Corrections

What a correction is

If a vendor changes its policy, or you believe an assessment is out of date, a correction triggers an autonomous re-run: the system re-fetches the live document, captures a new immutable snapshot, re-extracts clauses, re-applies the verification gate and two-model consensus, re-scores, and republishes — entirely without human editing. The data is only ever written by the pipeline.

How to trigger one

Send a request to the public correction endpoint:

POST /api/corrections/{platform-slug}

The platform's documents are enqueued for re-analysis and the worker (scripts/run-correction.mjs, also drained by the change-detection cron) executes the full re-run. Because every finding must re-pass the exact-substring verification gate against the new snapshot, a correction can only ever replace one machine-verified result with another.

Why no human edit path exists

A human-editable record could not carry the same guarantee. Every published claim traces to a verbatim quote anchored in an archived, hash-addressed snapshot; allowing manual edits would break that chain. The correction-as-re-run model keeps the entire dataset machine-verifiable end to end.