Six candidates, one test on 192 new works

More training was not automatically better

Our v2.3 repair corrected the earlier 1 human error and 57 AI misses in a consumed regression. Its separate new-work test detected only 23 of 72 AI originals. Those new-test files became development material for this round. Repeating them would no longer be an independent test.

We compared three nonlinear spectral/temporal/neural feature classifiers and two ways to fine-tune the existing small neural network. After those five results, we explicitly added one final experiment that emphasizes high-scoring human mistakes and consistency between two codec views of each recording. We added it before downloading this new validation audio; it was an adaptive follow-up, not part of the original five-way plan.

The selection rule was the largest number of AI detections on held-out development groups, at a threshold above every observed human score. All forms of a recording stayed in one fold; human aliases were united by normalized names and artist IDs. An initial alias-split error invalidated the first runs, which were archived and repeated with corrected groups. The old production model had already fitted much of this corpus, so incremental development replay is not independent evidence.

The selected model was expanded-c10, with fixed margin threshold 0.955892. The model was frozen at d999633; this validation protocol was initially committed at b382f5c before acquisition. After two infrastructure failures, the acquisition script and protocol reference were resealed at d5171cd before the final restart. Model, threshold and selected identities stayed unchanged. Its Node export reproduced all 2,335 development decisions, with maximum score drift below 0.000001. Zero human positives in that development table follow from threshold construction; they do not establish safety on new recordings.

Every original and transformation remains in the result

The completed test contains 108 new AIME recordings, 72 new FMA authors, and 12 new OpenGameArt authors. Each work has its original file, a clean eight-second excerpt, and four transcode chains. These are 192 works and 1,152 correlated files. This set was excluded from all six training and model-selection runs.

An earlier 1,512-file validation was voided before model scoring. The production decoder found 12 additional short files outside its declared exceptions. We kept that failure record and selected a disjoint replacement set under a new protocol. Every one of the 1152 new files passed the same 128,000-sample production-decoder gate; all received a model score, including 504 human files. None was silently retained as a baseline-only result.

Counts below are correct / inconclusive / wrong. Each row contains 108 AI and 84 human files. Inconclusive results never count as correct detections.

The new chain D is AAC 48k → Opus 24k → MP3 64k → AAC 96k → MP3 128k. It was absent from training. The model adds an AI signal only at or above its frozen threshold; otherwise it retains v2.3. Across this test, 0 files gained an abstention and 0 previously correct files lost their correct result.

Ordered reserves were sealed for each generator family and human source. Only acquisition or decoded-audio eligibility failures could trigger a replacement, before scoring. The completed run used all 192 primaries. Two earlier runs stopped on directory and output-format bugs; their progress was voided and the same selection restarted after documented repairs. The download preserves those records. No work was removed because of a detector result.

Original-file results by generator

Each row has 12 new originals from one historical configuration. Named versions can share architectures and are not the newest product versions.

Human controls from a second source

FMA and AIME differ in more than authorship. To challenge that source dependency, we added 12 OpenGameArt authors with pre-2019 publication records and CC0 licenses. They were excluded from development by author identity and normalized name, alongside the four older OpenGameArt control authors.

The 12 previously retained recordings from four other OpenGameArt authors remain separate, unfitted controls rather than new blind samples. Their result is 3 / 9 / 0. The five original heavy-transcode targets still receive AI results in the development replay.

The false positive is the original Retro Space BGM by Monster Logix Studio, from the human-reference OpenGameArt collection. v2.3's three-window neural recovery produces the AI result, with an uncalibrated score of 0.999257. The new classifier scores it −0.297701, below its positive threshold, but its addition-only policy preserves the earlier result. This is a detector error, not evidence that the artist used AI. Adding more positive signals cannot repair an existing false positive; the legacy recovery decision needs its own repair and another untouched validation.

Sources, limits and the next boundary

Generated audio comes from AIME under CC BY 4.0. Human references use per-track CC BY or CC0 records from FMA and per-work CC0 records from OpenGameArt. Source pages, file hashes and provenance are in the download; audio is not redistributed.

Our neural backbone comes from the MIT-licensed lofcz detector. Its upstream training overlap remains unknown. Publication metadata is not an independent attestation of creator identity or original stems. Different aliases may escape normalization. AIME prompt hashes were deduplicated within each generator family, but generation-session IDs are unavailable and cross-family prompt reuse remains possible.

The added classifier reads the opening eight seconds. It cannot locate splices or establish whole-track authorship. This experiment does not establish current-generation coverage, superiority to competitors, or zero population false positives. Its results are now consumed validation data: any further tuning must face another untouched set.