Audit a document collection for accessibility barriers and regressions

Proposed inputs

1. Name

   Bounded publication set

   Owner

   Document owner

   Minimum scope

   A few non-sensitive copies plus a manifest of expected files.

   Quality check

   Record identifiers, versions and exclusions; avoid recursively scanning unrelated folders.

2. Name

   Prior findings

   Owner

   Accessibility reviewer

   Minimum scope

   The earlier file versions and their observation locations, if a comparison is wanted.

   Quality check

   Ensure that a supposed fixed issue refers to comparable content, not a renamed or omitted document.

3. Name

   Review target

   Owner

   Qualified reviewer

   Minimum scope

   The intended audience, format-specific test scope and rules that can actually be checked.

   Quality check

   Separate automated structural checks, manual inspection and user testing.

Access and setup — proposed or unknown

Proposed minimum

1. Read-only originals with separate output storage; a small manifest must be approved before collection-wide work.

2. Normalize specialist results to one agreed schema before comparing them, and retain failed/unsupported scans as distinct states.

Compatibility unknowns

1. Unknown: installed scanners, rule-version consistency, encryption support, exact format coverage and report-renderer behavior.

External actions requiring approval

1. Uploading documents, expanding the folder scope, modifying any original or sending a compliance statement.

Why use this

1. A lower issue count can look like progress even when a file failed to open or disappeared from the scan. Comparing only headline totals hides this distinction.

2. Grouping verified patterns may help an owner decide whether to fix a shared template or one document. The report should preserve the evidence needed for that decision rather than turn a tool score into a claim of accessibility.

Why an agent / simpler alternative

Useful judgment

Connect format-specific observations to their likely reader impact and distinguish repeated authoring patterns from coincidental failures.

Simpler alternative

Use the native document application's accessibility checker and manual review for a small, stable set.

Unsuitable when

The requested result is a certification, a complete compliance determination, or an assurance of usability without assistive-technology testing.

Expected deliverable and review checks

Proposed artifact

A reviewer queue with file/version/location, observed barrier, evidence, confidence, check coverage and proposed next inspection.

Critical checks

1. Every expected file has a result state, including not opened and not tested.

2. A disappearing finding is labelled fixed only when equivalent content and the relevant check were actually examined again.

3. Suggested template fixes remain hypotheses until shared structure is confirmed.

4. Zero findings, score improvements and a successful parser exit are never treated as proof of conformance.

Missing output evidence

No document scan, assistive-technology session or comparison run was performed.

First check — proposed, unexecuted

Status

proposed_unexecuted

Synthetic fixture

Three original mock documents: one contains a deliberately unlabelled figure, another has a repeated structural problem, and the third is deliberately unreadable. A second snapshot fixes one problem but excludes a file.

Acceptance checks

1. Unreadable and missing files remain explicit; neither becomes a pass.

2. Only the re-examined changed observation can become a proposed fix; exclusions block a complete before/after claim.

3. Findings include reproducible locations and the report carries no compliance certification.

Stop conditions

1. The tool silently drops a file, overwrites an original or cannot preserve unknown states.

Failure signals and proposed responses

1. Signal

   The second report contains fewer files.

   Why it matters

   Omitted material can masquerade as successful remediation.

   Proposed response

   Reconcile the manifests and show an incomplete comparison until missing files are accounted for.

2. Signal

   Different specialists use incompatible severity names.

   Why it matters

   Aggregation can misrank serious barriers or drop rows.

   Proposed response

   Preserve raw labels, apply an explicit reviewer-approved mapping and reject unrecognized values.

3. Signal

   An observation has no inspectable location.

   Why it matters

   The owner cannot reproduce or responsibly correct it.

   Proposed response

   Move it to an unverified queue rather than treating the issue as established.

Effort and cost — unknown

Setup work

1. Select representative copies, agree check coverage and establish a comparable baseline.

Ongoing review

1. Inspect findings in the native applications and conduct separate reader/assistive-technology checks where needed.

Unknowns

1. Unknown: scanner cost, runtime, false-positive rate and specialist review effort.

Limits

1. Automated checks cover only part of accessibility. Source scores and optional compliance exports require independent scrutiny and are outside this proposed outcome.