CentSentry: Green Tests Aren't Proof

A payments app must never create, destroy or double-spend money: not under 50 concurrent requests, not on a retry, not when history is corrected. AI agents that write code and then test it only test what they built. CentSentry is four Claude Code agents in one BAND room with one rule: no agent accepts its own work. The Architect turns the spec into numbered requirements and never writes code. The Developer builds test-first and cannot approve itself. The Tester writes tests from the spec alone and never sees the code. The Reviewer rebuilds every handoff from a clean clone with no network and accepts or rejects. The mandates are generic; everything wallet-specific lives in one dispatch file. The proof: in stage 1, 663 of our tests and all 147 official checks were green. The reviewer probed the arithmetic limit by hand, a wallet at 2^53 plus one cent, and the cent vanished to floating-point rounding. Rejected. The architect recorded a rule, the tester added edge-case tests, the developer fixed it with exact integer math, and the reviewer accepted it. Ten real defects caught across six rejected builds, each one missed by the suites. Result: all four stages claimed in the official isolated harness, 193/193 shipped checks, verified from a fresh clone. Send, request, split, settle, hold and capture, bitemporal corrections and statements, refunds and batch corrections, plus a responsive UI with history, refunds and dark mode. Zero lines of product code written by a human. The full room log is in the repo.