MODEL CARD
breachlab-guard-v1: Qwen3-0.6B + LoRA fine-tuned with SFT, temperature-calibrated, served on CPU. Research demo; not a claim of production safety.
Measured on gold_test (v1)
- Attack success (ASR)
- 0.4%
- False positives (FPR)
- 1.8%
- F1
- 0.987
- AUROC
- 1.000
- p95 latency
- 2063 ms
qwen3-0.6b-sft-int4-v2: guard v2 (ADR-0094): same recipe on data-v2, int4 q4b32 ONNX as served on Vercel; T=1.124, own threshold=0.9646 fitted on v2 val (FPR<=2%); gold_test is unchanged from v1; held-out short benign opener FPR 6.3% (v1 54.3%); p95_ms = service-side model time, concurrency 1, warm, 1 vCPU, 120 random val rows (docs/lanes/T-v2-latency.json; v1 was 500 rows, so not a like-for-like comparison); the 300 ms target is NOT met
LINKS
Some links may be private until the public release.