← ClaudeAtlas

mellions-falsificationlisted

Load this before a test, a mutation, a revert arm or a green run is cited as proof that a fix holds — and again before writing "falsified" in a pull request. Triggers — "prove it holds", "falsify", "neutralise the fix", "watch it fail", "the test passes", "is the test real", "mutation", "revert arm", "the suite is green", "proof section". Not for deciding what to do with the result (mellions-reasoning) or for establishing what is true in the first place (mellions-deep-research).
LetA-Tech/mellions-coxen · ★ 1 · AI & Automation · score 67
Install: claude install-skill LetA-Tech/mellions-coxen
<!-- Mellions Engineer | LetA Tech Ltd. | leta@letatech.ca --> # Establishing that a fix holds A green run is a claim about the harness until something shows it could have been red. ## The copy you falsify in One copy per arm: a neutralisation left from the previous arm can mask the next. `git archive <rev>` carries that commit, not the index: a staged, uncommitted test is absent. Never falsify by restoring a working tree: it discards uncommitted work — yours or another session's. With one arm and nothing uncommitted at stake, toggling the edit in place — red, then green — is the arm. ## The oracle Where a test asserts the output equals what the mechanism under test computes, it has no independent side: whatever moves the mechanism moves both halves, and the test stays green through the defect it was written for. Compare against something the mechanism cannot reach — a literal, a figure from the requirement, a golden file, a reference implementation, an invariant, a value read across the boundary the fix crosses — never a call into the code under test. ## The revert arms Revert every part of the fix at once and watch the named tests fail. A complete revert that stays green has four readings, the third pointing at deleting the test: the edit did not land — confirm it where the test reads (off disk, in the database), since an edit that no-ops (a `sed` that matched nothing, an unexpanded variable) reads exactly like a pass; the tests never reach the fix; the fix is idle;