WAM19 Mac run notes

Machine summary:
- Apple M5 Max
- 18 CPU cores reported by hw.ncpu
- 137438953472 bytes memory (128 GiB)
- Darwin 25.5.0 arm64
- Runtime reported device_request=mps, mps_available=true, cuda_available=false, threads=1.

Setup and validation:
- Created local venv at .venv.
- Installed requirements from req.txt.
- Validation passed: PYTHONPATH=. PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/python -m pytest -q
- Result: 71 tests passed.
- No mathematical code, estimators, selectors, definitions, or theorem-facing measurement code were edited.

Exact fresh run command used:
source .venv/bin/activate
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
export OPENBLAS_NUM_THREADS=1
export VECLIB_MAXIMUM_THREADS=1
export NUMEXPR_NUM_THREADS=1
/usr/bin/time -p env PYTHONPATH=. ./mac.sh --fresh

Timing:
- /usr/bin/time real: 21517.94 seconds
- /usr/bin/time user: 202741.77 seconds
- /usr/bin/time sys: 2285.87 seconds
- Runner elapsed_seconds: 21465.249

Output files:
- out/mac.jsonl exists.
- out/mac_sel.json exists.
- out/mac_an.json exists.

Rows and child-correctness:
- Rows: 192
- child_ok count: 192
- Error row count: 0
- Benches: 20news, agnews, tiny, wikitxt

Per-benchmark ranges:
- 20news:
  - train_acc: 1.0 to 1.0
  - test_acc: 0.478 to 0.5195
  - full_primary.score: 0.11020770794297764 to 0.1549031107604381
  - full_score.score: 0.11020770794297764 to 0.1549031107604381
  - full_label.score: 0.3315291001562983 to 0.8699148640174156
- agnews:
  - train_acc: 1.0 to 1.0
  - test_acc: 0.2405 to 0.53
  - full_primary.score: 0.1544771664599878 to 0.22129444767689604
  - full_score.score: 0.1544771664599878 to 0.22129444767689604
  - full_label.score: 0.461857063981284 to 0.5969629538962525
- tiny:
  - train_acc: 1.0 to 1.0
  - test_acc: 0.051 to 0.677
  - full_primary.score: 0.1384954254675246 to 0.545884480057345
  - full_score.score: 0.1384954254675246 to 0.545884480057345
  - full_label.score: 0.9740856286330751 to 1.0
- wikitxt:
  - train_acc: 1.0 to 1.0
  - test_acc: 0.055 to 0.681
  - full_primary.score: 0.1336663791415479 to 0.5332838602904671
  - full_score.score: 0.1336663791415479 to 0.5332838602904671
  - full_label.score: 0.9774460351218633 to 1.0

Constant or missing full_primary.score:
- None. full_primary.score was present and non-constant for every benchmark.
- Unique full_primary.score counts: 20news=43, agnews=48, tiny=43, wikitxt=43.

Unknown-rate and coverage checks:
- No high unknown_rate observed.
- Max unknown_rate by benchmark:
  - 20news: 0.0
  - agnews: 0.0
  - tiny: 0.020833333333333332
  - wikitxt: 0.013020833333333334
- No low cell_weight_coverage observed.
- Minimum cell_weight_coverage by benchmark: 1.0 for all four benchmarks.

Within-benchmark Spearman against test_acc:
- 20news:
  - full_primary.score rho=0.5414805101164533, p=7.059333169113293e-05, n=48
  - full_score.score rho=0.5414805101164533, p=7.059333169113293e-05, n=48
- agnews:
  - full_primary.score rho=0.8806622366008454, p=1.569820664354294e-16, n=48
  - full_score.score rho=0.8806622366008454, p=1.569820664354294e-16, n=48
- tiny:
  - full_primary.score rho=0.8441822776066767, p=4.8068749227079703e-14, n=48
  - full_score.score rho=0.8441822776066767, p=4.8068749227079703e-14, n=48
- wikitxt:
  - full_primary.score rho=0.8832048224486589, p=9.841016349194067e-17, n=48
  - full_score.score rho=0.8832048224486589, p=9.841016349194067e-17, n=48

Analyzer within_benchmark aggregate:
- full_primary.score: rho=0.7670279814062134, p=0.000999000999000999, n=192
- full_score.score: rho=0.7670279814062134, p=0.000999000999000999, n=192

20news repaired spread check:
- 20news test_acc has 30 unique values and is not constant.
- Range is 0.478 to 0.5195; spread is 0.04149999999999998.
- out/mac_an.json marks by_dataset.20news.target_spread_ok=false with note low_spread_or_constant_target.
- Conclusion: repaired 20news is child-correct and selectable, but does not have adequate test spread by the analyzer's spread criterion.

Validity conclusion:
- This is a valid child-correct selection run in the mechanical sense: all 192 rows are child_ok, error row count is 0, all assigned Mac benchmarks ran, full_primary.score is present and non-constant for every benchmark, unknown rates are low, and cell_weight_coverage is complete.
- Caveat: 20news remains a low-spread benchmark according to out/mac_an.json, so downstream interpretation should retain that spread warning.
