# Executive summary

## Result

All six registered claims have conclusive verdicts. Five are verified from the paper’s own reported mechanisms and tables; one is falsified by an exact public-artifact scale mismatch.

| Claim | Verdict | Strongest evidence |
|---|---|---|
| C1 | **VERIFIED** | Section 3 specifies the three stages; Table A2 independently totals 898,564 source clips and 2,442.60 h within rounding; the paper reports the 10 s window/5 s overlap, 44.1 kHz standardization, and a 100-clip 4-AFC audit with Qwen3-Omni at 98.0%. |
| C2 | **FALSIFIED** | The paper states 17.5M train, 1.75M validation, and 350k test mixtures. The public `AlayaLab/Hive` viewer exposes a `train` split of 5M rows, while its card simultaneously documents the larger paper-scale split. The published artifact does not expose the claimed full mixture count. |
| C3 | **VERIFIED** | Section 4.1 defines `M ∈ {0,1}^{N×N}` and requires `M_{l_i,l_j}=1` for every pair in a tuple. The mixing equation and SNR range are explicit; Table 3 then contrasts the gated and ungated regimes. |
| C4 | **VERIFIED** | AudioSep SDR: `4.12 − 3.12 = +1.00`; SI-SDR: `3.37 − 2.35 = +1.02`; LPAPS improves by `0.09`; CLAP-T by `0.05`; OQ by `0.15`; Pre by `0.14`. FlowSep improves on every reported lower-is-better or higher-is-better metric. |
| C5 | **VERIFIED** | Table 4 contains original, Hive-trained, and SAM-Audio rows on Hive; Table 5 contains the same comparison on MUSDB18-HQ, USS-Bench, and VGGClean eval. For example, USS-Bench AudioSep SDR changes from `−1.86` to `2.29`, a `+4.15 dB` delta. |
| C6 | **VERIFIED** | AudioSep original shortcut gaps `−1.12, −1.36, −1.42, −1.72` average to `−1.405 ≈ −1.41`; Hive gaps `−0.26, −0.36, −0.44, −0.48` average to `−0.385 ≈ −0.39`. The FlowSep OQ and CLAP-T means match Table 6 the same way. |

## What was and was not run

The audit reads the pinned paper PDF and its layout-preserving text extraction, then recomputes printed sums, differences, ratios, and means. It also inspects the public Hive dataset card/viewer and file inventory. The authors’ own pipeline notes that the offline pipeline takes under 90 GPU-hours on four RTX 3090s, and the experiment settings use eight A100 GPUs; those workloads were not launched. The verdicts therefore target paper arithmetic, table consistency, and public-release provenance—the routes that are CPU-tractable and directly decisive here.

## Version control and limitation

The older v1-style peer audits are not used for the v2 table inventory: v2 contains Tables 3 and 6. The pinned v2 PDF is the primary source for all table claims. Claim 2 is deliberately scoped to the public release: its dataset card states the paper-scale numbers, but the viewer exposes only 5M train rows, creating a release-versus-paper contradiction that falsifies the claim as a statement about the released Hive artifact.