M6 Security Detector v2.4-dt (FINAL window delivery)

Six-task safety detection system, single trunk + single inference pass -> all results. Window delivery per Lead final adjudication (Option A, 2026-10-22): v2.4-dt @ QA operating point.

Six-generation arc (all zero-FP)

v0 baseline 0.0-2.4% -> v1-2 0.681 -> v1-3 0.818 -> v2 0.866 -> v2.4 0.910 -> v2.4-dt 0.935 macro recall

Gate card (QA point)

Gate	Result	Line	Verdict
G1 recall T1..T6	100 / 99.7 / 97.3 / 100 / 100 / 100 (CI95>95)	>95%	PASS
G2 FPR (T1/T2/T3/T5/T6)	0.000 / 0.000 / 0.021 / 0.000 / 0.000 %	<0.5%	PASS
G2 FPR (T4)	2.333%	<0.5%	4.9x gap (span-channel, post-window roadmap)
G3 challenge FPR	0.35-2.13%	<0.1%	gap (threshold/domain, characterized)
G4 OOD pass-rate	35.4%	>=95%	2.7x gap (dual-vector, post-window roadmap)
G5 red team	92.4% (157/170)	90%	PASS
Challenge pos recall	100% (345/345)	disclose	PASS

Architecture (ADR-1 + ADR-2 + ADR-2b chain)

Backbone: qwen3-0.6b-sec-m5a-r5 (frozen 0.6B, 28L), LoRA-adapted trunk (r16, L14-28, 1 epoch on v2.2 data wave)
Features: multi-layer {4,12,20,24,28}x{last,mean} -> 10,240-d; dual trunk-MLP probe routing (T4)
7 heads (PII = doc + BIO span dual head, span min-length gate), two-band logit-domain thresholds, Mahalanobis OOD suppression
Deterministic inference (determinism 97f2860a), 35 unit tests green

Validated findings

Probe architecture decisive vs r5-logprob baseline (68-98% vs 0-2% recall at zero-FP)
LoRA representation learning solves the four head-only capability-limit cases (longctx 26.7->100%, A3 0->8/8, red team 52.9->90.6%)
Supervised data is the active ingredient for T1 (challenge FP 23.0% -> 0.33%)
Head-level delivery preserves trunk (joint training structurally unstable: 4/5 ep2 basins); deterministic post-processing (frozen trunk + fresh heads) is the correct path
Span channel is load-bearing (9% heldout TP) - contextualization, not removal

Files

probe-heads.pt (90MB): dual-trunk + 7 task heads
ood_centers.json (32MB): Mahalanobis centers
calib.json: QA-point runpoint thresholds (logit domain, threshold_space per schema v0.3.24)
config.json / head_structure.json / metrics.json / build_manifest.json

Usage

Load backbone qwen3-06b-sec-m5a-r5 + LoRA adapter (training repo), extract features via scripts/m6_feat_probe.py, apply heads via scripts/m6_infer.py. Full reproduction commands in the companion GitHub repo (reports/m6/task13-final-adjudication.md section 7).

Post-window roadmap (ranked)

T4 context-gate (span contextualization) -> G2 ~0.1-0.3%
G3 calibration policy re-check (threshold domain, zero training)
G4: learned OOD discriminative head + span contextualization (both vectors)

Downloads last month: 35