Research log: UX and UI evidence passes

Append-only. Each pass records its date, agent, provider and model name, sources, verified earlier claims, changes, limitations, and handoff. Corrections are new entries that identify what they correct. The Principles document holds current synthesis; this log preserves how it changed.

Pass 01: 13 September 2026

Agent and model: one research agent. Mandate: research UI and UX evidence and place it in this organization according to its guidance. This is one model and one research session, not independent multi-provider review.

Starting state: the organization catalog contained only AGENTS.md. Its routes requested Principles, authority-specific Sources, Conflicts, Myths, and this log. Earlier private guides on UX and on UI were the prior synthesis and provenance. Their historical summaries were not treated as fresh primary evidence. The English documents are a new distillation, not wholesale copies.

Source coverage

Eight maintained source notes were created, each with original URLs, the date read, evidence class, application, and limitations: Nielsen Norman Group; Shneiderman; Tognazzini; W3C WCAG 2.2; Google web.dev; GOV.UK Design System; Google Research; and Microsoft Human-AI and HAX.

Earlier claims checked against the primary publisher

WCAG target sizes: corroborated the earlier correction. AA's minimum criterion uses 24 by 24 CSS pixels with exceptions; 44 by 44 belongs to enhanced AAA. The original blanket 44-pixel AA shortcut should not return. Verification: AA criterion and AAA criterion.

Core Web Vitals: corroborated the recorded thresholds and clarified completion. The publisher's current definitions support the stated good thresholds and percentile basis. INP does not cover the complete asynchronous transaction. Verification: thresholds and measurement boundary.

Shneiderman's second rule: corrected edition ambiguity. The starting guide's shortcut-focused wording is not the current author-page formulation, which is universal usability. The source note now identifies the sixth-edition formulation. This is a version distinction, not a claim that the older formulation never existed. Verification: current author-published rules.

Five-user advice: corroborated the qualitative and segment qualifications. It is not a completeness guarantee or an uplift sample plan. Verification: qualitative testing scope.

Synthesis and disagreements

Created Principles with 15 separately addressable principles, source references, scope, and evidence labels. Created Myths to prevent unsupported shortcuts, and Conflicts and open questions to retain discoverability versus disclosure, single-screen versus staged questions, and transfer questions about density and AI review. No product-specific trade-off has been ratified.

Limitations and handoff

Google's two research records and Microsoft's CHI publication were read at publication-record and abstract level. Full methods, raw data, effect sizes, and replication were not audited. Original psychology papers and several books from the starting guides remain uncovered. There was no live website usability test, independent-provider review, or legal-conformance assessment. These gaps and next-pass entry conditions are recorded in coverage gaps.

All eight source notes and the three synthesis owners returned confirmed writes and were read back by exact ID for outline and reference-key verification. Routing closure is recorded below after the entry point and originating references are updated. The organization remains private; this pass requests no publishing, membership change, or recurring schedule.

Pass 01 verification, 13 September 2026: The catalog contains AGENTS.md and all 12 research documents. All 12 are linked from AGENTS.md. Each new document was read back by its exact ID. The updated Principles navigation and both earlier guide introductions were also read back. The earlier Shneiderman summaries were corrected. Product-specific design owners were unchanged. This verifies storage and routing, not independent user testing or full-paper review. The evidence gaps above remain open.

Pass 02: Applied design guidance, 13 September 2026

Agent and mandate: the same research agent, in the same session. The question was what examples, evergreen rules, or other additions would make the organization more worthwhile. This is a follow-up research pass by the same model family, not an independent review.

Completed: added Baymard validation guidance and W3C APG tabs and dialogs. Extended the existing GOV.UK source with labels, comparison, and validation timing, and the WCAG source with reflow and text-spacing resilience. Original publisher pages and selected reference examples were read. The prior error-summary claim about focus and linked field errors and the AA target-size distinction were rechecked against their first-party pages.

New tension: validation timing retains the GOV.UK and Baymard defaults separately. No universal trigger or new product decision was ratified.

Recommended next development, not completed deliverables: a compact worked-example library, conditional decision guidance, a visual-craft review method, and an evaluation comparing design work with and without this organization. These should demonstrate applicability and limitations rather than add disconnected slogans or a large screenshot collection.

Write limitation: the agent's tool safety checks blocked two worked-example document creation drafts before execution; no reason was supplied. No example document was created and no alternative storage route was used for those drafts. A catalog check after the first block showed only the two new source documents in addition to the existing corpus. The examples remain recommendations, not delivered research artifacts.

