What the candidate may and may not change
This round tested whether a classifier fitted on consumed data could do two things the positive-only v2.3 stack cannot: add an AI signal at a frozen threshold of 0.973195, and correct the one known legacy neural false positive to human when a learned margin and both whole-track spectral heads corroborate human. Filenames, sources, declared genres and bitrates are never inputs. Abstention cannot increase: every rule either keeps v2.3's decision or replaces it with a definite one.
On this fresh set the corrective rule found nothing to correct — no row reached it through the legacy neural path — so every gain below comes from the additive threshold alone. The known Retro Space BGM false positive therefore also remains in production. Its repair still needs its own untouched validation.
Counts are correct / inconclusive / wrong per input form, 108 AI and 84 human files per row. Inconclusive never counts as a detection.
| Input | AI, v2.3 | AI, candidate | Human, v2.3 | Human, candidate |
|---|---|---|---|---|
| Original file | 36 / 7 / 65 | 50 / 7 / 51 | 55 / 29 / 0 | 55 / 29 / 0 |
| Clean, eight seconds | 34 / 6 / 68 | 47 / 6 / 55 | 71 / 13 / 0 | 71 / 13 / 0 |
| Transcode A | 22 / 13 / 73 | 28 / 12 / 68 | 61 / 22 / 1 | 61 / 22 / 1 |
| Transcode B | 19 / 8 / 81 | 27 / 8 / 73 | 65 / 18 / 1 | 65 / 18 / 1 |
| Transcode C | 28 / 6 / 74 | 43 / 6 / 59 | 62 / 21 / 1 | 62 / 21 / 1 |
| Transcode D | 37 / 6 / 65 | 45 / 5 / 58 | 67 / 16 / 1 | 67 / 16 / 1 |
On original files, AI detections rise from 36 to 50 of 108, and wrong human verdicts fall from 65 to 51. On the five transcoded forms, from 140 to 190 of 540. Human outcomes are byte-for-byte identical to v2.3 in both partitions. A separate unreleased round-two candidate was scored only as a non-gating benchmark and is included in the download.
The one error the gate counts
Lasso Lady (seamless loop) is a 2016 FamiTracker chiptune by congusbongus, published on OpenGameArt years before modern music generators existed, under CC0. Its original file stays inconclusive for both systems — an error neither makes into a verdict. Its four transcoded forms — MP3, AAC and Opus chains at low bitrates — all receive AI verdicts from production v2.3 today. The candidate inherits those same four verdicts: on two of them its own score stays below the threshold and the baseline result carries over.
This is the same structural lesson as the last round: a detector that can only add positive evidence cannot repair an existing false positive, and heavy transcoding of dense chiptune loops remains a real weakness of the whole v2.3 stack. Our gate is absolute — zero observed human errors on new audio — because a shipped detector that misjudges known-human music is worse than a smaller recall. The candidate stays out of production, and v2.3 keeps making these exact four errors until a repair candidate proves otherwise on untouched data.
Every other gate held: no previously correct file lost its result (0 regressions), no file gained an inconclusive result (0), and AI detection strictly improved on originals, on processed forms, and on the total separately.
Original-file results by generator
Each row is 12 newly drawn originals from one historical configuration, acquired for this test and excluded from every fitting run. Named versions are not the newest product versions.
| Generator | v2.3 | Candidate |
|---|---|---|
| AudioLDM 2 Large | 9 / 2 / 61 | 15 / 2 / 55 |
| MusicGen Large | 33 / 5 / 34 | 46 / 4 / 22 |
| MusicGen Medium | 16 / 15 / 41 | 20 / 15 / 37 |
| MusicGen Small | 36 / 13 / 23 | 42 / 12 / 18 |
| Mustango | 24 / 0 / 48 | 38 / 0 / 34 |
| Riffusion | 10 / 0 / 62 | 28 / 0 / 44 |
| Stable Audio v1 | 2 / 0 / 70 | 4 / 0 / 68 |
| Stable Audio v2 | 18 / 6 / 48 | 19 / 6 / 47 |
| Udio | 28 / 5 / 39 | 28 / 5 / 39 |
Where the human references came from
The 84 human works are new author identities: 68 general FMA tracks, 4 FMA electronic tracks, and 12 OpenGameArt electronic authors with pre-2019 publication records and CC0 licenses — the genre that resists this detector most. Two works failed acquisition quality gates before scoring (a constant opening after transcode; an original under 30 seconds) and were replaced by that pool's sealed next reserve; both failures ship in the download.
| Human pool | Works | Human, candidate (all six forms) |
|---|---|---|
| FMA-general | 68 | 315 / 93 / 0 |
| FMA-electronic | 4 | 10 / 14 / 0 |
| OpenGameArt-electronic | 12 | 56 / 12 / 4 |
Sources, limits and the next boundary
Generated audio comes from AIME under CC BY 4.0. Human references carry per-track license records from FMA (Attribution-ShareAlike 3.0 International, Attribution-ShareAlike 3.0 United States, or Public Domain Mark 1.0) and per-work CC0 records from OpenGameArt. Source pages, file hashes, license evidence and provenance are in the download; audio is not redistributed. Acquisition downloaded 1.7 GB under a sealed 3 GiB budget with score-blind identities: no model ran until an independent reviewer had reverified every decode gate.
Our neural backbone comes from the MIT-licensed lofcz detector; its upstream training overlap remains unknown. The candidate reads the opening eight seconds. It cannot locate splices or establish whole-track authorship. Passing or failing this gate does not establish current-generation coverage, superiority to competitors, or zero population false positives.
These 192 works are now consumed validation data. The next candidate — including any repair aimed at the transcoded-chiptune weakness this round exposed — must face another untouched set under a freshly sealed protocol.