Bound the task
Record scope, non-goals, allowed files, tests, and acceptance criteria.
Evidence-based adversarial review for coding agents
Your agent says “Done.”
Do you believe it? Meaning Assurance separates a worker’s claim from frozen evidence, independent challenge, moderated findings, and the human decision to accept.
Agents propose. Evidence is verified. Humans decide what to accept.
A protocol, not another agent
Confidence is not a control system.
A second agent agreeing with the first can still be synchronized confidence. The protocol makes the evidence and disagreement inspectable before a human accepts the result.
Isolate implementation
Keep worker changes in a Git worktree instead of the source checkout.
Freeze review evidence
Hash the exact snapshot reviewers receive so later edits cannot rewrite the review basis.
Keep authority human
Moderate findings and disagreement; never let an agent approve its own work.
Inspect the trail
One command. A complete bounded decision.
The deterministic demo uses a synthetic fixture, requires no API key, launches no external model, and writes only to a new output directory you choose.
$ powershell ... .\scripts\Run-Demo.ps1
Worker claim: complete
Evidence: one declared test missing
Scope: one changed file outside scope
Findings: 2 confirmed / 1 rejected / 1 not testable
Human decision: not accepted
Strategy is also reviewable
Brainstorm directions. Preserve the objections.
The first-party strategy-review case shows competing directions, adversarial questions, moderator decisions, and a read-only audit trail. It demonstrates the workflow—not market validation or proof that a strategy will succeed.
Read the bounded case- First-party internal dogfooding, not independent validation.
- Git worktrees isolate files, not operating-system permissions.
- Frozen evidence can still be incomplete.
- The protocol does not guarantee correct code or decisions.
Delegate the work. Keep the authority.