PDF accessibility audit with an evidence report

Run machine-detectable PDF/UA, WCAG 2.2 AA, and Section 508 checks against a PDF, get a JSON findings report with each finding located on the page, and keep judgment calls behind an explicit human-review gate.

Audit a PDF

The gap: a checker flags presence, not meaning

Accessibility checkers can tell you whether a PDF has a structure tree, a document language, a title, figure alt text, and table headers. They cannot tell you whether the alt text describes the image, whether the reading order matches the visual meaning, or whether a table conveys the right relationships. Those are human judgments. Most tooling blurs the two together, so a green score reads as a compliance claim it was never entitled to make. okraPDF A11y Audit keeps the line explicit: machine-detectable conditions are checked automatically and reported as evidence; the semantic items that require a person are routed to a human-review gate and stay open until someone resolves them. A machine PASS is a pre-assessment, not an attestation.

What the audit checks

okraPDF parses the PDF and evaluates veraPDF / Matterhorn-style conditions across PDF/UA, WCAG 2.2 AA, and Section 508. Each finding carries a criterion, a status, the standards it maps to, and its location on the page so it's citable in the source. Machine-detectable signals are decided automatically; everything that needs a person is marked needs-human-review rather than guessed.

Tags and structure tree present, MarkInfo set, marked content bound to structure
Document language set, document title present and configured to display
Real text layer vs image-only pages; Unicode mapping for custom fonts
Figure alt-text presence (not quality) and decorative-artifact marking
Table tagging and header cells (TH) present
Reading-order signals exposed for review — true reading order stays a human-review item
Meaningful alt text, table meaning, and form instructions: routed to the human-review gate

A findings report you can hand off

The audit returns a JSON evidence report. Each entry is a single criterion with a status of failed, warning, or needs-human-review, the standards it maps to, a suggested fix, and its location on the source page. failed and warning items are machine-decided; needs-human-review items are the gate that has to be resolved before the document is treated as remediated or attested. Drop the JSON straight into a remediation ticket or a reviewer queue.

{
  "document": { "pages": 12, "tagged": true, "lang": "en-US", "title": "Annual Report" },
  "gate": "needs-human-review",
  "summary": { "failed": 2, "warning": 1, "needsHumanReview": 3, "passed": 18 },
  "findings": [
    {
      "criterion": "figure-alt-text-present",
      "status": "failed",
      "standards": ["WCAG 1.1.1", "PDF/UA 7.3"],
      "page": 4,
      "bbox": [72, 118, 511, 402],
      "fix": "Add Alt to the chart figure; mark its background grid as an artifact."
    },
    {
      "criterion": "table-header-cells",
      "status": "warning",
      "standards": ["WCAG 1.3.1", "PDF/UA 7.5"],
      "page": 7,
      "bbox": [54, 240, 540, 388],
      "fix": "Table is tagged but no TH header cells were found. Verify header associations."
    },
    {
      "criterion": "reading-order-matches-meaning",
      "status": "needs-human-review",
      "standards": ["WCAG 1.3.2", "WCAG 2.4.3", "Matterhorn human checks"],
      "page": 1,
      "bbox": null,
      "fix": "Walk the tag tree with a screen reader; machine checks cannot prove order."
    }
  ]
}

The human-review gate

The gate is the point of the product. Conditions a machine can detect are decided automatically. Conditions that need a person — does the alt text describe the image, does speech follow the intended order, does the table convey the right relationships — are surfaced as needs-human-review and block attestation. The status only flips from audited to attested when a reviewer resolves the gate. okraPDF gives you the evidence and the preflight; it does not certify legal compliance and does not pretend a clean machine run is a sign-off.

Who it is for

Higher-ed accessibility offices, public-sector and government teams working toward the ADA Title II deadline, and compliance teams that need a repeatable preflight plus a paper trail for remediation. Use the free public web tool to audit a PDF and read the findings; once you want this in a pipeline, mint an API key to run the same audit programmatically. This is preflight and evidence for a remediation workflow, not a guarantee of compliance.

Higher-ed accessibility / disability services producing accessible course and admin documents
State and local government working toward Title II
Compliance and QA teams that need page-located evidence per criterion for handoff

Frequently asked questions

Does a passing audit mean the PDF is legally compliant?

No. The audit is preflight plus an evidence report. It decides machine-detectable PDF/UA, WCAG 2.2 AA, and Section 508 conditions and routes everything else to a human-review gate. A machine PASS is a pre-assessment; it is not an attestation or a legal compliance guarantee. Attestation only happens after a person resolves the human-review items.

What is the difference between a failed check and a needs-human-review item?

A failed or warning check is something a machine can detect from the document structure — a missing structure tree, no document language, no title, a figure with no alt text, a table with no header cells. A needs-human-review item is a judgment a machine cannot make reliably: whether alt text is meaningful, whether the reading order matches the visual meaning, whether a table conveys the right relationships. Those stay open until a reviewer confirms them.

Which standards does it map findings to?

Findings carry the standards they map to across PDF/UA (Matterhorn-style failure conditions), WCAG 2.2 AA success criteria, and Section 508. veraPDF / Matterhorn-style coverage is the basis for the machine-detectable conditions, and each finding lists its standard references so the report is usable as remediation evidence.

What does each finding point to in the source?

Each finding points at the exact region of the source page it concerns — for example the chart that is missing alt text or the table that is missing header cells. That makes it locatable and citable in the original document rather than a vague page-level note. Some criteria, like document-wide reading order, have no single location and report none.