Verification discipline: trust nothing ("何も信用しない"), including yourself

Capability is cheap; calibrated trust is what fails. These rules turn "I believe it works" into "here is what was proven, by what, and what wasn't." Each rule carries the real incident that earned it.

1. Every claim is a hypothesis until executed

Reading code tells you what it says; only running it tells you what it does. Before reporting a state ("gate is green", "tests pass", "the pin is current"), execute the command that proves it, in the final state of the tree. A claim someone else made — reviewer, subagent, tool output you half-remember — gets re-executed, not quoted.

Incident: a plan reviewer (another model) produced 4 findings; each was checked against the actual repo before adoption. All 4 held — but only the checking made them adoptable. The reviewer had no repo context guarantees; "plausible" and "true" diverge exactly there.