How SpecLock catches AI violations that keyword matching misses

[sgroy10]

How SpecLock catches AI violations that keyword matching misses

Most code review tools look for keywords — "delete", "remove", "drop". But AI assistants are trained to be helpful, and they've learned to euphemize destructive actions in ways that bypass simple pattern matching.

SpecLock's semantic audit catches these.

The 10 euphemisms that slip past keyword filters

Here are real examples of how AI assistants describe destructive changes in ways that sound benign:

What the AI says	What it actually does
"Clean up old data"	Deletes production records
"Temporarily disable authentication"	Removes auth middleware
"Simplify the architecture"	Removes microservice boundaries
"Consolidate duplicate logic"	Merges distinct modules that happen to look similar
"Streamline the API surface"	Removes endpoints that external consumers depend on
"Modernize the data layer"	Replaces battle-tested ORM with untested alternative
"Remove dead code"	Deletes feature flags and fallback paths
"Optimize the build pipeline"	Removes test stages from CI
"Refactor for consistency"	Rewrites working code to match AI's preferred style
"Update dependencies"	Major version bumps with breaking changes

How SpecLock catches these

When you run speclock_semantic_audit or speclock_protect, SpecLock doesn't just look for keywords. It:

1. Understands intent — Maps the proposed change to its actual effect on the codebase
2. Checks against constraints — Compares the effect against your locked rules
3. Scores confidence — Rates how likely the change is to violate a constraint (LOW / MEDIUM / HIGH)
4. Explains the reasoning — Shows exactly which constraint is at risk and why

Example: Catching a "cleanup" that deletes auth

You have a lock: "Authentication middleware must be present on all API routes"

The AI proposes a commit with message: "Streamline request handling for cleaner code"

The diff removes authMiddleware from three routes.

Keyword matching: No match. No one said "delete" or "remove".

SpecLock semantic audit:

⚠️ VIOLATION DETECTED (Confidence: HIGH)
   Lock: "Authentication middleware must be present on all API routes"
   Change: authMiddleware removed from /api/users, /api/orders, /api/admin
   AI description: "Streamline request handling"
   Actual effect: Removes authentication from 3 protected endpoints

Try it yourself

npx speclock init
# Add a constraint
# Make a change that violates it using innocent-sounding language
# Run: npx speclock protect
# Watch it catch the violation

---

SpecLock is free and MIT licensed. Semantic analysis that goes beyond keyword matching.