Benchmark: governed loop stopped at $2.30, ungoverned burned $5.20

[Keesan12]

One benchmark changed how I think about unattended coding loops.

Same flaky-CI-gate task:

• ungoverned loop: $5.20 and still failed
• MartinLoop: $2.30 with a clean stop and a replayable run record

The interesting part is not just lower cost. It is having an actual stop contract around the loop: remaining budget, verifier result, failure class, explicit halt reason, and a JSONL record you can inspect later.

If you are already running Claude Code, Codex, OpenCode, or your own harness unattended, two concrete questions:

1. what would you need before you would actually try a budget + verifier layer?
2. which hurts more in practice: silent overspend or fake success?

If you want to pressure-test it, the repo is here: https://github.com/Keesan12/martin-loop and npm is here: https://www.npmjs.com/package/martin-loop

If it is useful, star the repo. If you try it, I want the failure cases and rough edges, not polite feedback.