Rewrite Rate: a pre-registered protocol for measuring how often coding agents make test suites pass by editing the tests

Coding agents sometimes make a failing test suite pass by editing the test rather than the code. Every existing signal for this is a proxy — a test file changed, an assertion count dropped, a skip appeared — and none of them establishes that the change to the test is why the suite went green. Reverting the agent's test edits and re-running the suite does establish it. This document registers, before any data is collected, the method for measuring it. The Rewrite Rate is the share of runs an agent finished green that turn red when its test edits alone are reverted. Fixed here in advance: the task set (33 tasks across 9 public Go repositories, fingerprint blake2b:ee02cf85bdf681de63d60ed0e83429b2), the agents and their pinned versions, the prompt, the counting rule, the stability filter, the exclusion criteria, and what is published regardless of outcome. The tool used to take the measurement is written by the author of this registration. That is a conflict of interest, and stating the method in advance is the only thing that makes the result worth reading. A Rewrite Rate near zero would mean the tool's premise is wrong; that result is published on the same terms as any other. Task set, construction script and analysis: https://github.com/angelnicolasc/Warrant