AI WORK AUDIT & VERIFICATION HARNESS
Review Bridge
AI coding you can audit.
Run coding agents with approved plans, enforced budgets, and independent review. Review Bridge records approvals, code changes, test results, and review findings so you can check what was authorized and what was delivered.
The ideas behind Review Bridge: AI adoption: managing delivery and verification responsibilities (HackMD, in Chinese)
MODULE / 01
Work from a sealed contract
Bridge creates and seals the contract from your work order without calling a model. You approve the card before work starts, giving the writer and reviewer the same requirements, scope, and budget to work from.
MODULE / 02
Independent, read-only review
The reviewer checks the work in a separate session with read-only access. Bridge checks its read-only configuration before work starts and saves the verification results.
MODULE / 03
Know what you authorized
Each approval records the work order and plan you authorized. The record cannot be changed afterward, and an edited work order needs a new approval before it can run.
MODULE / 04
Check changes against Git
For projects using Git, the report lists actual file changes, the writer's own report, and edits that predate the run in separate sections. You can compare them when reviewing the results.
MODULE / 05
Enforce budgets and time limits
Policy sets the writer's token budget, iteration limits, and time limits. Work orders can lower those limits, and agents cannot raise them. Bridge stops when a limit is reached; reviewer usage is recorded separately from the token budget.
MODULE / 06
Query the execution history
PostgreSQL stores work order versions, approvals, contracts, review verdicts, and results. Query records across runs to trace what was approved, reviewed, and delivered.
- Write a work order
- Approve the plan
- Implement and verify
- Independent review
- Review the results