7 October 2026 (the Material pages did not render on 29 October 2026).

The validation tension now states the direct disagreement: GOV.UK forbids validation on field exit, while Baymard recommends it for most fields, checks fixed-length entries at their length and clears errors on the next keystroke. The starting context names GOV.UK for government-style transactions and Baymard for commerce checkouts, and each product decides for its own context. Source: https://design-system.service.gov.uk/patterns/validation/ and https://baymard.com/blog/inline-form-validation, read 2026-10-29.

A second dark mode tension was added after the validation entry, saying Apple and Material are transfer evidence on a website, not requirements, with a proposed resolution and the open evidence. Why: the 7 October pass proposed these disagreements for this document, but they existed only in the Research log. Source: https://developer.apple.com/design/human-interface-guidelines/dark-mode and https://www.nngroup.com/articles/dark-mode/, read 2026-10-29.

References to "this pass" and "the starting guides" now read "The first pass (13 September 2026) did not verify" and "Earlier private summaries cited them". The status line adds that the validation timing and dark mode entries changed on 29 October 2026. The review date stays, because no full review took place.

The note now says GOV.UK tells you not to validate when the user leaves a field and to wait until they continue or submit. Its research exception covers only checking before the user has finished an entry, with the character count as the example. Why: the old text called field exit a default that research could relax. Source: https://design-system.service.gov.uk/patterns/validation/ and https://baymard.com/blog/inline-form-validation, read 2026-10-29.

The text input example link now carries "read 2026-09-13". Why: every URL needs its read date.

The tension now says GOV.UK does not validate on field exit, that its research exception covers checking during entry, and that the two sources disagree directly. Why: the old wording made a contradiction look like a difference of defaults. Source: https://design-system.service.gov.uk/patterns/validation/ read 2026-10-29.

The note now gives field exit as Baymard's trigger for most fields, with validation at the expected length as an extra option for fixed-length entries, and errors that disappear on the keystroke that makes the input valid. It adds the author, Edward Scott, and the date, 9 January 2024. Source: https://baymard.com/blog/inline-form-validation read 2026-10-29.

The note now says the guide recommends automatic tab activation when panels appear without noticeable latency, and manual activation when panels lag, because automatic activation then slows focus movement. It adds Up and Down for vertical tab lists. Why: "permits" had turned the recommended default into one of two equal options. Source: https://www.w3.org/WAI/ARIA/apg/patterns/tabs/ read 2026-10-29.

The header now spells out ARIA (Accessible Rich Internet Applications) and the Authoring Practices Guide (APG), with read dates 2026-09-13 and, for tabs, 2026-10-29.

The heading is now "What the maintained library holds". The body spells out HAX, summarizes all 18 guidelines, including social biases and change notices, and says the guideline list was checked but the individual patterns were not. Why: the heading claimed verification that the body denied. Source: https://www.microsoft.com/en-us/haxtoolkit/library/ read 2026-10-29.

The header now names its evidence classes, the access scope names the pages on research.google, HEART is spelled out with the authors' paper as source, and the Tuch study names its journal. Why: this was the only source note without an evidence class, and the Tuch study read as Google's own work. Source: https://research.google/pubs/the-role-of-visual-complexity-and-prototypicality-regarding-first-impression-of-websites-working-towards-understanding-aesthetic-judgments/ and https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/36299.pdf, read 2026-10-29.

Conflicts and open questions: two dark mode tensions had been added in parallel; the shorter duplicate was removed and the fuller entry, which the Research log links, was kept.

Conflicts and open questions: a new tension, "Where a forward action sits", sets the owner rule in Principles against GOV.UK's left-aligned form button, with the current resolution (forms follow GOV.UK) and the open evidence. Source: https://design-system.service.gov.uk/components/button/, read 2026-10-29.

Source: W3C WCAG 2.2: SC 1.1.1 Non-text Content is added to the criteria other documents rely on, since Principles now requires text alternatives for pictures. Source: https://www.w3.org/WAI/WCAG22/Understanding/non-text-content.html, read 2026-10-29.

Skill: run a site coherence pass: the shared brief step now names the visitor, editor and draft views, matching the Site coherence pass.

Worked examples: in E4, automatic tab activation is now "the recommended default" when panels appear without delay, matching the APG note. Source: https://www.w3.org/WAI/ARIA/apg/patterns/tabs/, read 2026-10-29.

Design review playbook: "a screen coherence pass" now reads "a site coherence pass over every screen" and links the 16 October run record. The pointer to Principles adds that the rule stays a candidate until a second site tests it. Why: Principles had already renamed the method and marked the rule a candidate. Source: Principles, "One picture style per role", read 2026-10-29.

Design review playbook: the smallest slot is now "a product cutout shown at icon size, such as an ingredient", "icon variant" became "small-size variant", and a new sentence says that when icons are a role with their own style, as line drawings were on that site after 16 October 2026, that style is tested at its smallest size instead. "Icon" joins the role list. Why: the section used "icon" two ways, so a reader could mix two styles in one role. Source: Principles, "One picture style per role" ("line drawing for icons and the map").

Worked examples: the E3 evidence clause now says GOV.UK waits for continue or submit and does not validate when a person leaves a field; it allows checking during entry only when user research shows that helps on balance, with its character count as the example. Baymard recommends checking on field exit. Why: the old text read GOV.UK's exception as covering field exit. Source: https://design-system.service.gov.uk/patterns/validation/ read 2026-10-29: "Do not validate when the user moves away from a field."

