WAM19 Dell CUDA run notes

Machine summary
- Host: pop-os
- OS/kernel: Linux pop-os 6.18.7-76061807-generic x86_64
- System memory: 30 GiB RAM, 19 GiB swap
- Profile: dell / RTX A5000 16GB CUDA profile
- CUDA device: NVIDIA RTX A5000 Laptop GPU, 16384 MiB VRAM, driver 580.159.03, CUDA 13.0, compute capability 8.6
- Runtime reported by run10.py: cuda_available true, cuda_device "NVIDIA RTX A5000 Laptop GPU", cuda_mem_gb 15.61, device_request cuda, mps_available false, threads 1

Setup and validation
- Created and used wam19/.venv.
- Installed req.txt into the venv.
- nvidia-smi succeeded on the host.
- PYTHONPATH=. PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/python -m pytest -q passed: 71 tests.
- No math, estimator, selector, or theorem-facing measurement code was edited.

Run command
- Initial requested command failed immediately in the elevated shell because dell.sh invokes python and the venv was not on PATH: ./dell.sh: line 17: python: command not found.
- Rerun command used:
  PATH="$PWD/.venv/bin:$PATH" OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 NUMEXPR_NUM_THREADS=1 /usr/bin/time -p env PYTHONPATH=. ./dell.sh --fresh
- This rerun is the successful fresh Dell profile run recorded here.

Wall clock
- run10.py elapsed_seconds: 24448.631
- /usr/bin/time -p: real 24593.28, user 208022.00, sys 719.33

Artifact check
- out/dell.jsonl exists, 288 rows.
- out/dell_sel.json exists.
- out/dell_an.json exists.

Overall validity
- Rows: 288
- child_ok count: 288
- Error row count: 0
- Benchmarks present: cifar10, fmnist, kmnist, mnist, stl10, svhn
- Selection analysis used child_ok_filter true with n_used 288 and n_child_ok 288.
- This is a valid child-correct selection run.

Metric interpretation
- Headline metric: full_primary.score.
- Score-gap full extension surface: full_score.score.
- Label-anchor diagnostic: full_label.score.
- Legacy full.score was not used as the headline in these notes.

Per-benchmark ranges
- cifar10: train_acc 1.0 to 1.0; test_acc 0.1 to 0.2355; full_primary.score 0.12753164964069877 to 0.55121410760694; full_score.score 0.12753164964069877 to 0.55121410760694; full_label.score 1.0000000000000002 to 1.0000000000000002.
- fmnist: train_acc 1.0 to 1.0; test_acc 0.0935 to 0.6005; full_primary.score 0.13734138963032252 to 0.4839928087535181; full_score.score 0.13734138963032252 to 0.4839928087535181; full_label.score 1.0000000000000002 to 1.0000000000000002.
- kmnist: train_acc 1.0 to 1.0; test_acc 0.097 to 0.436; full_primary.score 0.13007493635747502 to 0.543079641518963; full_score.score 0.13007493635747502 to 0.543079641518963; full_label.score 1.0000000000000002 to 1.0000000000000002.
- mnist: train_acc 1.0 to 1.0; test_acc 0.094 to 0.6665; full_primary.score 0.14504001815410522 to 0.5463261855365469; full_score.score 0.14504001815410522 to 0.5463261855365469; full_label.score 1.0000000000000002 to 1.0000000000000002.
- stl10: train_acc 1.0 to 1.0; test_acc 0.1 to 0.251; full_primary.score 0.127318583610898 to 0.5340542340653817; full_score.score 0.127318583610898 to 0.5340542340653817; full_label.score 1.0000000000000002 to 1.0000000000000002.
- svhn: train_acc 1.0 to 1.0; test_acc 0.0995 to 0.136; full_primary.score 0.12710551732603292 to 0.5546062666872963; full_score.score 0.12710551732603292 to 0.5546062666872963; full_label.score 1.0000000000000002 to 1.0000000000000002.

full_primary.score constant or missing check
- No benchmark had missing full_primary.score.
- No benchmark had constant full_primary.score.
- Unique full_primary.score counts: cifar10 46, fmnist 48, kmnist 48, mnist 48, stl10 44, svhn 48.

Unknown-rate and coverage checks
- No benchmark showed high unknown_rate in the inspection output.
- unknown_rate max by benchmark: cifar10 1.575100806451613e-05; fmnist 4.7253024193548386e-05; kmnist 1.575100806451613e-05; mnist 1.575100806451613e-05; stl10 1.575100806451613e-05; svhn 1.575100806451613e-05.
- No benchmark showed low cell_weight_coverage.
- cell_weight_coverage min by benchmark: 1.0 for cifar10, fmnist, kmnist, mnist, stl10, and svhn.

Spearman checks
- Within-benchmark pooled/z-scored full_primary.score: rho 0.7095188574306119, p 0.000999000999000999, n 288.
- Within-benchmark pooled/z-scored full_score.score: rho 0.7095188574306119, p 0.000999000999000999, n 288.
- Per-benchmark full_primary.score Spearman rho: cifar10 0.8497; fmnist 0.6769; kmnist 0.6004; mnist 0.8539; stl10 0.8481; svhn 0.7538.
- Per-benchmark full_score.score Spearman rho: cifar10 0.8497; fmnist 0.6769; kmnist 0.6004; mnist 0.8539; stl10 0.8481; svhn 0.7538.

Selection summary
- full_primary.score and full_score.score were selectable for all 6 benchmarks.
- Selector sign test for full_primary.score: rho_positive k 6 / n 6, p 0.015625; selector_gain_positive k 6 / n 6, p 0.015625; mean_selector_gain 0.18395138888888887.
- Selector sign test for full_score.score: rho_positive k 6 / n 6, p 0.015625; selector_gain_positive k 6 / n 6, p 0.015625; mean_selector_gain 0.18395138888888887.
- full_label.score was constant by design in this run and non-selectable; it was treated only as a diagnostic.