Verification: both new source documents returned confirmed writes and were read by exact ID. AGENTS.md links both; their routes were checked. The added GOV.UK label guidance, WCAG text-spacing block, and validation conflict were read back. No live component trial, user study, cross-provider review, or measured capability improvement was performed. The organization remains private.

Pass 03: Worked examples, review method, and evaluation preparation, 13 September 2026

Agent and mandate: the same research agent. The proposed practical additions were accepted and their completion requested. This is a further preparation pass in the same session, not independent-provider corroboration.

Delivered: Worked examples contains eight original fictional cases, counterexamples, acceptance checks, fixture data, and textual comparison exhibits. Design review playbook contains six conditional evergreen prompts, visual-craft checks, a complete-journey procedure, finding and prioritization guidance, and a correction-to-learning writeback method. Evaluation pack contains a with/without-corpus protocol, six disclosed practice briefs, fixed data, a common accessible-prototype brief, scoring, cost and correction records, and contamination boundaries. These three documents now exist; the earlier missing worked-example deliverable was not left as a recommendation.

Primary-source rechecks: verified GOV.UK's persistent-label guidance and table use, the existing GOV.UK/Baymard validation-timing disagreement, and W3C APG's latency-dependent tab activation and modal focus guidance. Re-read W3C reflow and text-spacing explanations, NN/g's visual principles, web.dev's INP boundary, and the Google HEART publication record/abstract. Original URLs and the read date are beside relevant example claims. Sources: GOV.UK text input, table, validation, Baymard validation, W3C tabs, modal dialog, reflow, text spacing, visual principles, INP, and HEART record, all read 2026-09-13. The manual-tabs source example was read as documentation, not executed.

What changed: AGENTS.md routes the three new owners and the correction-to-learning method. Principles now references the playbook instead of maintaining another copy of the review procedure. The original anecdotal claim about improved agent work is labeled as motivation, not a measured effect size. Conflicts now points to the prepared evaluation while keeping independent review and actual trials outstanding. Existing standards and validation tensions remain unchanged.

Verification: all three creations returned writeConfirmed. Each document was read back by exact ID, including the refined evaluation text. The live content contains eight numbered example sections and six practice task sections. Each new document belongs to this UX/UI organization and is linked from its AGENTS.md. The method and evaluation are original proposed applications, not additional externally validated findings. No native-app scope, product brand, membership, or publication setting changed.

Limits and receiving step: the examples are text exhibits, not rendered mockups, screen recordings, implementation-tested components, or user-observed redesigns. No paired agent trials or user studies ran. The next operator can use the review method on actual work immediately; the paired evaluation requires fresh sessions, an independent reviewer, a fixed resource budget, and an artifact-testing environment. The experiment is specified but no runner, spending, recurrence, or held-out task store is installed. A request to operate a product follows that product's own authority, not this research record.

Pass 04: 7 October 2026, dark mode

Agent and model: one research agent, with one research subagent of the same model. Mandate: research what good UX and UI is for a UI pass over the screens of a product, with dark mode as the main gap. This organization held no dark mode guidance, so this pass adds evidence for it. One model, one session; not independent multi-provider review.

Sources read on 7 October 2026: Material Design 2, Dark theme (design-system documentation); Material 3, color roles and applying elevation (design-system documentation); Apple HIG, Dark Mode and Color (platform guideline); Radix Colors, Understanding the scale (design-system documentation); WCAG 2.2 criteria 1.4.3, 1.4.11, 2.4.11 and 2.4.13 (normative standard); MDN, color-scheme and prefers-color-scheme (platform documentation); web.dev, prefers-color-scheme and color-scheme (practitioner articles); NN/g, Dark Mode vs. Light Mode (practitioner research summary); Trimble Modus v1, Dark mode (archived design-system documentation, on halation).

What the sources support, in our words:

Disagreements, proposed for Conflicts and open questions: how dark the base is (Material 2 dark grey, Radix tinted near-black, Apple gives no value in its text); the contrast target (WCAG AA 4.5:1 against Apple's "strive for 7:1"; Radix states contrast in APCA, which does not map one to one to WCAG ratios); elevation (Material 2 keeps shadows plus overlays, Material 3 tone first, Apple a colour swap); an in-app theme switch (Apple says do not offer one; Material 2 and NN/g endorse letting people switch; a web app reasonably defaults to the system setting).

