laya-nli-conflict-v10-l1 — research archive (round 10, "remove block B" lever), NOT delivered

⚠️ Research archive — NOT a delivered model, and this checkpoint is a local reconstruction, not a kernel-direct artifact: after a Kaggle output-pull failure the weights were recovered via CRC-verified partial zip salvage, metrics were recomputed on local CPU and τ(noul) refit locally (1.0288). Deviation disclosure: HANDOFF_NLI_V10.md §10.0. The current production head is slow-stack/laya-nli-conflict-v12 (v12, delivered 2026-10-07; it superseded v4).

Round 10 (2026-09-30 → 10-01) ran a preregistered three-run protocol. L1 = v9 corpus with block B deleted (single-lever experiment).

Headline results

• B2 passed for the first time ever (0.1216) — but B1 flipped 0.9078 → 0.2648, the swap-B1 variant also flipped, and the old-20 name-contradiction case missed; both deltas far outside the noise band (B2 −77pp, B1 −64pp) ⇒ block B is the load-bearing wall for "same-subject attribute contradiction ⇒ true" and must not be deleted
• verdict 10 PASS / 2 FAIL (old-20 19/20; bias diagnosis 11/14); main val 0.900 (recomputed)

Artifacts

file	value
model.safetensors	SHA256 ed77fc7d…aed7e7 (reconstructed weights; full hash in archive_sha256_manifest.txt)
rl_agent_config.json	τ(noul) = 1.0288 (locally refit); encoder jhu-clsp/mmBERT-base; bf16
metrics.json	val_accuracy 0.900, n_val 1000, no_rl true, reconstructed: local CPU recompute from CRC-verified salvaged weights
val_probs.json	frozen-val probability dump (CPU-recomputed)

Provenance

• Training: Kaggle GPU kernels daphnelaurent/laya-nli-conflict-ce (three-version sequence), dataset daphnelaurent/nli-conflict-pairs v15
• Round record & full gate table: kaggle_eval/HANDOFF_NLI_V10.md
• Sibling runs: laya-nli-conflict-v10-s2 (noise baseline), laya-nli-conflict-v10-l2 (v9-config rerun)