A repair of the model, tested beyond the old errors
Our previous experimental candidate falsely flagged one human file and missed 57 AI files. It was never deployed. All 58 errors are now corrected in the 204-file regression: 108 AI detections, 89 human results and seven human inconclusive results. Those works were reused for development; this is a regression result, not an independent accuracy test.
A linear model of average spectral energy was insufficient. We compared nonlinear spectral classification, temporal features and frozen neural embeddings on grouped development folds. The selected RBF classifier combines all three. We also rejected direct short-clip activation of the upstream CNN and two tree-based probes. The alternatives and numerical export checks remain in the evidence.
Weights were committed at faa165e, and the complete serving verification gate at e8a4af3, before scoring the new works. The threshold and decision rule were not changed after reading their results. Training features use the same frontend as serving; the integrated detector reproduced all 840 frozen validation verdicts exactly.
New works, including an unseen transcode chain
We selected 96 new FMA artist identities and 72 new AIME parent tracks before scoring: eight tracks per generator configuration. Human authors and parent works do not overlap the new head's development set. Each source has its original file, a clean eight-second excerpt, and three matched transcode variants. That is 168 works and 840 files, not 840 independent works.
Each row below contains 72 AI and 96 human files. Counts are correct / inconclusive / wrong, including all failures.
| Input | AI, v2.2 | AI, v2.3 | Human, both |
|---|---|---|---|
| Original file | 5 / 8 / 59 | 23 / 6 / 43 | 80 / 16 / 0 |
| Clean, eight seconds | 1 / 3 / 68 | 20 / 3 / 49 | 84 / 12 / 0 |
| Transcode A | 1 / 6 / 65 | 13 / 5 / 54 | 73 / 23 / 0 |
| Transcode B | 0 / 5 / 67 | 10 / 3 / 59 | 76 / 20 / 0 |
| Transcode C (unseen chain) | 0 / 6 / 66 | 15 / 4 / 53 | 77 / 19 / 0 |
Across all 360 generated files, correct AI results rise from 7 to 81, inconclusive results fall from 28 to 21, and false human results fall from 325 to 258. All 480 human files retain their previous outcomes: 390 human signals, 90 inconclusive, zero observed false AI signals. No individual file changes from a determinate result to inconclusive.
Chain A uses MP3 → Opus → MP3 → Vorbis → MP3; B uses AAC → MP3 → Opus → MP3. Chain C, absent from training, uses MP3 96k → AAC 48k → Opus 24k → MP3 128k. Exact parameters and file hashes are in the JSON. Matching duration and processing across both labels prevents codec or length alone from serving as the label.
A separate set of 12 previously retained recordings from four OpenGameArt authors was never fitted to this head. All outcomes remain unchanged: three human and nine inconclusive, no false AI. These are additional controls outside FMA, not newly acquired or wholly blind data.
Original-file results by generator configuration
Each row has eight newly selected originals. Versions may share architectures; these are historical configurations, not a benchmark of the newest products.
| Generator | AI | Inconclusive | False human |
|---|---|---|---|
| AudioLDM 2 Large | 1 | 0 | 7 |
| MusicGen Large | 3 | 1 | 4 |
| MusicGen Medium | 2 | 0 | 6 |
| MusicGen Small | 2 | 1 | 5 |
| Mustango | 3 | 1 | 4 |
| Riffusion | 3 | 0 | 5 |
| Stable Audio v1 | 3 | 0 | 5 |
| Stable Audio v2 | 1 | 2 | 5 |
| Udio | 5 | 1 | 2 |
How v2.3 changes the product
The existing v2.2 checks run first. If they already return AI, that result is retained. Otherwise, a frozen additional classifier analyzes the opening eight seconds. A margin at or above 0.872304 adds an AI signal; a lower score, short input or unavailable model preserves the previous result. This release adds no abstention rule. The displayed score is an uncalibrated screening score, not a probability of authorship.
This update applies to single-song screening. Batch screening still uses the original acoustic model. The new head does not examine every section of a track or locate AI-generated splices. Agreement and finite zero-error counts cannot prove an individual recording's origin.
Sources and what this test cannot establish
Human references come from the historical FMA archive, restricted to entries listed as CC BY or CC0. Generated audio comes from AIME at revision b84d4be5, under CC BY 4.0. We do not extract its MTG-Jamendo recordings or description tags. Source IDs, artists, titles, license records and hashes are published; audio is not redistributed.
The frozen neural backbone comes from the MIT-licensed lofcz detector. Its training overlap is unknown, including possible FMA overlap. FMA versus AIME remains a dataset-origin confound, and generation-session or prompt overlap is unknown. Historical human references are not per-artist attestations. The sample is small and transformed files from one work are correlated.
We have not demonstrated broad current-generation coverage, superiority to competitors, or a zero population false-positive rate. The remaining 258 false-human results across all variants are real failures. These new works are now consumed validation data; further tuning must be tested on another untouched set.