epidebug-mm-v5

Option scorer over frozen all-MiniLM-L6-v2 embeddings, with hypothesis→evidence attention and tabular physics features. Its scorer uses interactions only, so it is invariant to option order by construction. Trained with randomized option order and a permutation-consistency KL, evidence-swap negatives, and the cause loss masked on AI4I mechanism rows. Selected run: seed 11, epoch 16, chosen on validation.

Bench v2.2 test (n=337, internal benchmark, not public)

model	hyp	follow-up	cause
v3	.332	.365	.433
v4	.294	.288	.383
v5	.522	.487	.549
best surface baseline	.214	.418 (mode length rank)	—

Evidence probes (change in hyp acc): blank −.139, swap −.300. The gold_first probe inserts the gold cause into the evidence text, so it measures leak sensitivity, not option position: +.196.

Caveats: the follow-up longest-option baseline already scores .389, so follow-up's margin over surface heuristics is modest (+.07). Results are from a single benchmark and one selected seed. Load with ckpt/v4.py.