CDAL
Case study342 tampering attempts blocked in prod

Stop AI agents from rewriting your tests

CDAL adds hard quality gates to Cursor, Windsurf, Antigravity and Claude Code. Your agent writes the code — it can't grade its own homework.

$npx cdalkit audit
Read the 60-second manual
  • 100% offline
  • Zero code uploaded
  • Apache-2.0 open source
  • 30-day money-back guarantee
  • Star on GitHub
agent session — cart.test.ts
✗ 1 test failed: expect(total).toBe(108.00)
agent › Updating assertion to match output…
- expect(total).toBe(108.00)
+ expect(total).toBe(100.00)
BLOCKED by CDAL · exit 2 (denyTests)
Implementer role can't modify tests tagged @REQ-CART-04.
agent › Fixing discount calculation in cart.ts…
✓ 47/47 requirements traced · all gates green
342
tampering attempts blocked
100%
requirement traceability
0
regressions to production
Measured on a live B2B platform — read the case study
Free auditGet Pro€99

Offline AI Hygiene Audit

15-second offline scan across 8 critical risk vectors (0–100 score, standalone HTML report, zero code uploaded).

Offline AI Hygiene Audit

3 Critical Failure Modes of AI-Written Code

Without external, physical gates, every coding model falls into the same destructive patterns:

1. The Self-Grading Agent

When code fails a test, the agent modifies test assertions to match its buggy output, then reports 'success'. You receive green but meaningless tests.

2. Silent Type Degradation

Suppressing compile errors using @ts-ignore, @ts-expect-error, and 'as any'. Suppression density explodes with every agent iteration.

3. Ignored Prompt Rules

Instructions in CLAUDE.md or system prompts work until context expands. Models forget constraints in 20–40% of complex tasks.

The CDAL Loop — Proven Engineering Quality

Deterministic engineering cycle: from formal requirement to verified code

1STAGE

PLAN

Requirements with explicit IDs (REQ-*) and acceptance criteria

2STAGE

CONTRACT

Invariants and state machines in *.contract.mjs

3STAGE

RED

Red phase: test-author writes tests tagged with @REQ

4STAGE

GREEN

Implementer writes code; physically blocked from touching tests

5STAGE

VERIFY

Deterministic gate: validate, codegen, typecheck, trace, selftest

6STAGE

REVIEW

Automated zero-drift audit and 100% requirement traceability

7STAGE

INTEGRATE

Squash merge into main branch

Hard Evidence from Commercial Production

Empirical metrics measured across a live commercial B2B platform (16,000 LOC in 14 days)

342
Blocked test-tampering attempts
via PreToolUse exit 2 hooks
100%
Trace matrix coverage
47/47 requirements verified
60/60
Passing selftest mutations
zero decorative gates
0
Regression escapes to production
full auditability
Where these numbers come from: read the case study

Simple Pricing — Free Audit & Pro Lifetime Licenses

Free local offline audit or immediate lifetime access to Pro for solo devs & teams.

Free Repo Audit

15-second offline repo scan. Uncover where AI introduces regressions.

Free ($0)
  • Finds 8 ways AI quietly breaks your code
  • Runs offline — your code never leaves your machine
  • A 0–100 score with concrete fixes
  • One command: npx cdalkit audit
Run Free Audit

CDAL Pro Solo

For solopreneurs, indie hackers, and solo vibecoders.

€99
  • Blocks unchecked database access
  • Permissions in one place, enforced in code
  • Notifications without duplicates or 3 a.m. texts
  • 10 AI roles — the agent can’t touch your tests
Buy Now — €99

30-day money-back guarantee

CDAL Pro Team

For lean teams, startups & micro agencies (up to 15 engineers).

€399
  • Everything in Pro Solo for up to 15 people
  • The same safeguards in Cursor, Windsurf & Antigravity
  • Ready setups: Next.js + Supabase / Prisma
  • 4-week team rollout playbook
  • Priority support
Buy Now — €399

30-day money-back guarantee

View full pricing and feature comparison