[frapercan]

What was wrong

compute_confusion_matrix_exclude_sparse restricted coverage to eligible rows and left every mass summing over all of them:

tp_totals   = tp_at_tau.sum(axis=0)      # all rows
pred_totals = pred_at_tau.sum(axis=0)    # all rows
eligible_rows = ((gt_sub != 0) & toi_mask[None, :] & (~excluded_mask)).any(axis=1)
metrics[:, 0] = ((pred_at_tau > 0) & eligible_rows[:, None]).sum(axis=0)   # eligible only

A protein whose TOI ground truth was entirely already known in t0 has nothing left to score. normalize() drops it from every denominator, and its predictions still landed in fp. Every one of them is a false positive by construction, because no term it could emit would have counted. Micro precision was charged for proteins the evaluation was not measuring.

The fix

Restrict pred_totals to eligible rows. tp_totals is restricted with it: a no-op under the current invariant, since a column can only be a TP when it is in TOI and not excluded, so an ineligible row already contributes zero. It is included so the two totals share one population instead of resting on that invariant.

Measured, on the fixture this repo already carried

tests/test_pk_coverage_bug.py::_make_pk_scenario has exactly the case: protein 1 is ineligible and predicts terms 2 and 3, neither excluded for it, neither in its truth.

column	before	after
n	2	2
tp	2	2
fp	4	2
fn	5	5
macro precision	1	1
macro recall	5/6	5/6

Only fp moves. The new test pins all six so a later widening of the restriction cannot pass unnoticed.

Direction and size, for the record

This raises PK micro precision, so published PK figures were conservative and not optimistic. On the 2026-08-30 campaign run the excluded mass is 3,385 of 2,648,460 in BPO (0.128 %) and 576 of 214,491 in MFO (0.269 %), from ten and eight proteins respectively. Order 1e-4 on a base of 0.15.

The macro columns the external board reports do not move at all, since an ineligible protein already contributed zero to both.

It does change stored f_micro_w figures, which is why this was raised as a decision rather than applied as a patch.

Out of scope, flagged not changed

The dense fallback compute_confusion_matrix_exclude has neither this restriction nor the earlier coverage one. It is gated behind CAFAEVAL_SPARSE=0 for A/B comparison, raises NotImplementedError when a per_protein_sink is requested, and has produced no published number. Bringing it into line would change what the A/B measures, so it is left alone and named here.

Downstream

PROTEA's _run_cafa_grid_artifact.py refuses to write a namespace while this defect stands, and tests/test_per_protein_grid_producer.py:1306 is a canary that fails the day this lands. Both need updating in the same migration, leaf first: this merges, PROTEA repins the resolved commit, then the guard and the canary follow.

Checks

Full suite before and after on origin/main: identical, no failures. With this branch: 36 passed, 6 xfailed, 0 failed. ruff check . clean.