Evaluation pack: the WCAG 2.2 link in the common builder brief now carries "(read 2026-10-29)". Why: every link carries a read date.

Principles: the dialog sentences now say Apple puts the default button on the trailing side and Cancel typically on the leading side, which matches the rule, while Microsoft puts OK before Cancel, so Cancel sits to the right of OK, against the rule. A website dialog follows this rule and the Apple order, and neither convention cites a study. Source: https://developer.apple.com/design/human-interface-guidelines/alerts (JSON data endpoint) and https://learn.microsoft.com/en-us/windows/win32/uxguide/win-dialog-box (order: OK, No, Cancel, Apply, Help), read 2026-10-29.

Principles: "the kitchen's people" became "the people who share the account", "the host" became "the account owner", and "leaving the kitchen" became "leaving a shared account". Why: these product terms meant nothing to a public reader, and the same terms had already been removed elsewhere.

Principles: "a missing tag is the usual cause" became "on the site this rule comes from, a missing tag was the cause". Why: the section rests on one site-wide check of one recipe site (21 October 2026), so nothing backs a claim about frequency.

Principles: two group intro paragraphs without headings are removed, since each sat at the end of the section before it. Each of the six 21 October Source lines now reads: "a site-wide UX check of one recipe site, 21 October 2026, approved by its owner; the pattern showed up on several pages at once. Evidence class: a craft rule from review; not user-tested." Source: full read of Principles, 29 October 2026.

Principles: the remaining "lens" terms now use the plain wording chosen in this review, such as "three reviewers who each measured the page", "a reviewer reading as an editor" and "a reviewer comparing the pages with each other". One section that spells out its three lenses was left alone. Why: the document used two vocabularies for the same thing.

Principles: the Norman conceptual model sentence now reads: "Mental models: people hold one belief about how a product works, and a mismatch calls for changing the structure or explaining the difference better (see mental models); two homes for one job create that mismatch and blur which control does what." Why: the organization has not verified Norman's book, and the linked NN/g note does not mention him. Source: the note "Source: NN/g", read 2026-10-29.

Site coherence pass: in the lens table, the persona's job is restored to its earlier text ("Walks the site's main journeys as a real visitor: arrives from a social video, finds a dish, cooks it, saves it, comes back"), and the designer now measures "at both capture widths". Why: the second round's width fix had landed in the persona row by mistake. Source: the table's earlier versions; Capture once holds only 1440 by 900 and 390 by 844.

Site coherence pass: the review date is now 29 October 2026 instead of 14 October 2026, since the whole document was reread and edited in this review.

Skill: run a site coherence pass: step 1 now says "390 by 844 and 1440 by 900 widths, the two the capture set holds" instead of "390, 768 and 1440 widths". Why: the 768 width sent reviewers to a state the capture set never holds. Source: Site coherence pass, Capture once and the brief, read 2026-10-29.

Skill: run a site coherence pass: the PADDED flag now says "a video service may pad an upright video's landscape thumbnail with side bars" and links the Principles section "Pictures: never twice, never through a face, video frames in their own ratio". Why: it stated the padding as fact and called the bars blurred, with no source. Source: https://github.com/yt-dlp/yt-dlp/issues/9837 (read 2026-10-29), as cited in Principles.

Research log: the 29 October paragraph on changes made elsewhere now lists How to do UX and UI and three Conflicts entries (validation timing rewritten, forward action added, form pre-selection added). It says eight of the ten source notes were corrected, and only the NN/g and Tognazzini notes are unchanged since 13 September 2026, so the queued item's statement that the notes have not changed no longer holds. Source: update dates of each note, read 2026-10-29.

Conflicts and open questions: a new source tension, "Pre-selecting the answer to a form question: GOV.UK versus NN/g", gives the current resolution (defaults apply to settings; a consequential form answer stays empty until testing settles it) and links the Principles section. GOV.UK says "Do not pre-select radio options"; NN/g says "Always offer a default selection for radio button lists." Source: https://design-system.service.gov.uk/components/radios/ and https://www.nngroup.com/articles/checkboxes-vs-radio-buttons/, read 2026-10-29.

Conflicts and open questions: the status line now reads: "on 29 October 2026 the validation timing entry was revised, and the dark mode, forward action and form pre-selection entries and the dark mode coverage gap were added." Why: the dark mode entry was new, not revised, and two new entries were missing. Source: a read of the document on 28 October 2026, which showed no dark mode entry.

AGENTS.md: the Baymard entry now reads: "timing of inline feedback, and its direct disagreement with GOV.UK on validating when a person leaves a field." Why: the old "GOV.UK default" framing is the one this review rejected. Source: the note "Source: Baymard on validation timing", read 2026-10-29.

Research log: the 29 October corrections now add that the WCAG note covers SC 2.2.2 since 29 October 2026, which the 15 and 16 October entry said it did not. The log is append-only, so the earlier entry is corrected only here. Source: the note "Source: W3C WCAG 2.2", read 2026-10-29; https://www.w3.org/WAI/WCAG22/Understanding/pause-stop-hide.html.

The nine copy rules in Principles belong to the sibling organization Copy and Messaging Principles, which owns the words. They stay here, with a pointer to that owner, until a writer there merges them; then this section becomes one pointer line.

Principles kept two process paragraphs about how the 19 to 23 October rules were made. The Research log now records their provenance, and the third round removed them from Principles, so this item is closed.