RepairGate
Trust a code fix only after it passes checks it could not have been tuned against.
- Type
- Open-source tool
- Focus
- Code-repair verification
- Optional
- Local Ollama adapter
- Repo
- vanovri/repairgate
Problem
A proposed code fix should be trusted only after it passes checks it could not have been tuned against.
Approach
Candidate patchFrom a recorded or local model
Protected testsCannot be edited by the patch
Holdout checksIndependent verification
DecisionExported patch verdict
Build gallery
Results
- Overfitting patches are caught: in the recorded demonstration, a candidate that hardcodes the visible examples passes the visible tests but is rejected by the independent holdout checks.
- Tests cannot be gamed: patches that edit or remove tests are blocked before execution.
- Reproducible decisions: every attempt starts from the same original in a fresh copy, and recorded proposals replay deterministically without a model.
- Auditable output: the exported decision report carries diffs, test logs and file digests, and accepted fixes are never applied automatically.
This demonstrates gate behaviour, not an AI model's repair success rate.
Lessons
- Separating protected tests from holdout checks makes a pass result mean more.