Skip to content

v5.2.6 — 100/100 on Claude's Adversarial Test Suite

Choose a tag to compare

@sgroy10 sgroy10 released this 10 Mar 05:29
· 40 commits to main since this release

SpecLock v5.2.6 — AI Constraint Engine

Developed by Sandeep Roy

Highlights

  • 100/100 on Claude's independent 7-suite adversarial test battery — first ever clean sweep
  • 929 tests across 18 suites, all passing — enforced via pre-publish gate
  • 15.7ms average per semantic conflict check
  • Zero false positives, zero missed violations

What's New in v5.2.6

  • Safe-intent Pattern 6 — technology maintenance verbs (refactor, update, upgrade, optimize) are correctly allowed against exposure/secrets locks when the action doesn't touch secrets
  • Direct-overlap guard — prevents false negatives on buried data exposure (e.g., "update endpoint to include email" is still caught)
  • Pre-publish gatenpm publish now automatically runs all 18 test suites (929 tests). If any fail, publish is blocked. No more shipping regressions.

Test Results (v5.2.6)

Suite Score
False Positives (Suite A) 20/20 — 100%
Indian Gateways (Suite B1) 11/11 — 100%
DB/Auth/Security (Suite B2) 13/13 — 100%
Adversarial (Suite C) 15/15 — 100%
Question Framing (Suite D) 11/11 — 100%
Patch Gateway (Suite E) 9/9 — 100%
Diff Analyzer (Suite I) 7/7 — 100%
Overall 100/100 — 100%

Full Feature Set

42 MCP tools · Semantic conflict detection (65+ synonym groups, 80+ euphemisms, domain concept maps) · AI Patch Firewall (diff-native review) · Patch Gateway · Spec Compiler · Code Graph · Typed Constraints · Python SDK · ROS2 Integration · REST API v2 · Hard enforcement · HMAC-SHA256 audit chain · AES-256-GCM encryption · RBAC (4 roles) · SOC 2/HIPAA compliance exports · Policy-as-Code · SSO (Okta, Azure AD, Auth0)

Install

npx speclock setup --goal "Build my app"

Works with: Claude Code, Cursor, Bolt.new, Lovable, Windsurf, Cline, Aider