| Role | Machine | Used for |
|---|---|---|
| GPU sweeps | NHR@FAU Alex — 1× NVIDIA A40 per job (SLURM a40 partition) | deep baselines (mlp/gru/cnn/resnet), 24 h wall per grid sweep (hpc/sweep.sh) |
| Classical baselines | NHR@FAU Alex — A40 MIG slice | majority/threshold/random_forest sweeps auto-routed by hpc/submit.sh |
| Login node | cluster frontend (no compute) | submission, git, log reads, deterministic re-rendering only |
Every cell aggregates seeds 0–4 (5 seeds); results report mean ± 95% CI half-width (Student t, ddof=1). What is and is not bit-reproducible is documented in docs/determinism.md.
Window memmaps are staged to node-local $TMPDIR before training; sidecar manifests (feature hash, shape, contract fingerprint) travel with the data and are re-verified at load time.
env/results-v1.0.0.lock — sha256 9a44a46b4474f9b592870c31df2c7837ab9a5e906c42a70cfbd1f34e75ead201. The conda prefix that actually produced these numbers (torch + CUDA stack included), captured by hpc/capture_results_env.sh. Every run record independently captured its own package versions at run time; agreement between the two is verified by hpc/verify_results_env.py — see the row below.env/results-v1.1.0.lock — sha256 371de4e6bb54c83841b507c22aea446e7e918a368717adde6849cb6807cb4da4. The environment for the pretrain/transfer re-scores under the corrected multi_grid splits; the held_out results keep the lock above.requirements-ci-py3.10.lock / requirements-ci-py3.12.lock — hash-pinned, installed by CI with --require-hashes. Torch-free: they pin the test gate, not the results.requirements.lock — sha256 2685c1c2c60b595a851478312d95b383123c36f4ce2dfe208eabdc8bdddeaa92. The v1.0.0-era CI pin, recorded as lockfile_sha256 in every run record and kept byte-identical for that reason. It describes neither the results environment (no torch) nor current CI.76a8f6eb814c… (src/evemtbench/contract/contract.json)src/evemtbench/splits/v1.0 · v1.1 (multi_grid_pretrain, topology-grouped) — src/evemtbench/splits/v1.1Aggregated at build time from every run_record.json under outputs/ — the audit reads the records, it does not assert them.
| run records | 1876 |
| with no-tuning attestation | 1876 |
| distinct git SHAs | 011ee01cd33f0007571994d4e1172901d33b1448, 055c212302bc0a5a9ae44ae7e1ad7b085e2c7450, 0740bad210eef4786ee220d6d2b74b5767960491, 097d7505e0a46f1ab85e0a209112398c3b97f45d, 09a30596a0316e1ed7c7b8a2ecfe4fa41756e7e2, 0dacb15beb0f7052423033d9e731417a8b6a65d5, 13f8cdac87378b21f935ec3355b5d3452d15454a, 15eec5ba46f8ba35f3fafd803333066ce6ac3038 … |
| distinct lockfile SHAs | 2685c1c2c60b595a… |
| agree with the results-env capture | 1876 of 1876 (every record's own captured versions match) |
| python versions | 3.12.13 |
| torch versions | 2.6.0+cu124 |
| recorded hosts | a0122, a0123, a0124, a0125, a0224, a0226, a0227, a0228, a0229, a0321, a0323, a0325, a0422, a0423, a0424, a0425, a0427, a0521, a0605, a1622, a1722 |