Sample Compliance Evidence Pack

Why this pack does not rely on an LLM

• Detection: 670+ ATR rules with 97.2% Garak recall (650 samples); 0.2% FP on the 498-sample SKILL.md corpus, and lane-based benign-gate FP of ~0.24% enforce / ~9% hunt — zero LLM tokens
• Report narrative: rendered directly from rule YAML metadata (title / severity / description / compliance citation)
• Assessment findings: assessors read firewall, encryption, TLS, and logging configuration on the customer machine
• Audit integrity: every file carries SHA-256 + HMAC-SHA256 signatures (production signatures use a customer-bound key)
• Pilot / Enterprise customers never need to bind an OpenAI / Anthropic key to the detection path