Engineering · core domain

Coding agents that hand back proof.

Scope a code change, assign worker and verifier roles, and specify the tests required for review.

Bounded assignment

Each task declares the files it may touch and the check that proves it. Work outside the bound is rejected, not merged.

Independent checks

The verifier is not the worker. It reruns tests on the exact resulting state, in isolation, and records the outcome.

Release hold

Passing checks lift nothing automatically. A person approves an exact, reproducible state, or sends it back.

Verification fixture

A failed check holds release.

Synthetic test results only. No repository access or agent execution.

Evaluation fixture · no real repository

Independent verification

Simulation
  • Patch stays within declared files
  • New tests fail on pre-state
  • Full suite passes on exact post-state
  • Static checks clean

Scope the next change.

Draft tasks, required inputs and human review checkpoints.