laya-nli-conflict-v5 — research archive (round 5, arm A), 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 5 (2026-09-28) of the laya NLI memory-conflict head program by modusensus ran two arms on the same corpus: arm A = the v4 recipe (RL-style term + CE), this checkpoint; arm B = pure CE (see laya-nli-conflict-v5-ce).

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

• main val 0.893 (gate ≥0.896, ✗), old-20 acceptance 19/20 (✗), negation 4/5 (✗), new-10 8/10 (✗)
• passed: polarity diagnostics, val_soft 1 error/300
• Key finding: the RL term clashes with graded soft targets — arm B (pure CE) overtook arm A on every axis ⇒ from v6 on, the main arm is pure CE

Artifacts

file	value
model.safetensors	SHA256 8ac971b1…40f413 (full hash in archive_sha256_manifest.txt)
rl_agent_config.json	τ(noul) = 1.1384; encoder jhu-clsp/mmBERT-base; bf16
metrics.json	val_accuracy 0.893, val_ece 0.0321, n_val 1000, no_rl false
val_probs.json	frozen-val probability dump (calibration analyses)

Provenance

• Training: Kaggle GPU kernel daphnelaurent/laya-nli-memory-conflict-fine-tune v10, dataset daphnelaurent/nli-conflict-pairs v10
• Round record & full gate table: kaggle_eval/HANDOFF_NLI_V5.md
• Base: fine-tuned from convaiinnovations/laya-multilingual lineage; bf16 weights; checkpoint_latest/ (optimizer/step state) intentionally not uploaded