verifylisted

# Verify An agent earns trust in exact proportion to how mechanically its output can be checked. This skill is that rule pointed at the finished work. Read `${CLAUDE_PLUGIN_ROOT}/SKILL.md`, then `${CLAUDE_PLUGIN_ROOT}/references/phase-verification.md` and `${CLAUDE_PLUGIN_ROOT}/references/laws-hard-gates.md` (L7 to L10). ## Run verify once, through the command that mints its own evidence ``` "${CLAUDE_PLUGIN_ROOT}/bin/sbe" verify <dir> ``` This single command already runs the design completeness check (`sbe_design.py --strict`), the four hard gates together (`sbe_gate.py`: numbers, migration, approval, ran), and the scored surface (`sbe_score.py --strict`), in that order, and prints every verdict line each one produces. These four gates plus the silent-failure lints are refused rather than waived: an operator instruction in session can override a default, never a hard gate. It then mints one evidence receipt per delegate (design, gate, score) into `.sbe/evidence`, the same store `sbe status` reads (CR-08, `design/lifecycle-blockers/03-adr.md`), so a clean run leaves proof behind instead of a PASS `sbe status` cannot see. A receipt minted against a dirty tree still reads NO-DATA, naming the dirty state: that is correct, not a bug, the first time it is surprising. For the stricter soft-finding surface `bin/sbe verify` does not itself request, also run: ``` "${CLAUDE_PLUGIN_ROOT}/bin/sbe" score --strict --strict-soft <dir> ``` ## How to read a verdict - **PASS** means