epidebug-mm-v6

This is v5, an interaction-only option scorer over frozen all-MiniLM-L6-v2 that is invariant to option order, with evidence-swap negatives and a permutation-consistency loss, plus:

• Option-only debiasing: a debiased product-of-experts with a train-time bias head that sees only the option texts (embedding plus row-relative style features: length rank, commas, digits, hedges). The main scorer is trained on main + log p_bias.detach(), so it only has to explain what option style can't. A gradient-reversal adversary (λ=0.3) predicts the gold from the scorer's option projections. Both heads are dropped at inference.
• Temperature scaling fitted on validation (hyp 0.99, follow-up 0.94, cause 4.32). The temperatures are stored in the T buffer and already applied to the outputs.

The selected run is seed 11, epoch 22, chosen by validation score. Load it with ckpt/v4.py.

Bench v2.4 test (n=337; internal benchmark, not public)

model	hyp	follow-up	cause	ECE hyp / fu / cause
v5	.401	.472	.549	.121 / .151 / .325
v6	.478	.665	.585	.091 / .037 / .139
best surface baseline	.211	.377	—
option-only style_gb (data probe)	.389	.579	—

Memorization caveat: 75 of the 337 test hyp golds also appear as train golds. On those rows v6 scores .933 hyp / .987 fu (v5: .840 / .773). On the other 262 rows v6 scores .347 / .573 (v5: .275 / .385). Most of the headline hyp number comes from the overlap rows.

Probes (change in hyp acc): blank evidence −.240, evidence swap −.318, gold cause inserted into the evidence +.300. The permutation probe gives identical predictions under every ordering.

Style reliance (v6 accuracy when style_gb is right vs wrong): hyp .534 vs .442 (v5: .481 vs .350); follow-up .667 vs .662. The hyp gap is smaller than v5's but not gone.

Caption pairs (n=24, private): hyp .667 vs v5's .917, a regression on this small set. There were 0 flips among the 8 clean-correct pairs.