Aelira Core v0.9.6: accessibility remediation that knows when to stop

The dangerous kind of accessibility automation is not the kind that fails loudly. It is the kind that produces a plausible-looking fix without enough evidence to know whether the fix is right.

Aelira Core v0.9.6 is built around that distinction. This release expands what the open-source remediation engine can do with equations, PDF tables, queued document work, and Canvas content. More importantly, it makes the boundary between safe automation and human judgment much harder to cross by accident.

The whole self-hosted product remains free and open source. You can inspect the implementation, run it in your own environment, and review the evidence behind each remediation decision.

Image equations become structured MathML

An equation printed as an image is inaccessible to screen readers unless the document provides a usable text or mathematical alternative. Recognizing the pixels is only the beginning. The system also has to prove that it changed the right image occurrence, that the recognized expression can be converted safely, and that the saved document still contains the intended association.

In v0.9.6, Aelira can address eligible standalone equation images by page, image reference, index, bounding box, and exact occurrence identity. It can recognize an equation as LaTeX, convert that expression to MathML, round-trip it through deterministic verification, and associate the result with the correct Formula content in the PDF.

Recognition does not equal approval. Image-derived equations are marked as AI vision output, capped at 0.55 confidence, and always require human acceptance. That acceptance is checked again after processing and under the final artifact lock before the document can be published.

If purpose-bound vision configuration is missing, Aelira detects the equation and leaves it for manual remediation. It does not borrow an unrelated provider or invent MathML. Unsupported Form XObjects, multiple equations in one screenshot, handwriting, chemistry, vector-only equations, and mixed STEM visuals also remain explicit review cases.

PDF tables are automatic only when their structure is provable

Tables are another place where a visually convincing result can hide a broken semantic structure. A row of text that looks aligned is not enough. Accessible PDF tables need real Table, TR, TH, and TD relationships tied to marked content in the document.

The new table path binds visible cells to their actual PDF content through MCID and MCR references, checks ParentTree integrity and object identity, saves the result, then verifies the semantic structure again from the output file. If any part of that chain fails, the write is rolled back and the table stays in review.

Verified ordinary tables can be remediated automatically. Ragged, merged, ambiguous, unbound, or excessive structures fail closed before mutation. The inclusive limits are 64 columns, 10,000 cells, and 200 tables, which prevents a malformed structure from turning into an unbounded memory or document-growth problem.

Document remediation now has a durable execution boundary

Long-running document work should survive a browser reload, a worker restart, or a slow file without leaving callers guessing. v0.9.6 moves remediation through a durable queue with active-job deduplication, bounded waiting for legacy callers, tenant-fenced status and latest-job endpoints, and artifact-gated downloads.

Each remediation runs in a killable subprocess with a configurable hard deadline and bounded termination. A hard timeout is terminal and is not retried automatically. This matters because retrying a timed-out document mutation can duplicate work or publish the wrong output unless the entire claim and artifact lifecycle is rebuilt carefully.

Artifacts are identified by stable descriptors and checksums rather than reopened from a mutable file path. They remain unavailable until validation succeeds, and external errors are reduced to stable codes instead of exposing provider responses or filesystem details.

The same release also preserves OCR-generated searchable text in delivered PDFs, tightens passive HTML reconstruction from PDF content, and fixes tenant scoping and CORS behavior around integrations.

Canvas remediation becomes restart-safe

Canvas stored-content remediation now persists durable jobs and immutable, bounded source snapshots. Candidate fingerprints capture the complete target, approvals are revalidated against the current source, writeback intents are fenced, and unfinished work can be discovered after a restart.

That design protects against a subtle but important failure: approving one version of a course item and later writing the result into a changed version. If the source is no longer the source that was reviewed, the approval is no longer enough.

A release you can inspect

The v0.9.6 release is immutable and tied to the exact source commit. It includes signed release evidence, seven SBOM files, the @aelira/cli package, and API and dashboard container images for both linux/amd64 and linux/arm64.

There is still human work in accessibility. Mathematical meaning, unusual document structures, reading order, and author intent cannot be reduced to a green check without evidence. Aelira's job is to automate the cases it can prove, preserve a clear trail, and stop when it cannot.

Read the Aelira Core v0.9.6 release notes, review the source on GitHub, or follow the self-hosting documentation to run the open-source edition in your own environment.