================================================================================ RUN.log — THE CORPUS-COHERENCE BENCH (PREREG-COHERENCE.md, D-20260903-02) ================================================================================ --- SEAL VERIFICATION (§0.4), performed before anything was read --- method: restore the §0.4 line to its placeholder text `sha256(this file at registration) = ⧖ TO BE FILLED AT SEAL` in a temp copy, then sha256 that copy. computed = 4bbe9548945d1bd849b355a080b49c481c067e21fd105e98d8d6d30c9f2fe5f2 §0.4 seal = 4bbe9548945d1bd849b355a080b49c481c067e21fd105e98d8d6d30c9f2fe5f2 VERDICT = MATCH — the registration this run executes is the sealed one. --- amendments filed by this lane (appended after §13/A-R4, never edited in place) --- A-C1 the S3 crop budget (§4.4 registers the arm; §10.2 budgeted only training-set crops) A-C2 §2.3 path resolution (music/datasets/) and Round 4's landing site A-C3 §2.4's disjointness claim is MEASURED FALSE (21 of 24 shared) ================ the evaluation box:logs/step0.log ================ [2026-09-03T03:19:28Z] STEP 0 — instrument check (no corpus touched) [2026-09-03T03:19:28Z] STEP 0 freeze-A1.txt sha256 = 258a10ee9966ed60eb7839ce9ae9695c431277e4e324564c9fe2bf4e2550ebb1 [2026-09-03T03:19:28Z] STEP 0 freeze-A1.txt lines = 105 MATCHES A1 [2026-09-03T03:19:31Z] STEP 0 music_audioset_epoch_15_esc_90.14.pt bytes = 2352471003 sha256 = fae3e9c087f2909c28a09dc31c8dfcdacbc42ba44c70e972b58c1bd1caf6dedd [2026-09-03T03:19:31Z] STEP 0 checkpoint MATCHES A1 pin [2026-09-03T03:19:31Z] STEP 0 weights.sha256 on disk: a4931a7a01c3e6733352e9d85371835f03bf9135f8b31e1583c23538811d4a32 audiobox-aesthetics-checkpoint.pt fae3e9c087f2909c28a09dc31c8dfcdacbc42ba44c70e972b58c1bd1caf6dedd music_audioset_epoch_15_esc_90.14.pt [2026-09-03T03:19:37Z] STEP 0 python 3.12.14 torch 2.8.0+cpu numpy 1.26.4 laion_clap n/a [2026-09-03T03:19:37Z] STEP 0 torch.cuda.is_available() = False (CPU-only by construction, §3.1) [2026-09-03T03:19:38Z] STEP 0 verification clip: Ballade no. 1 - Op. 23.m4a dur=613.589s start=294.294573s 25 s @ -16 LUFS / -1 dBTP / 48 kHz stereo sha256=2fe2595d7d31067cb0597a1b5acb00502083a05a2cd036163e9ed298669ce6c2 music/venvs/eval/.venv/lib/python3.12/site-packages/torch/functional.py:554: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /pytorch/aten/src/ATen/native/TensorShape.cpp:4322.) return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined] You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. Load the specified checkpoint music/venvs/eval/weights/music_audioset_epoch_15_esc_90.14.pt from users. Load Checkpoint... [2026-09-03T03:19:44Z] STEP 0 CLAP_Module(enable_fusion=False, amodel='HTSAT-base') + explicit ckpt loaded in 6.7s [2026-09-03T03:19:45Z] STEP 0 RECEIPT: CLAP-laion-music embedding: shape (1, 512), L2 norm 1.0000 [2026-09-03T03:19:45Z] STEP 0 A1 REQUIRES: CLAP-laion-music embedding: shape (1, 512), L2 norm 1.0000 [2026-09-03T03:19:45Z] STEP 0 PASS — A1's receipt reproduced exactly; the corpora may now be read. ================ the evaluation box:logs/plan-training-box.log ================ ================ the evaluation box:logs/plan-evaluation-box.log ================ [2026-09-03T03:21:13Z] PLAN chopin: dataset.json sha256 = baf091ea1d251a84c6b96652874d706d78d0da54edc63d14357f76633b112d7d [2026-09-03T03:21:13Z] PLAN chopin: train n=19 (expect 19), corpus n=24 (expect 24) [2026-09-03T03:21:16Z] PLAN bach: dataset.json sha256 = 5460a13eaf487e0bd2689cec495e0883735b598adddc7e3b86cb265ebd37cbcf [2026-09-03T03:21:16Z] PLAN bach: train n=63 (expect 63), corpus n=79 (expect 79) [2026-09-03T03:21:23Z] PLAN sousa: dataset.json sha256 = 5de3991c18e356be0cb82a9b08c0f8699b24306b3acea0a5285664245d077054 [2026-09-03T03:21:23Z] PLAN sousa: train n=84 (expect 84), corpus n=106 (expect 106) [2026-09-03T03:21:27Z] PLAN 1125 crops planned on the evaluation box (music/out/coherence-bench/crops-plan-evaluation-box.csv) sources shorter than 10s: 0 [14.2s] ================ the evaluation box:logs/crop-training-box.log ================ ================ the evaluation box:logs/crop-evaluation-box.log ================ crop_windows.sh: finished plan=music/out/coherence-bench/crops-plan-evaluation-box.tsv rows=1125 par=6 [2026-09-03T03:23:43Z] CROP crop_windows.sh rc=0 in 54.4s [2026-09-03T03:23:44Z] CROP 1125 crops written and hashed, 0 bad, total 55.5s ================ the evaluation box:logs/verify.log ================ [2026-09-03T03:27:54Z] VERIFY 2433/2433 crops re-verified by sha256 after transfer ================ the evaluation box:logs/verify2.log ================ [2026-09-03T03:33:31Z] VERIFY 2577/2577 crops re-verified by sha256 after transfer ================ the evaluation box:logs/embed.log ================ [2026-09-03T03:27:59Z] EMBED 2433 crops to embed music/venvs/eval/.venv/lib/python3.12/site-packages/torch/functional.py:554: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /pytorch/aten/src/ATen/native/TensorShape.cpp:4322.) return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined] You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. Load the specified checkpoint music/venvs/eval/weights/music_audioset_epoch_15_esc_90.14.pt from users. Load Checkpoint... [2026-09-03T03:28:04Z] EMBED instrument loaded in 5.0s [2026-09-03T03:28:06Z] EMBED 8/2433 5.77 crops/s eta 7.0 min [2026-09-03T03:28:25Z] EMBED 168/2433 8.02 crops/s eta 4.7 min [2026-09-03T03:28:45Z] EMBED 328/2433 8.13 crops/s eta 4.3 min [2026-09-03T03:29:04Z] EMBED 488/2433 8.14 crops/s eta 4.0 min [2026-09-03T03:29:23Z] EMBED 648/2433 8.21 crops/s eta 3.6 min [2026-09-03T03:29:42Z] EMBED 808/2433 8.26 crops/s eta 3.3 min [2026-09-03T03:30:01Z] EMBED 968/2433 8.29 crops/s eta 2.9 min [2026-09-03T03:30:20Z] EMBED 1128/2433 8.31 crops/s eta 2.6 min [2026-09-03T03:30:39Z] EMBED 1288/2433 8.33 crops/s eta 2.3 min [2026-09-03T03:30:58Z] EMBED 1448/2433 8.34 crops/s eta 2.0 min [2026-09-03T03:31:17Z] EMBED 1608/2433 8.34 crops/s eta 1.6 min [2026-09-03T03:31:36Z] EMBED 1768/2433 8.35 crops/s eta 1.3 min [2026-09-03T03:31:55Z] EMBED 1928/2433 8.36 crops/s eta 1.0 min [2026-09-03T03:32:14Z] EMBED 2088/2433 8.37 crops/s eta 0.7 min [2026-09-03T03:32:33Z] EMBED 2248/2433 8.37 crops/s eta 0.4 min [2026-09-03T03:32:52Z] EMBED 2408/2433 8.38 crops/s eta 0.0 min [2026-09-03T03:32:55Z] EMBED 2433 forwards in 290.5s [2026-09-03T03:32:55Z] EMBED crop-vector L2 norms: min=1.0000 max=1.0000 mean=1.0000 [2026-09-03T03:32:55Z] EMBED wrote embeddings.npz 4616398 bytes ================ the evaluation box:logs/embed2.log ================ [2026-09-03T03:33:37Z] EMBED 2577 crops to embed