Whose Oracle Is It? The Authority Problem in Deployed Agentic Program Repair

Abstract

Tests play two roles in automated program repair (APR): they are the evidence that accepts a patch, and they are repository artifacts that may themselves need repair. Benchmarks resolve the tension by making evaluator-owned tests immutable; deployment cannot, because tests legitimately evolve with the code they check. We argue that the resulting question, who may authorize a change to the evidence that accepts a patch, is a problem of allocating authority, distinct from whether a patch is correct, and that no diff separates an authorized change of requirements from an agent weakening its own acceptance test. We propose an admissible repair witness, evidence that exists before the agent runs and that the agent cannot rewrite, and a two-zone policy that protects it while the rest of the test surface evolves under review. In 230 merged bug-fix pull requests from nine repositories, 160 of 172 test-changing repairs (93.0%) edit a test artifact that already exists, so locking the test surface is not an option. Of the 172, 87 delete test lines, but only 53 delete an assertion, a test case, an expected output, or a whole test file: escalating on deleted lines alone sends half of all test-changing repairs to a human, while reading what was deleted sends at best under a third.