Evidence must let an independent qualified reviewer identify the frozen target, reproduce the action and evaluate the decision without receiving credentials, secrets or unnecessary personal data.
Minimum record
Every formal result records the procedure ID and version, procedure checksum, catalogue version and checksum, target ID, product version, state, environment, tester identity, timestamps, actions, actual results, target outcome, execution state, applicable WCAG decision, evidence references, limitations and review state.
The browser worksheet begins with four tester-facing controls: target, optional scenario, result and observation. The interface maps the single result choice to separate execution-state and rule-outcome fields without asking testers to understand the schema. Evidence, environment, expectation, attestation, WCAG-decision, limitation and formal-completion fields are secondary. A shorter initial form does not reduce the formal execution schema or the evidence required to validate a completed record.
Agent records use accessibility-agent-execution.v2. They bind the catalogue, procedure, plan and embedded browser-probe evidence with SHA-256 checksums, distinguish the bounded decision from the overall run blocker, and include an independent agent-verification result. They remain not_ready for formal claims. Hash validation and agent verification establish recorded-scope integrity and internal decision consistency; they do not turn diagnostic evidence into real-platform or participant evidence.
Real-platform records use accessibility-platform-execution.v2. They additionally bind the platform matrix, exact environment inventory, current calibration fixture, frozen multi-target inventory, action log and independently agent-verified evidence files. The validator rejects missing files, unsafe paths, hash or byte-length mismatches, simulated coverage, placeholder versions, environment drift, missing physical-device identity, same-agent execution and verification, and stale evidence. These controls do not make an unintegrated platform real or prove a false attestation.
Method-specific evidence
- Semantics: DOM or tag-tree extract, accessibility-tree extract where useful, computed accessible name and relevant state.
- Contrast: sampled foreground and background values, location and state, calculation method and ratio; retain the separate applicability judgement.
- Zoom and reflow: starting viewport, zoom level, resulting CSS-pixel viewport, orientation and evidence of information and functions after reflow.
- Keyboard: exact keys, starting point, focus sequence, revealed state, visible focus and activation result.
- Screen reader: exact product, version, browser, operating system, mode, commands and announcement or speech-history output.
- Mobile: physical device model, operating-system version, orientation, settings and gesture sequence.
- PDF: immutable file checksum, PDF version, PDF/UA identifier, authoring and remediation tools, checker export and manual tag-tree, order, keyboard, reflow and assistive-technology observations.
- User evaluation: approved protocol, relevant participant-characteristic summary, consent basis, tasks, facilitation, observations, limitations and protected evidence location.
Evidence references
Store references or content-addressed identifiers in execution JSON. Do not embed large or sensitive files. Use SHA-256 checksums for immutable documents, fixture packages and exported tool reports where integrity matters. Record redactions without destroying the reviewable relationship to the original controlled evidence.
Review rules
A screenshot alone rarely proves programmatic exposure or an interaction sequence. A checker report alone does not prove accessibility or PDF/UA conformance. A passed target must have evidence for every applicable expectation; a failed target identifies the unmet expectation. inapplicable and cantTell also require evidence and a reason. The browser worksheet derives a single atomic expectation result from the recorded target result; composite procedures expose their separate expectation controls.
Retention
Apply the engagement retention schedule, access controls and deletion process described in Privacy and security. The repository is not an evidence vault.