When Human Review Matters in Accessibility Remediation

Human review matters whenever an automated change can affect meaning, navigation or access. A heading tag, a reading-order change and an equation conversion deserve review alongside an image description.

There is no published Aelira benchmark establishing that 90% of fixes can be accepted automatically or that a fixed percentage of documents is enough to review. The appropriate checks depend on the actual material and the evidence available for that workflow.

Start with the saved file

Aelira Core is beta and applies supported partial fixes. Review the actual saved candidate and recorded unresolved findings. Some documents need manual remediation or cannot produce a downloadable remediated file. A successful scan or a higher score does not by itself establish accessibility conformance.

For structure, check heading hierarchy, reading sequence, list semantics and table-header associations. For conversions, compare the exported mathematics and text against the source. Confirm that content has not been omitted, reordered or given an incorrect role.

Bring subject expertise to educational content

The following examples are illustrative, rather than descriptions of actual customer files.

An economics chart description such as “two lines over time” may omit the comparison the lesson depends on. The reviewer must identify the actual variables, units and relevant relationship from the source before approving a longer description.

A biology process diagram needs its meaningful stages and relationships described. A fluent generated paragraph that invents a reaction or reverses an arrow can be more misleading than an unresolved finding.

Context also determines whether an image is decorative. Do not mark an informative figure decorative simply to remove a scanner warning. Conversely, avoid adding unnecessary descriptions to purely decorative content.

Use a risk-based review plan

1. Inventory the formats, complexity and intended student tasks.
2. Define acceptance checks with the accessibility team, including structural and content-preservation checks.
3. Process only supported, authorized inputs; retain originals and record failures.
4. Review uncertainty flags and consequential generated content. Absence of a flag is not proof of correctness.
5. Test the saved output with the applicable validators, manual checks and assistive technologies.
6. Assign unresolved issues to a named owner and provide an appropriate interim access route with disability services where needed.

Sampling can help evaluate a large corpus, but its coverage and limits must be agreed and documented. A sample of three files or 10% of descriptions is not a universal acceptance rule.

Make review sustainable

Build semantic headings, meaningful image descriptions and accessible templates into new material. Keep review records with the file version, tools, findings, decisions and remaining work. Recheck changed material and provide a channel for students to report barriers.

Estimate staff time from a representative trial, including review and corrections. Do not assume that one hour of checking flagged items completes a semester's accessibility work.

Know the legal and product boundaries

Institutional obligations depend on the applicable law, technical standard and content scope. A review workflow supports the work; it does not replace those obligations or guarantee a compliant outcome.

Read the Aelira evidence and readiness reference and review workflow before planning a rollout.