verification before completion

five-step gate before any status claim:

1. Identify — what command proves this claim?
2. Run — the full command, fresh, now.
3. Read — the complete output, the exit code, the failure count.
4. Verify — does the output actually confirm the claim?
5. Only then claim, with the evidence attached.

each claim type has its proof: "tests pass" needs the test command's zero-failure output; "build succeeds" needs exit 0; "bug fixed" needs the original symptom re-tested; "agent completed" needs the VCS diff, never the agent's self-report. regression tests carry their own bar — red-green verified: write, pass, revert the fix and watch the test fail, restore, pass again.

satisfaction is a feeling; verification is a fact. the failure mode is claiming on momentum — the work feels done, the words "great, perfect, done" arrive before the command does. red flags: exactly those words, plus the hedges ("should," "probably," "seems to") that smuggle in unverified claims.

• verification must be fresh — a green run from earlier in the session proves only the tree it ran on.
• partial verification proves nothing.
• agent success reports are inputs to verify, not evidence.

before every success claim, commit, task completion, or "moving on" — any moment something positive is about to be said about work state.

bad

statement: the route is fixed because typecheck passed earlier

the prior check covered a previous tree, not the live behavior claimed. a later edit, merge, or stale checkout silently invalidates the remembered run, and the claim asserts something nobody observed. the gap surfaces downstream — review, deployment, someone else's debugging — where it costs far more than one fresh command would have.

good

claim: the route returns a typed active status
fresh evidence: current deployment probe returned 200 with that payload

fresh output tied to the exact assertion: every completion statement is reproducible by anyone, and hedges lose their place. held as a habit, "done" stops depending on who is speaking.

verification without tests — the local resolution · what counts as verification evidence · debugging before fixing