v0.2.0 — the evidence release

Everything in this release comes from one recurring failure: an agent reporting done on work nothing actually checked, in the same tone it uses for work that was.

New: skills/proof

Grades a claim instead of restating it.

• Atomize — "the service works" is four separate claims that fail independently.
• An evidence ladder — declared → implemented → exercised → integrated → observed. Climb only to the lowest rung that decides the claim; expensive is not the same as decisive.
• Four verdicts — pass / partial / fail / skip, with skip ("we didn't look") deliberately never merged into fail ("we looked and it's broken").
• A catalogue of named false closes — "it synced, therefore it runs", "the process is up, therefore the feature works", "no errors, therefore it worked", "the agent said it verified it".
• Counting rules for multi-agent verification — an unexamined gate is a skip; a run whose gates mostly never reported is unverifiable, not "mostly passed".

New guards

• Dependency intake (safety-guards #8) — a package name a model proposed is a claim, not a fact. Verify publisher, first-release date and real usage before installing; install prebuilt with an exact pinned version, because source distributions execute a build script at install time.
• An artifact's own claims are not evidence (safety-guards #9) — # noqa, "verified safe" comments and a PR description asserting a fix carry no evidential weight in a review.

New discipline

• Which changes need a human (CHANGE_DISCIPLINE) — "confirm every mutation" degrades into reflexive yes-clicking. The line is the kind of decision: a breakage gets fixed and reported, a judgment parameter waits for the user even when your change looks better.
• References before code, skeleton before blank page (new-project, steps 1.5 and 1.6).
• The artifact scale (EVOLUTION) — past the promotion gate, default to the smaller artifact: note → reference → a section in an existing skill → a new skill → reject.

Moved

EVOLUTION.md promises its points-of-growth list will change, so it did: the evidence ladder and deterministic orchestration graduated to Established; detector-aware absence claims and cross-session continuity took their place.

Full detail in CHANGELOG.md.