laya-nli-conflict-v8 — research archive (round 8), NOT delivered

⚠️ Research archive — NOT a delivered model. This checkpoint failed its round's acceptance gates and was never shipped. The current production head is slow-stack/laya-nli-memory-conflict (v4). Uploaded 2026-10-02 for provenance/backup while round 11 (multi-run verdict protocol) waits for Kaggle GPU quota.

Round 8 (2026-09-29) added 120 "same-subject compatible attribute ⇒ false" rows on the v7 corpus (carried blocks byte-identical): pet_attr 80 (= 40 pet_name B2-same-shape with swapped literals + 40 pet_benign) + doctor_attr 40.

Headline results (frozen 1000-pair main val unless noted)

• B2 dose-response confirmed: 0.8985 → 0.7491 (−15pp, the only large diagnostic shift) — the v7 mechanism attribution was in the right direction, but the +40 same-shape dose was insufficient to cross 0.5
• main val 0.896 (pass, err 104); failed: val_soft 7 (✗, all holdout degree/barber consid rows that flip across runs — training variance, not attr regression), polarity val_soft +6 (✗), conformal s2 abstain 35.4% (✗, 0.4pp over the 35% line; capture 42/54 = 77.8% was strong)

Artifacts

file	value
model.safetensors	SHA256 bcd1b158…174088 (full hash in archive_sha256_manifest.txt)
rl_agent_config.json	τ(noul) = 1.1240; encoder jhu-clsp/mmBERT-base; bf16
metrics.json	val_accuracy 0.896, val_ece 0.0319, n_val 1000, no_rl true
val_probs.json	frozen-val probability dump (calibration analyses)

Provenance

• Training: Kaggle GPU kernel daphnelaurent/laya-nli-conflict-ce v4, dataset daphnelaurent/nli-conflict-pairs v13
• Round record & full gate table: kaggle_eval/HANDOFF_NLI_V8.md
• Base: fine-tuned from convaiinnovations/laya-multilingual lineage; bf16 weights; checkpoint_latest/ intentionally not uploaded