Conflicts and open questions: UX and UI evidence boundaries
Status: live register of source tensions, transfer hypotheses, and evidence gaps. First pass: 13 September 2026; 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. No disagreement is resolved by silently replacing one authority with another. A research tension is not automatically a decision for a product owner; a concrete product trade-off belongs with its product owner.
Source tensions
Discoverability versus a smaller initial display. Tognazzini warns against hidden controls; NN/g recommends selective disclosure with a findable secondary path. These are competing risks rather than a demand to choose one slogan. Proposed resolution: retain frequent actions and meaningful entry points, then test whether essential less frequent tasks remain findable. Open evidence: task-specific discovery and efficiency observations. Sources: Tognazzini's limits and progressive disclosure.
Single-screen forms versus staged questions. Shneiderman's memory discussion favors keeping a long form together. GOV.UK starts from one question per page but explicitly permits related questions together when research supports it. Proposed resolution: group information needed for comparison, and stage independent work when that helps the actual audience. Open evidence: novice completion and repeated expert use, including error recovery and Back behavior. Sources: single-screen guidance and qualified staging guidance.
Historical speed slogans versus modern metrics. A 50-ms acknowledgment target and a Core Web Vitals threshold measure different things. This is primarily a scope mismatch, not an empirical contradiction. The earlier 400-ms shortcut has not been re-established as a general completion rule. Proposed resolution: name the measured interval, percentile, device context, and meaningful result. Sources: historical advice and modern responsiveness scope.
Validation on submission versus on leaving a field. GOV.UK's service pattern says not to validate when the user leaves a field, and to wait until they try to continue or submit. It allows checking a field before the user has finished entering it only when user research shows that solves more problems than it causes. Baymard's commerce research recommends checking each field when the user leaves it, checking fixed-length entries such as postal codes once they reach their length, and removing an error on the keystroke that makes the input valid. So the two sources disagree directly on the field-exit trigger. Both warn against errors that appear before the user could have finished an entry. Proposed decision method: start from the source closest to the service being built: GOV.UK for a government-style transaction, Baymard for a commerce checkout. Before adopting an earlier trigger, compare correction success, interruption, and assistive-technology behavior. Test the slow-typing and already-corrected states explicitly. Each product decides this tension for its own context; no general ruling is needed. Sources: GOV.UK: Recover from validation errors, read 2026-10-29, Baymard: Usability testing of inline form validation, read 2026-10-29, and Baymard validation research.
Dark mode: how dark, how much contrast, how to raise surfaces, and whether to offer a switch. The dark mode pass of 7 October 2026 found four disagreements. Base darkness: Material 2 recommends a dark grey and allows true black only to save power on OLED screens; Radix suggests step 1 or 2 of a grey or tinted scale for a dark app background; Apple's text gives no value. Contrast target: WCAG 2.2 AA asks 4.5:1 for body text and 3:1 for interactive boundaries; Apple sets 4.5:1 as a minimum and asks designers to "strive for a contrast ratio of 7:1" for custom colours, especially small text; Radix states contrast in APCA, which does not map one to one to WCAG ratios. Elevation: Material 2 keeps shadows and adds a lighter overlay per level, Material 3 separates surfaces by tone first, and Apple swaps between base and elevated background colour sets. Theme switch: Apple says to avoid an app-specific appearance setting; NN/g strongly recommends letting people switch to dark mode, and Material 2 endorses a switch. Proposed resolution, our synthesis: a website follows the system setting by default and meets WCAG AA in both themes; a darker base, a higher contrast target or an in-page switch is a product choice to test, not a general rule. Apple's guidance is a native-platform convention, and its transfer to websites is not yet stated. Sources: Apple HIG, Dark Mode, NN/g, Dark Mode vs. Light Mode and Radix Colors, Understanding the scale, read 2026-10-29; Material Design 2, Dark theme, Material 3, applying elevation and WCAG 2.2, read 2026-10-07 (the Material pages did not render for a reread on 29 October 2026). The full source list is in the dark mode pass of the Research log.
