[hallelx2]

Follow-up measurement on HAL-4071, recorded because it is the first question a reviewer asks after reading the probe.

n_poses sits on the boundary the probe draws. It is the number of modes Vina returned — a property of the run — yet it is in the feature matrix and arguably says something about the complex. The 13 Uni-Dock rows ran at exhaustiveness 16, so it could plausibly encode the engine.

Measured 2026-10-25

              count   mean    std   min  max
CPU Vina      574     8.833   0.728   2    9
Uni-Dock GPU   13     8.538   1.664   3    9

CPU values: {9: 531, 8: 21, 7: 9, 5: 6, 4: 4, 6: 2, 2: 1}
GPU values: {9: 12, 3: 1}

The distributions overlap almost entirely; the only GPU-only value is a single row.

n_poses alone, receptor-grouped macro-F1   0.1892
majority-class floor                       0.1894

Marginally below the floor, which is noise. No label signal, not proxying the engine — so it stays a feature rather than being removed on suspicion.

Also recorded

assert_provenance_is_not_a_feature was run against the rebuilt matrices, not only unit-tested: it passes on ml_ready_dataset.parquet and dataset_holdout.parquet, and none of the three provenance columns reaches X_train_eligible.parquet (478 × 2547), which is what the model actually sees.

Writing an assertion is not the same as it holding, and the doc should say which one was done.

No-Linear-Issue: a measurement recorded against HAL-4071, which is already Done; too small to warrant its own issue.