Correctless

Correctless

AI coding agents are velocity machines. Correctless is the correctness system around them.

It does not try to make one agent smarter, faster, and more confident. It makes wrong answers harder to merge. A spec agent defines what correct means, a reviewer attacks the spec, a test agent writes failing tests without seeing the implementation plan, an auditor checks whether those tests would catch real bugs, a separate implementer makes them pass, QA hunts for defects, and verification checks the final code against the original rules.

Most AI workflows optimize for throughput. Correctless optimizes for trust.

Why This Exists

Fast AI development has a recurring failure mode: the same model plans the work, writes the code, writes the tests, reviews the result, and then reports that everything looks good. That is not review. That is self-confirmation with extra steps.

Correctless breaks that loop.

The point is not ceremony. The point is to create independent pressure from multiple angles before bad code reaches main.

What Correctless Gives You

A spec-to-merge correctness pipeline. /cspec -> /creview -> /ctdd -> /cverify -> /cdocs turns vague requests into testable rules, adversarial review, enforced TDD, verification, and documentation.

Agent separation by design. The spec author, reviewer, test writer, implementer, QA agent, mini-audit agents, and verifier are different agents with different lenses. The test writer does not know the implementation plan. The QA agent did not write the code. The verifier checks against the spec, not against the implementer's story.

Mechanical enforcement where prompts are weak. workflow-gate.sh blocks source edits during RED and all writes during QA. audit-trail.sh records what actually happened. sensitive-file-guard.sh…