Where a forward action sits: owner judgement versus GOV.UK for forms. Principles places an action that moves the reader on at the right of its group or centred under what it acts on; that rule rests on owner judgement (23 October 2026), not on a study. GOV.UK tells services to align the primary button to the left edge of a form. Current resolution: cards, sections and dialogs follow the owner rule; forms follow GOV.UK (Principles, Calls to action point forward). Open evidence: whether a right-aligned or centred forward action helps people outside forms on a website; no source here has tested it. Source: GOV.UK button, read 2026-10-29.
Pre-selecting the answer to a form question: GOV.UK versus NN/g. GOV.UK says not to pre-select a radio answer, because people then miss the question or submit the wrong answer; Nielsen Norman Group advises a default in every radio list. Current resolution: Principles keeps defaults for settings and leaves a consequential form answer empty until testing settles it (Principles, Sensible defaults). Open evidence: error and completion rates with and without a pre-selected answer on a website form. Sources: GOV.UK Radios and NN/g, Checkboxes vs. Radio Buttons, read 2026-10-29.
Transfer questions that need product evidence
Visual simplicity versus comparison efficiency. Screenshot preference findings do not decide the right density of a repeatedly used working view. Hypothesis: compare first-impression comprehension and actual task completion separately; do not select a card layout solely because it looks calmer. Unresolved: which density serves each real task and audience. Source boundary: aesthetic-study limits.
Human control versus repeated interruption. Review can expose important consequences but may interrupt routine work. Hypothesis: calibrate review to consequence, recoverability, and scope. Unresolved: which checks help users detect real errors without unnecessary operating work. No evidence here supports a universal confirmation count. Basis: Human-AI control tension.
Conversation versus structured interaction. The proposed transfer uses conversation for open-ended goals and structured views for precise comparisons or change review. This is not a finding established by the 2019 Human-AI guidelines. Next evidence: compare representative tasks with real users; preserve outcome quality and mistake detection, not only preference. Basis: explicit website application.
Coverage gaps and next-pass priorities
Full research methods are not yet reviewed. The Google HEART record, the screenshot-study record, and the CHI 2019 Human-AI record were read at publication-record and abstract level. Their raw data, full methods, effect sizes, and replication history were not audited. Any stronger quantitative recommendation must first review the relevant full text. Sources: Google access scope and Human-AI access scope.
Original authorities and psychology need a separate pass. The first pass (13 September 2026) did not verify the original books of Norman, Krug, Cooper, Rams, or Wroblewski, or the original Fitts, Hick, Miller, and Doherty research. Earlier private summaries cited them; that is provenance, not proof of a fresh primary-source review. Do not introduce numerical UI rules from those summaries. Native-platform guideline updates remain outside the current website scope.
Dark mode has no principle or source notes yet. The dark mode pass of 7 October 2026 found sourced rules (a designed dark palette rather than an inverted one, 3:1 boundaries in both themes, following the system setting and declaring color-scheme), but they live only in the Research log. A later pass rechecks those sources, writes a dark mode principle, and adds source notes for Material, Apple and Radix. Until then, read that log entry for dark mode work.
Independent corroboration and measured capability improvement remain outstanding. The recorded preparation passes use one model family from one provider in one session, not independent or cross-provider replications. Another authorized provider should recheck a consequential source claim. The Evaluation pack now supplies a paired protocol, six disclosed tasks, and scorecards, but no trials have run. Its tasks are not hidden holdouts because this organization exposes them to consuming agents. Record source corroboration and measured product-task results separately. No recurring schedule is installed.
No product usability study has been performed here. These documents provide methods and evidence boundaries. They do not establish that any particular site satisfies the principles. A consuming agent should record the actual task, audience, observations, and remaining uncertainty with that product's work.