Earlier claims rechecked: the WCAG 2.2 levels of the criteria cited here, against the Recommendation of 12 December 2024 (1.4.3 and 1.4.11 AA, 2.4.11 AA, 2.4.13 AAA). Principle 10 needs no change.

Applied to: a product pass that measures one product's dark palette against these rules. That is product-specific; no product brand value is imported here.

Not done and receiving step: no Principles text changed and no Source note per authority was created; adding a dark mode principle and the Material, Apple and Radix source notes is proposed for a later pass. Apple's exact dark values, Material 3's dark tones and a primary vision-science source on halation were not found in this pass.

Learning from use: site coherence pass, 13 and 14 October 2026

Agent and model: one coordinating agent and its lens subagents, all on one model; this entry was written by a writeback agent in the same session on 14 October 2026, after the site owner approved moving the general lessons here.

What happened: two review panels (four lenses, then three including a consistency auditor that measured computed styles) and an attention heatmap validation, on one recipe and cooking-lessons website. Parallel builders built the findings between the panels.

What changed here: Principles gained ten principles, from "Recognise, don't recall" to "Motion only to confirm", each with a quick test and an anonymised example. New owners: the Site coherence pass method and the Skill: run a site coherence pass. The Design review playbook, the modelled minds method and AGENTS.md now point to them.

Evidence class: owner judgement, modelled lens panels and a computed-style audit of one site. No user observation. The heatmap section cites one validation run on 30 UEyes web pages (CC BY 4.0): the model beat centre bias and lost to the dataset's mean fixation prior.

Not done: no earlier claim was rechecked against a primary source in this writeback; that check is owed by the next pass. Before and after audit numbers for the built fixes are not yet recorded. The audit and heatmap tools are being generalised into a reusable toolkit.

Learning from use: one illustration style for a whole site, 14 October 2026

What happened: a recipe and cooking-lessons website replaced three generated picture families (near-photographic cutouts, soft painted thumbnails, busy painted scenes) with one style, chosen by the owner from candidates shown inside the real page, then regenerated all 118 pictures from one written prompt template and swapped them as drop-in files.

What changed here: the Design review playbook gained the section "Choosing and keeping an illustration style": one style per site, choose in place, write the style as a template, test at the smallest size first (icon variant), no words in pictures, and how to keep it consistent over time.

Evidence class: implementation-tested on one site; owner judgement plus checks at 48, 40 and 32 px on light and dark themes. No user observation and no outside source.

Not done: no earlier claim was rechecked against a primary source in this writeback. A second site should test the method before any part of it moves into Principles.

Learning from use: fit pass on one first screen, 14 October 2026

Agent and model: one coordinating agent and three lens subagents, all on one model. The first screen of a recipe site's landing page, which the owner judged good but overwhelming, was measured by three lenses on the live page and five reference cooking sites; changes were tried as injected CSS and the owner accepted all five. What changed here: the Site coherence pass gained the section fit pass and its Skill a step; Principles gained "Quiet chrome before content" (chrome before content, the pale band that reads as a toolbar, wordmark below the headline) and two rules under "One focal path per screen" (count the voices, accent budget). Evidence class: measurement of one site plus owner judgement, not user-observed. Not done: no earlier claim was rechecked against a primary source; the effect on visitors is not measured.

Learning from use: screen pass and hero motion, 15 and 16 October 2026

Agent and model: one coordinating agent with lens and builder subagents, all on one model. On the same recipe site: two fit passes in cook mode (15 October), a fit pass on the moving landing hero and a pass over every screen with 116 screenshots and three lenses (16 October); the owner delegated the picks. What changed here: Principles gained "A motion budget for the first screen", "Session controls stay where the reader looks", "The word first, the mark after it", "One picture style per role", a rule "One action row per screen" and a rule "One label per destination", and its picture rule now prevents a repeat by choosing a different item; the Design review playbook's "One style per site" became "One style per role"; the Site coherence pass gained a capture-once step, a three-lens screen pass, slice-per-pull-request building, delegated picks, a motion step in the fit pass, lessons and three runs; How to do UX and UI gained capture once in its panel input and a delegation paragraph. Evidence class: owner judgement, modelled lenses and measurement of one site, not user-observed. Rechecked: the five second threshold of WCAG 2.2 success criterion 2.2.2 was applied from the criterion as the team knows it, not reread at the source in this writeback; the WCAG source note does not yet cover 2.2.2. Not done: no before and after audit numbers; the effect on visitors is not measured.

