laya-nli-conflict-v6 — research archive (round 6), 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 6 (2026-09-28) was a data-hygiene reset: all 35 acceptance sentences scrubbed from the training corpus (leak_audit.py 0/35 assertion enforced), unrelated control rows added (negation-family 400 / counterfactual 300), five negation families ×80 (pet/never/relapse/dneg/nocar), pure-CE main arm (RL term dropped after round 5's arm-A collapse). This round also re-baselined the numbers of v2–v5 under the sentence-reuse-free accounting.

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

• main val 0.903 ✓ (program best at the time; refuted "leakage inflates the score"), old-20 20/20 ✓ (unrelated-supplement false-positives fixed), negation 5/5 ✓ (first time), confidence band 16.29pp kept
• failed: new-10 8/10 (✗), val_soft 5 (✗, all holdout degree/barber rows), polarity +2 (✗), conformal FAIL — root cause: 52 of 97 val errors sit at confidence ≥0.92, which no confidence-threshold rule can catch (the compression-band disease; abstention moved to a surface-consistency rule in v7)

Artifacts

file	value
model.safetensors	SHA256 11facce6…84b4e0 (full hash in archive_sha256_manifest.txt)
rl_agent_config.json	τ(noul) = 1.0905; encoder jhu-clsp/mmBERT-base; bf16
metrics.json	val_accuracy 0.903, val_ece 0.0192, 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 v2, dataset daphnelaurent/nli-conflict-pairs v11
• Round record & full gate table: kaggle_eval/HANDOFF_NLI_V6.md
• Base: fine-tuned from convaiinnovations/laya-multilingual lineage; bf16 weights; checkpoint_latest/ intentionally not uploaded