Keeping Open Decisions Open: Executable Rules for Agent-Built Software — a Field Case

A coding agent asked to implement a requirement has no way to leave an open question open: it decides, and nothing in the artifact marks the choice. We report a field case of a small product (a mutual-attention app for older adults living alone, Android and iOS) built almost entirely by coding agents under one human ratifier, observed over about ten weeks. We describe three mechanical checks run against the code: CANON (every string shown to a person exists verbatim in a governed catalog), COVERAGE (every verifiable contract clause cites a test that still exists) and NO-IMPLICIT-CLOSE (for every decision recorded as open, a test fails if the code behaves as if one alternative had been chosen). The first two adapt established practice; the third is, to our knowledge, not described elsewhere. We report what the checks found on first execution, including errors in the verification artifacts themselves. One case, observed by its own author: a documented method and its first field record, not a validation.