LHTB’s official numbers require the modified Harbor in that repo. The 0.20.x drop-in is harbor/patches/single_step.py.harbor-0.20.0. 30 of the 46 tasks set continue_until_timeout. Stock Harbor ignores the flag, so those tasks run single-shot and score lower.

That module freezes the agent’s process tree during each verifier pass. An audit of one 46-task sweep found 14 of 17 perfect scores came from reading grader artifacts: /logs/verifier/pytest.log, scorecard.json, leftover /tmp/pytest-of-root fixtures, and a background loop that copied /tests while it was mounted. HB_VERIFIER_FEEDBACK_MODE=diagnostic is not benchmark-comparable. The default is binary.

The published July snapshot predates that isolation. I would not mix a run that used stock Harbor, diagnostic feedback, or an unfrozen sandbox with those numbers. Until those three are named, the score is a harness result.