Spaces:

---

gaurv007

/

ClauseGuard

like 0

Sleeping

ClauseGuard v4.2 — Deep Audit Fixes

🔴 Critical Bug Fixes

1. NLI Contradiction Detection was BROKEN — pipeline("text-classification") with dict input silently failed for cross-encoder. Replaced with CrossEncoder.predict() from sentence-transformers which accepts (text_a, text_b) tuples correctly. Contradictions now actually work.
2. BoundedCache Race Condition — OrderedDict compound operations are NOT atomic. Added threading.RLock to prevent crashes under concurrent Gradio requests.

🟠 High-Severity Bug Fixes

3. Extension Risk Formula Mismatch — Local fallback used (weighted/clauses)*100 (normalized). Backend uses 100*(1-1/(1+w/30)) (diminishing returns). Same contract got completely different scores. Fixed to match.
4. Extension API_BASE URL Wrong — Was pointing to non-existent -api subdomain. Fixed to correct Space URL.
5. Missing Regex Labels — Indemnification, Confidentiality, Force Majeure, Penalties had regex patterns but no entries in RISK_MAP/DESC_MAP. Added.
6. Inconsistent Model Name — compare.py used "all-MiniLM-L6-v2" without prefix while chatbot.py used "sentence-transformers/all-MiniLM-L6-v2". Could cause duplicate downloads.

⚡ Performance Improvements

7. Pre-compiled ALL regex patterns — Clause classification (45 label patterns), obligation extraction (25+ patterns), compliance negation (8 patterns), false positive filters, time patterns, party patterns — all compiled once at module level instead of per-call.
8. API Rate Limiter Memory Fix — Stale IPs now cleaned up every 60s regardless of dict size (was only cleaning when >1000 entries).

🔒 Security

9. API CORS localhost — localhost:3000/3001 origins now require explicit CORS_ALLOW_LOCALHOST=true env var instead of being always allowed.

📝 Housekeeping

10. Version updated to v4.2 across app.py docstring, README.md, and changelog.