Stop reviewing lines. Verify claims.
An agent's pull request lists its claims. A second AI tries to break each claim with tests it runs. You read only the code that the evidence cannot cover, answer what the AI refused to decide, and sign.
Review REQ-2283228 lines agent pr · 7 claims · 4 verified by tests · 2 refused by ai · 25 lines read by you
reviews
req-2283Fix the closed generic struct crashreq-2234Analyzer with SuppressSourceGenerationprinciples
Evidence or silenceNo evidence, no verdict. The AI refuses instead of guessing.
Falsify, don't summarizeRevert the fix and the tests must fail. Mutate a line and a test must notice.
Show the evidence, not a scoreExecuted, probed, traced or read. A person can check each one.
Attention follows riskPeople read the risky lines. Everything else gets an evidence check.
Refusing is a resultProduct intent and trade-offs go to a person as one clear question.
One name per mergeA person signs. The AI never approves.
Trust is earned per areaSpot checks find misses. One escape sends an area back to full reads.
+ trade-offs
Overhead on tiny changes. One-click sign when an area is trusted and every claim is verified. In return: every merge has an owner and a record.
Needs runnable tests. Without tests, claims go to a person. Slower, but correct.. In return: proof anyone can rerun.
Spot checks take human time. 10% at first, less where the record is clean. In return: we can measure what the ai misses.
A second model and CI minutes. Probe only the risky hunks. In return: fewer shared blind spots.
+ limits
It cannot check intent that nobody wrote down. It can only ask.
A bug that both the tests and the reviewer miss gets through. Spot checks measure this.
Weak tests from the author agent pass. Probes catch some of them.
Risk rules are written by people and stay wrong until escapes correct them.
Checkers are deterministic, but writing the rule or the spec is not. A rule counts only after it hits the original bug, a spec only after it catches a planted one.