music/venvs/eval/.venv/lib/python3.12/site-packages/torch/functional.py:554: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /pytorch/aten/src/ATen/native/TensorShape.cpp:4322.) return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined] You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. Load the specified checkpoint music/venvs/eval/weights/music_audioset_epoch_15_esc_90.14.pt from users. Load Checkpoint... [2026-09-03T03:33:42Z] EMBED instrument loaded in 4.5s [2026-09-03T03:33:43Z] EMBED 8/2577 9.10 crops/s eta 4.7 min [2026-09-03T03:34:01Z] EMBED 168/2577 8.56 crops/s eta 4.7 min [2026-09-03T03:34:20Z] EMBED 328/2577 8.57 crops/s eta 4.4 min [2026-09-03T03:34:39Z] EMBED 488/2577 8.56 crops/s eta 4.1 min [2026-09-03T03:34:57Z] EMBED 648/2577 8.54 crops/s eta 3.8 min [2026-09-03T03:35:16Z] EMBED 808/2577 8.55 crops/s eta 3.4 min [2026-09-03T03:35:35Z] EMBED 968/2577 8.55 crops/s eta 3.1 min [2026-09-03T03:35:54Z] EMBED 1128/2577 8.53 crops/s eta 2.8 min [2026-09-03T03:36:13Z] EMBED 1288/2577 8.49 crops/s eta 2.5 min [2026-09-03T03:36:33Z] EMBED 1448/2577 8.46 crops/s eta 2.2 min [2026-09-03T03:36:52Z] EMBED 1608/2577 8.43 crops/s eta 1.9 min [2026-09-03T03:37:12Z] EMBED 1768/2577 8.42 crops/s eta 1.6 min [2026-09-03T03:37:31Z] EMBED 1928/2577 8.40 crops/s eta 1.3 min [2026-09-03T03:37:51Z] EMBED 2088/2577 8.38 crops/s eta 1.0 min [2026-09-03T03:38:10Z] EMBED 2248/2577 8.37 crops/s eta 0.7 min [2026-09-03T03:38:30Z] EMBED 2408/2577 8.36 crops/s eta 0.3 min [2026-09-03T03:38:49Z] EMBED 2568/2577 8.35 crops/s eta 0.0 min [2026-09-03T03:38:50Z] EMBED 2577 forwards in 308.6s [2026-09-03T03:38:50Z] EMBED crop-vector L2 norms: min=1.0000 max=1.0000 mean=1.0000 [2026-09-03T03:38:50Z] EMBED wrote embeddings.npz 4889733 bytes ================ the evaluation box:logs/metrics.log ================ [2026-09-03T03:39:05Z] METRICS 2577 crop vectors, 2577 index rows [2026-09-03T03:39:50Z] METRICS permutation test: obs=0.211063 p_one_sided=0.000100 (43.6s) [2026-09-03T03:39:50Z] METRICS wrote music/out/coherence-bench/out/RESULTS.json ================ the training box:logs/plan-training-box.log ================ [2026-09-03T03:19:48Z] PLAN fm-control: dataset.json sha256 = 13b16b5314807d0c4d45d61e37ded51e5cd4e176a761e2e043f1303d58d358ac [2026-09-03T03:19:48Z] PLAN fm-control: manifest.json sha256 = 11a12a746a929b8046d09905bf46ed98abefa3ae160859ee334760769563754c [2026-09-03T03:19:48Z] PLAN fm-control: train n=24 (expect 24), corpus n=30 (expect 30) [2026-09-03T03:19:52Z] PLAN mnml: dataset.json sha256 = d26114c0d3828f1b2aaab947db6e676fc5a617cfbe3def62508c1924b8ae8333 [2026-09-03T03:19:52Z] PLAN mnml: manifest.json sha256 = 73040ef2e45694efef61e308ecb333c10f7aa21580fc086826a42245bd64e41f [2026-09-03T03:19:52Z] PLAN mnml: train n=159 (expect 159), corpus n=208 (expect 208) [2026-09-03T03:20:13Z] PLAN renders set A: 15 distinct sha256 across 3 dirs [2026-09-03T03:20:13Z] PLAN renders set A: all 15 registered sha256 prefixes located [2026-09-03T03:20:14Z] PLAN 1308 crops planned on the training box (music/out/coherence-bench/crops-plan-training-box.csv) sources shorter than 10s: 0 [26.4s] ================ the training box:logs/crop-training-box.log ================ Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped Incorrect BOM value Error reading comment frame, skipped crop_windows.sh: finished plan=music/out/coherence-bench/crops-plan-training-box.tsv