Detector changelog

Short answer: the engine currently shipping is build ensemble-1. This page lists every build, what changed in it, and which published parameters moved. Detection results are only comparable when they come from the same build, so every measurement on this site is stamped with one.

Last updated August 2026 · engine build ensemble-1

Why a detector needs a changelog

A probability from a detector is only meaningful alongside the version that produced it. If the combination rule, the thresholds or the feature set change, an old score and a new score are not the same measurement. Tools that publish a single accuracy figure with no version history are asking to be trusted about something they cannot show. Every number on this site carries the build it came from, and this page is where those builds are defined.

service-v1 · 2026-08

Current build. Classification moved to a specialist third-party AI music detection service.

Retired the in-browser ensemble engine in favour of a specialist third-party classification service.
Audio is now uploaded over HTTPS to our server and forwarded for analysis; audio is not written to storage by us.
Maintains the probability and confidence reporting format, with results keyed by SHA-256 hash for 14-day caching.
Increased file limit to 25 MB and supported duration to 15 minutes.

ensemble-1 · 2026-08

Retired build. Four detectors combined by weighted vote, with confidence reported separately from probability.

Replaced the single acoustic detector with a four-model ensemble: spectral-artefact (weight 2.2), embedding (1.6), acoustic (1.0) and extended-features (0.9).
Vote weights, thresholds and shrinkage moved into public/config/ensemble.json so the combination rule is inspectable.
Added a disagreement penalty: when models disagree by more than 0.20, confidence is reduced instead of the probability being moved.
Reported probability capped to 15-85% and shrunk toward 50% by a factor of 0.94.
Added the measurement-driven explanation panel that lists which signals moved the estimate and in which direction.
2026-08-24: first measured results published for this build — engine behaviour study 1 (168 detections, 12 reference tracks, 7 encode conditions) at /benchmark. Repeat determinism 0.0 pp, mean encode shift 2.4 pp, median analysis time 406 ms. No accuracy figure is claimed.

single-1 · 2026-07

Historical build (retired). One acoustic/spectral detector running entirely in the browser.

Client-side analysis via the Web Audio API: spectral rolloff and ceiling, high-frequency energy, stereo correlation, dynamic range, transient statistics and loudness behaviour.
Probability and confidence reported as separate outputs, with an explicit inconclusive band.
Downloadable PDF report generated in the browser from the same values shown on screen.
No account, no upload, no server-side storage of audio at any point.

What this changelog will never say

Changelog entries describe behaviour changes, never accuracy claims. An accuracy claim requires a measured benchmark run, and every result cell on /benchmark is still null.
Because audio is never written to storage by us, no user audio is collected, so no build in this list was tuned on files submitted through the site.
When the engine build changes, any published measurement is re-stamped with the build it was produced on rather than carried forward silently.

Where to go next

Benchmark protocol — the metric definitions and corpus design each build will be evaluated against.
Limitations — the failure modes that no build in this list removes.
Methodology — what the engine measures and how the signals are combined.

Cite this page

Quotation with attribution is welcome. Please keep the wording of factual claims intact and link back to the source page.

AI Music Detector. “Detector changelog.” Updated August 2026. https://aimusicdetector.co/changelog