Design QA Evidence Pack

Verify the scoped implementation and produce evidence that another reviewer can reproduce. Report findings without changing the target.

Enforce the boundary

• Work read-only against the target, repository, design, data, and configuration.
• Do not edit files, generate patches, fix defects, refactor code, change tokens, or update content.
• Do not install dependencies or run commands that write build output, caches, snapshots, or lockfiles into the target.
• Use an existing URL, running environment, static file, or already-built artifact.
• When build verification is required, use a disposable copy or tool-supported output and cache paths outside the target. Record the exact command and disposable location.
• Allow only transient, non-destructive UI interactions. Avoid submissions, uploads, deletions, purchases, messages, or any action that changes persistent data.
• Store screenshots only in a temporary or tool-managed evidence location outside the target when capture is available.
• Mark a check Blocked when verification requires mutation, missing access, unavailable tooling, absent requirements, or an unusable environment.
• Stop after the evidence pack. Do not include repair instructions or implementation recommendations.

Establish the test contract

Collect:

• target URL or artifact;
• build, commit, version, or timestamp when available;
• routes, components, user states, and interactions in scope;
• approved reference and acceptance criteria;
• required browsers, devices, viewports, themes, locales, and motion preferences;
• exclusions and known limitations.

Use 1440 × 900, 834 × 1194, and 390 × 844 as coverage defaults only when the user supplies no viewport set. Label them as test defaults, not product requirements.

Do not invent acceptance criteria. Mark subjective or unspecified comparisons Blocked and identify the missing approval source.

Build the matrix

Create one check for each criterion × route/component × state × viewport combination that matters. Assign stable IDs such as V-01, R-03, or I-02.

Read qa-evidence-schema.md for the status rubric, evidence fields, and coverage checklist.

Prioritize:

• reference fidelity;
• hierarchy, spacing, alignment, crop, and token use;
• responsive recomposition and overflow;
• required interaction states and recovery;
• motion sequence, continuity, interruption, and reduced-motion behavior;
• text clipping, asset failures, and console errors;
• keyboard reachability, focus visibility, semantics, and obvious contrast failures.

Keep factual content verification outside this skill unless the acceptance source contains the exact expected text. Verify presence, truncation, and placement; do not research truth claims.

Execute reproducibly

1. Record the environment, viewport, route, initial state, and timestamp.
2. Reset to a known state before each check when reset is non-destructive.
3. Perform one scoped interaction sequence at a time.
4. Capture the smallest evidence that proves the result:
   • screenshot or frame;
   • DOM, computed-style, bounding-box, or overflow measurement;
   • console entry;
   • focus order or keyboard trace;
   • exact observed text or state transition;
   • reference comparison with both sources identified.
5. Record expected and observed outcomes separately.
6. Preserve raw observations. Label interpretations.

Never use visual similarity alone to verify behavior. Never use a working interaction alone to claim visual fidelity.

Assign one status

• Use Pass only when direct evidence satisfies the stated criterion.
• Use Fail only when direct evidence contradicts the stated criterion.
• Use Blocked when the check cannot be completed or judged.
• Use NA when the criterion does not apply to the scoped surface or state; include the reason.

Do not infer Pass from absence of an observed failure. Do not average multiple checks into one status. Split mixed outcomes into separate checks.

Return the evidence pack

Return sections in this order:

1. Test identity — target, version, environment, time, and evidence location.
2. Scope and acceptance sources — tested and excluded surfaces.
3. Status summary — counts for Pass, Fail, Blocked, and NA.
4. Evidence matrix — ID, criterion, context, expected, observed, evidence, and status.
5. Failures — reproduce each failed observation without prescribing a repair.
6. Blocked checks — name the missing access, artifact, requirement, or environment.
7. Coverage gaps — disclose untested routes, states, viewports, or preferences.
8. Read-only statement — confirm that no target changes or fixes were made.