A public record of changes to the deployed detection engine and to the site around it. 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. Engine entries change what a result says. Site and interface entries change how it is presented and never touch the scoring path. Every entry states its release id.
Engine changes
5 September 2026 — engine V3, the current detector
Release rel-v3-stage1-2026-09-05-308f3a03fb05 — the active detector. V3 keeps the V2 fusion core unchanged (no classifier was newly trained) and replaces the decision layer around it: frozen decision thresholds that depend on the length of the text, a frozen monotonic 0-100 display map, a family-bootstrap score stability range, and a new result contract. A result is now a 0-100 AI-writing signal score with its stability range, one of three bands, the length band and the measured false-flag rate for that length. Its sealed confirmation was scored once and missed two of six bars — human texts of 100 to 149 words at about 2.4 wrong AI signals in every 100 against a bar of 2 in 100, and issue-tracker human text at about 4.8 in every 100 against a bar of 3 in 100. The publisher released with both published rather than re-tuning against an opened test set. Full figures: evidence; thresholds and display map: methodology.
4 September 2026 — V2 conservative flagger (superseded)
Release rel-2026-09-04-c3caa5009358 — superseded by V3 on 5 September 2026. V2 introduced the fusion core that is still in production, and returned one of three bands for English text of at least 100 words with no displayed number, using a single pair of thresholds on the internal scale. It stated with every result that it did not prove AI or human authorship. Entries below this one describe the previous score-based engine and are kept as history.
26 August 2026 — clamp recovery, and launch hardening (previous engine)
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 a separate non-English path (since retired), 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.
Site and interface changes
None of the entries in this section changes a score, a threshold or a band. They are recorded here so a result seen on one date can be matched to the site that presented it.
6 September 2026 — content and information-architecture pass
Release rel-site-content-2026-09-06-b2edaedd6c18 — theme 9.3.0. The content-type hub was rebuilt from a 34,000-word merge of nine former pages into a twelve-card hub; the methodology, training-and-validation and three content-type guides were rewritten for V3; the frozen thresholds are now published per length band in both internal and displayed units; stale V2 descriptions were removed from every page that is not dated history; the one page under /guides/ moved into the blog; and in-body “Last reviewed” lines were removed from articles, where the theme’s review-date field is now the only mechanism. Engine untouched.
6 September 2026 — interface pass 2
Release rel-site-ux2-2026-09-06-a266b771b1e0 — theme 9.2.0, plugin 1.5.0, UI contract r66.3. Fixed the result auto-scroll, the empty card in compare mode, the band-strip labels (now positioned from the result’s own band edges) and a sticky summary that could show a stale result; rebuilt the phone layout of the tool and the model guides; removed the fixed example section and the in-tool example chips; gave every post and guide a cover image in the current identity. Engine untouched.
6 September 2026 — interface pass 1
Release rel-site-ux-2026-09-05-97ec62650924 — theme 9.1.0, plugin 1.4.0. The product was renamed AI Detector Checker throughout, the logo became a single outlined image, the result card gained a per-band “What this means for you” box, a print report and designed error states, and the evidence chart was generated directly from the sealed confirmation file. Engine untouched.
5 September 2026 — site release, Stage 2
Theme 9.0.0 — the information architecture that accompanied V3: the evidence page replaced the separate benchmark, model-card and limitations pages, the help and model hubs were introduced, and 23 retired URLs were given permanent redirects. Stage 1 (engine, result card and content truth) and Stage 2 (identity and architecture) were released separately, with separate rollbacks, so that a site rollback could never touch the engine.