v2.3.1

v2.3.1: Prompt extraction false-positive fix

API Relay Audit v2.3.1 is a focused patch release. It improves refusal
handling without changing the 14-step audit surface, runtime profiles, threat
taxonomy, risk matrix, or Claude stream-identity baseline.

What Changed

• Normalize common Unicode apostrophes before matching refusal phrases.
• Treat explicit refusals containing weak words such as assistant,
  developer, or system prompt as non-leaks in Step 4.
• Keep structural prompt disclosures and strong markers such as kiro or
  hidden_prompt actionable even when refusal language is also present.
• Preserve self-corrected Test C output as a review signal rather than a
  confirmed prompt leak.

This fix was derived from #78
by @techotaku39.

Quick Start

AUDIT_SCRIPT_REF=v2.3.1
curl -fsSL "https://raw.githubusercontent.com/toby-bridges/api-relay-audit/${AUDIT_SCRIPT_REF}/audit.py" -o audit.py
python3 audit.py --key <YOUR_KEY> --url <BASE_URL> --output report.md

Verification

• Full pytest suite, version sync, standalone drift, and metrics drift must pass.
• Standalone audit.py SHA-256: 9cbab3b300bcca6e759500fbfb04f0d920c293556003ac2fc1080f398979de76
• The tagged audit.py, generated standalone output, release asset, and
  published SHA-256 file must be byte-identical before publication.
• Published release checks: /releases/latest and /releases/tags/v2.3.1.

Limits

• This release does not broaden Step 10 to arbitrary model families.
• API Relay Audit does not certify that a relay is safe.
• Do not publish API keys, private relay URLs, headers, raw responses, wallet
  material, or user data with audit reports.