rows=1308 par=16 [2026-09-03T03:22:46Z] CROP crop_windows.sh rc=0 in 25.3s [2026-09-03T03:22:48Z] CROP 1308 crops written and hashed, 0 bad, total 27.0s ================ STEP 1 — cross-box crop equivalence (§4.3) ================ nominated track: music/corpus/fm-control/audio/Monokrak205FloatingMind_VoyagerEP__[monokrak205] FloatingMind_Subaquatika.wav source sha256 identical on both boxes: 5231ee65b50c120a4b0894b78f9ece78ec41996674500252d8b7cf615f9c56cb ffprobe duration identical on both boxes: 346.295465 s ffmpeg on both boxes: 8.0.1-3ubuntu2 the training box w25 start=81.573866 sha256=aed726381f6950df1044a98b21902edbe2c0316429a44a553679bf6e141acdb9 bytes=960194 the evaluation box w25 start=81.573866 sha256=aed726381f6950df1044a98b21902edbe2c0316429a44a553679bf6e141acdb9 bytes=960194 the training box w50 start=168.147732 sha256=95d5612ea37380e576afd7df6e658669087c65226ed8da0ed46c25b8797884ae bytes=960194 the evaluation box w50 start=168.147732 sha256=95d5612ea37380e576afd7df6e658669087c65226ed8da0ed46c25b8797884ae bytes=960194 the training box w75 start=254.721599 sha256=ce5b8bf9472a3d3b2fd455b749a7bf33ecc0d081db29753882752d80a4a3f167 bytes=960194 the evaluation box w75 start=254.721599 sha256=ce5b8bf9472a3d3b2fd455b749a7bf33ecc0d081db29753882752d80a4a3f167 bytes=960194 VERDICT: 3/3 byte-identical. Cross-box cropping is PROVEN for this run; the crops travel and the §4.3 fallback (copy source audio, cut everything on the instrument box) is NOT used. ================ STEP 3 — transfer (§10.2/§10.4) ================ transport: rsync -a, two legs via the laptop (the training box and the evaluation box do not hold each other's ssh host keys and the bench changes no box configuration). rsync is the registered transport; the receipt is the post-transfer sha256 re-verification, which is what STEP 3 actually requires. leg 1 the training box -> laptop 1,256,331,078 B in 21 s (58.4 MB/s) leg 2 laptop -> the evaluation box 1,256,331,214 B in 17 s (109.2 MB/s) Set B 144 crops, both legs, ~10 s post-transfer: 2433/2433 then 2577/2577 crops re-verified by sha256 (see VERIFY lines above). ================ the training box:logs/crop-training-box-setB.log (Set B, run in the foreground) ================ [2026-09-03T03:32:57Z] PLAN renders set B: 48 distinct sha256 across 2 dirs [2026-09-03T03:33:02Z] PLAN 144 crops planned on the training box, Set B sources shorter than 10s: 0 [4.4s] [2026-09-03T03:33:06Z] CROP crop_windows.sh rc=0 in 4.4s [2026-09-03T03:33:06Z] CROP 144 crops written and hashed, 0 bad, total 4.5s ================ Round 4 landing receipt (A-R4) ================ driver: ALL 49 RENDERED in 1042s wall (incl. per-process model load) ROUND 4 END 2026-09-03T03:30:03Z manifest: sha256sum -c MANIFEST.sha256 -> 49 lines, 0 FAILED records: 49 clips, 48 distinct output_sha256, 12 prompts x 4 arms + 1 determinism repeat determinism_control: byte_identical = true (PASS) pair_checks: 72 of 72 passed (every arm pair byte-distinct) ================ FINAL metrics run (with Set B) ================ [2026-09-03T03:40:06Z] METRICS 2577 crop vectors, 2577 index rows [2026-09-03T03:40:50Z] METRICS permutation test: obs=0.211063 p_one_sided=0.000100 (42.1s) [2026-09-03T03:40:50Z] METRICS Set B: 49 records over 12 prompts [2026-09-03T03:40:50Z] METRICS wrote out/RESULTS.json