Human review

Why review exists

Detection is layered and measured, but no detector should be presented as perfect. Review is the answer to the question every security team asks: what if the model misses something? Salus gives the people who understand the documents a place to see what was detected, record what was wrong, and add what was missed.

The playground

The console’s detection playground runs the full pipeline on pasted text and shows every span with its class, source layer, confidence and verifier verdict. It is the fastest way to validate detection on your own samples before and after go-live, and to test a custom rule before publishing it.

Detection reviews

Captured detection runs appear in the review screen with the original input (stored encrypted) and every span. A reviewer marks each span correct or a false positive and can mark a span the engine missed. Completion and reviewed precision are shown as cards over the last 24 hours and the last 7 days, and the feedback can be exported as JSON for your own analysis or for the Salus team.

Feedback is evidence. It never retrains a model or edits a rule silently; releases are pinned, and a release is either active or rolled back.

Custom rules

Where coverage is missing for an identifier specific to your organisation, a rule manager adds a regular-expression rule with a category and a confidence, tests it against sample text in the console, and publishes it. Rules are per tenant, or global when published by a platform administrator. They are read by the rule service on every request, so a published rule applies to the next call without a restart. Active rules are disabled before they are edited, and every version is kept.

Review on endpoints

In the endpoint deployment, the device component can show the person what was detected in a file before it leaves, so an incorrect detection can be removed and a missed value marked. In the API path, that step belongs to your application: Salus returns the protected content and the span list, and your workflow presents its own approval step where policy requires a human decision before egress.