Learning from use: owner rulings and a site-wide check, 19 to 23 October 2026

Agent and model: coordinating agents in design sessions on one recipe and cooking-lessons website, with the owner ruling on each change; provider and model name not recorded. This entry was written on 29 October 2026 from the dated source lines in Principles and its version history, because no entry was written at the time.

What happened: reviews of that site's screens, each change ruled on by the owner, and on 21 October 2026 a site-wide UX check for patterns that showed up on several pages at once.

What changed here: Principles gained One home per action on 19 October, with its exception for a shortcut to the same control the same day; Sell at the few moments that grow the product and Make it look like what it is on 20 October, then seven sections from Empty states show what will be there, and the first step to Speed is part of the design, proposed by the coordinating agent and approved by the owner the same day; Every page has one job, and its first screen does it on 21 October, then six principles from the site-wide check, from A browse link promises a shelf to Review first screens as a visitor sees them, with the owner's yes the same day; One visible action per element and Calls to action point forward on 23 October. Principles also gained Copy rules on 16 October, which the entry for 15 and 16 October does not list.

Evidence class: owner judgement and modelled review on one recipe website; some sections cite general heuristics. Not user-observed.

Not done: no earlier claim was rechecked against a primary source; no before and after numbers; no second site has used these rules.

Review and corrections, 29 October 2026

Agent and model: a review of every document in this organization on 28 and 29 October 2026; a finding was fixed only when at least two of three independent checks upheld it. This entry and the changes listed under "This agent changed" were written by one fix agent on Anthropic Claude Opus 5.5. The Review log: 2026-10-29 lists every finding and change.

Method: one review per area against the owning documents and, where a claim was rechecked, against the primary source, read 2026-10-29.

This agent changed:

Other sessions of the same review corrected Principles (including moving the blocks of "One home per action" back under their heading), How to do UX and UI, the Design review playbook, Worked examples, the Evaluation pack, the Site coherence pass, its Skill, three entries in Conflicts and open questions (the validation timing tension was rewritten, and tensions on where a forward action sits and on pre-selecting the answer to a form question were added), and eight of the ten source notes (web.dev, WCAG 2.2, Shneiderman, GOV.UK, Baymard, W3C ARIA Authoring Practices, Google Research, Microsoft HAX). Only the Nielsen Norman Group and Tognazzini notes are unchanged since 13 September 2026, so the queued block's statement that the notes have not changed no longer holds. The queued deepening pass below is still owed for all ten notes.

Corrections to earlier entries: the dark mode pass of 7 October 2026 is the fourth numbered pass; its heading said 02. The organization was private when passes 01 and 02 were written on 13 September 2026 and is now public, so their "remains private" lines describe that time. No entry before this one recorded the provider and model name, and they cannot now be recovered; read "one model" in those entries as one model whose name is unknown. The entry for 15 and 16 October says the WCAG source note does not cover success criterion 2.2.2; since 29 October 2026 it does, with the criterion reread at its Understanding page.

Earlier claims rechecked, read 2026-10-29: NN/g on the three-click rule still reports no published data for the rule and a study that found no rise in dropoff after three clicks. Apple's Dark Mode guidance still says to avoid an app-specific appearance setting and to strive for 7:1 with custom colours; NN/g still recommends letting people switch to dark mode; Radix still states contrast in APCA. The Material 2 and Material 3 pages did not render for a reread.

Still open: a dark mode principle and source notes for Material, Apple and Radix; the deepening of all ten source notes; an independent review by an agent from another provider.

Queued passes

Queued: deepen the ten source notes from 13 September 2026

Queued 27 October 2026; not run yet. The ten notes in the AGENTS.md source directory (Nielsen Norman Group, Shneiderman, Tognazzini, WCAG 2.2, web.dev, GOV.UK Design System, Google Research, Microsoft HAX, Baymard, W3C ARIA Authoring Practices) were all written on 13 September 2026, at 1 to 3 thousand characters each, and have not changed since. Goal: for each note, reread the primary sources and deepen what was read, with URLs and dates, the evidence class, transfer limits and tensions, so that Principles can lean on the note rather than on memory. Method: ordinary research passes as AGENTS.md describes them (one agent, one model, one session; several notes per pass are fine), each verifying at least one earlier claim and recorded here; disagreements go to Conflicts and open questions. Run at the next research pass in this organization, or before a Principles change that rests on one of these notes.