A public record of changes to the deployed detection engine. Entries appear here only when a change actually reached production. Proposals that were evaluated and rejected are not listed as improvements, because they never shipped.
26 August 2026 — clamp recovery, and launch hardening
Release rel-2026-08-26-00ad353f8f70
- Added a recovery step for a specific false-negative case: where the internal false-positive correction had suppressed a document that the underlying classifier was maximally certain about, the original score is now restored. The step can only raise a score, never lower one.
- Validated against reserved human-safety data from sources never used in development: no new false positives, and no existing correct detection lost.
- The measured benefit is real but narrow. It is concentrated in a particular class of open-weight text generators and is close to zero for several widely used hosted models. It is not a broad accuracy improvement, and is documented here as such.
- Added aggregate serving counters for operational monitoring. No submitted text is recorded.
- Hardening: disabled a development flag that could return internal error detail, and closed a schema endpoint that had been left reachable.
14 August 2026 — short-document correction
Release rel-2026-08-14-a31fa69cd325
- The false-positive correction was found to misfire on short documents that the classifier was already confident about, collapsing them onto a fixed low score. It is now skipped in that specific case.
- Verified to leave all other documents, and all non-English output, bit-for-bit unchanged.
29 July 2026 — reliability
- Analyses were moved off the request event loop and given an explicit admission-control limit, fixing a long-standing timeout problem under concurrent load. Requests that arrive when capacity is exhausted now receive an immediate, honest “busy” response rather than hanging.
27 July 2026 — release integrity and traceability
- Every release now records a cryptographic hash of each artefact the scoring path loads, and re-verifies them at start-up. The service refuses to report itself ready if anything differs from what was attested.
- Every response carries a release identifier and a request identifier, so any individual result can be traced back to the exact engine that produced it.
Earlier work
Before this record began, the engine moved from a rule-based scoring system to the calibrated ensemble described in the model card. That programme also introduced the dedicated multilingual path, defences against invisible-character and look-alike-character manipulation, order-invariant scoring so that reordering paragraphs no longer changes the result, and the register controls that reduced false positives on formal human writing.
For how changes are evaluated before they reach this page, see training and validation.