laya-nli-conflict-v12

A Laya decision head that answers one question about a memory pair:

新信息(new)与已有记忆(known)是否冲突？冲突=矛盾需更新旧记忆，兼容=一致或无关

noul question, two custom labels {"false": "兼容", "true": "冲突"} — "should the stored memory be superseded by what the user just said, or not?". Built for memory plugins that must decide whether to update a user's stored facts, preferences and constraints.

This is the delivered head as of 2026-10-07, superseding slow-stack/laya-nli-memory-conflict (v4). It is run r2 of a preregistered n=3 equal-status training protocol (HANDOFF_NLI_V12.md) whose 14 gate axes all passed — the first round to clear the full protocol. The weights in this repo are r2's; r1/r3 are published unmodified as sibling archives (v12-r1 / v12-r3).

Why r2 ships

Preregistration treats the three runs as equal for gating (no run selection), but exactly one set of weights must serve traffic. r2 is chosen by an objective, pre-stated criterion: it is the only run with all three acceptance batteries at full marks (old-20 20/20, negation-5 5/5, new-10 10/10); every other reading sits at or near the three-run median.

Three-run readings (the card must carry all of them — §1.3)

axis (gate)	r1	r2 (ships)	r3	gate	verdict
main val acc (median ≥ 0.896)	0.904	0.903	0.907	median 0.904	PASS
main val ECE (report)	0.0207	0.0209	0.0190	—	—
old-20 real cases (median = 20, no case missed in ≥2 runs)	19	20	20	20	PASS
negation-5 (median = 5, no repeat miss)	5	5	4	5	PASS
new-10 (median = 10, no repeat miss)	10	10	10	10	PASS
soft-conflict val_soft errors ≤3	2	1	1	≤3	PASS
label-surface swap	PASS	PASS	PASS	all PASS	PASS
polarity real·diag ≤1	PASS	PASS	PASS	all ≤1	PASS
polarity val_soft ≤1	PASS	PASS	PASS	all ≤1	PASS
band ≥8pp	PASS	PASS	PASS	all	PASS
conformal adopted rule ≥50% @ ≤35%	36/50@34.2% (s1)	34/50@33.4% (s1)	33/47@34.0% (s1)	all	PASS
bias diag ≥13/14 in ≥2/3 runs	13/14	13/14	13/14	3/3 runs	PASS
family rate, compat (median mean-p <0.5, ≤1 case p≥0.9/run)	0.0806	0.0919	0.0814	median 0.0814, 0 high-conf	PASS
negation family rate (median mean-p ≥0.5, ≤1 case p<0.5/run)	0.7872	0.8022	0.7930	median 0.7930, 1 low-conf/run	PASS
τ(noul) post-hoc fit (report)	1.0887	1.0914	1.1080	—	—
automation@5% (auxiliary, non-gating)	0.814	0.781	0.781	—	—

Same-config variance band: main val 0.903–0.907; acceptance thin-edge flips are single-run and never repeat across runs (r1 old-20 one miss; r3 negation one miss) — exactly the bimodal thin-edge behaviour the preregistered median aggregation models.

Protocol (multi-run, preregistered)

• n=3 equal-status runs, same kernel, same corpus (Kaggle dataset v17 = 15,950 rows = v11's 15,914 carried verbatim + 36 negation-known-conflict lever rows across 6 domains), same pure-CE recipe; no run selection/substitution/addition for gating.
• 14 gate axes per HANDOFF_NLI_V12.md §1.2; threshold numbers are frozen across v9–v12. Thin-edge axes (new-10 / negation-5) use median = full marks AND no case missed in ≥2 runs — preregistered before the runs (single-run flips at q≈0.2–0.25 made "all-three" a per-round lottery; the repeated-miss clause keeps detection power for真 weakness).
• Family-rate instruments (both preregistered): compat family (9 cases, gold=false) — median over runs of family-mean p_conflict < 0.5 AND every run has ≤1 case at p ≥ 0.9; negation family (8 cases, gold=true) — median over runs of family-mean p_conflict ≥ 0.5 AND every run has ≤1 case at p < 0.5. Both families are eval-side only, never trained on.
• Hygiene: the 35 acceptance cases have zero verbatim / field-level intersection with any training corpus since v6 (leak_audit 0/35; v12 gen asserts 0 overlap against 35 cases
  • 14 diag pairs + both families = 0/51). The chronic acceptance texts themselves are excluded from training.
• Verdict tool: protocol_verdict.py --protocol v12 (claims pack claims_v12.json + verify_claims_v12.py verified exit 0 before this publish).

Weights integrity

repo	run	model.safetensors SHA256
this repo	r2	6648d892449c13f81107b0b5891aca260293d176567ccbe432b8a6ef4af3f72b
v12-r1	r1	4269ec6bf8a64eee8aed9693395ac2f3d37119531f7d582869b05f048a27dc9b
v12-r3	r3	b2bf87bee5a0d379a1f8a633458389c389ceee7e61fe532ff52203d37d79f9f8

All three weights are 643,835,524 bytes (same architecture).

Usage

import laya, json

agent = laya.load('slow-stack/laya-nli-conflict-v12')   # or a local checkpoint dir
q = {"type": "noul",
     "instructions": "新信息(new)与已有记忆(known)是否冲突？冲突=矛盾需更新旧记忆，兼容=一致或无关",
     "labels": {"false": "兼容", "true": "冲突"}}
a = agent.predict({'known': '用户对花生过敏', 'new': '用户说他可以吃花生酱了'}, {'conflict': q})
p_conflict = float(a['answers']['conflict']['noul'])   # >= 0.5 -> conflict (update the memory)

The per-bucket post-hoc temperature (τ noul ≈ 1.091) ships in rl_agent_config.json and is picked up automatically by the laya runtime. Keep the label words exactly as above — option polarity is part of the wire contract (see the handoffs' polarity notes).

Provenance

• Training corpora (v3–v12, one per round, SHA256 manifest): slow-stack/nli-conflict-train-lineage · current-round rolling mirror: slow-stack/nli-conflict-pairs
• Frozen evaluation set (66 cases, 7 instruments, scoring recipe): slow-stack/laya-nli-conflict-eval
• Full index: the "Laya NLI memory-conflict head — v4 & three-run protocol" collection
• Base model: convaiinnovations/laya-multilingual; tooling & handoffs: kaggle_eval/
• All content synthetic; no personal or real-user data.