verifying-completionlisted

# Verifying Completion Evidence before claims. "Should work", "looks right", and "probably fine" are not verification — they're guesses in a confident voice. This is a discipline skill: you will be tempted to claim done without checking, especially when tired or rushed, and the point is not to. ## When to use - Before any statement that work is complete, fixed, passing, or done — including paraphrases and expressions of satisfaction ("great, that's sorted"). - Before committing or opening a PR. - This gate does **not** scale down with task size; only *which* check is appropriate does. ## The gate 1. **Identify** — what check would actually prove this claim? 2. **Run** it fresh and in full — not a cached result, not a similar run from before. 3. **Read** the real output: exit status, pass/fail counts, the actual values. 4. **Confirm** the output supports the *exact* claim — not a near neighbour. 5. **Only then** state it, with the evidence. Skip any step and you have not verified — you have asserted. ## When no automated check exists Some claims — visual output, real-browser flows, realtime UI, subjective usability — have no command that returns pass/fail. The gate is then human-run, but it is still a gate: structure it as a traceable acceptance matrix with evidence, never self-certify a step a human must judge, and treat an unrun row as pending, not passed. Automate everything that *can* be; this covers only what genuinely can't. See `manual-acceptance.md`. ## Hard s