Verify
Reach for Verify the moment anyone, including you, says something is done.
- A task came back with a summary and no evidence you can point at. Ask what the summary is standing on before you accept it.
- The message says "should", "probably" or "seems to". Each of those words means nobody ran the check.
- You are about to publish, send or merge on a model's word. Run the check first, so the claim carries a tier.
- Another habit just reported back. A search, a debate, a fix or a review each makes a claim, and each claim needs its grade.
Verify never skips. It ranks first because it audits the other four habits.
Done is the word that causes the most trouble in a working session, and most false claims of done are not lies. They are small claims that sound large. "The file is there" turns into "it works" somewhere between the work and the message, and nobody checks the step in between. Verify makes that step visible. Before anyone says a thing is finished, it asks one question: what tier did this reach, and what evidence proves it, in the same message?
The Four Tiers Of Done
Every claim sits on one of four tiers. Each one asks a harder question than the tier below it.
- Tier 1, Exists. Is it there at all, and not empty? Most done claims stop here.
- Tier 2, Substantive. Is it real, or a placeholder shaped like the real thing?
- Tier 3, Wired. Does anything reach it? Work that nothing calls fails here.
- Tier 4, Functional. Did it run end to end just now, with the evidence in the same message?
Reading the work can reach tier 3 at most. Only a run reaches Functional, so a file that looks correct can still fail once it runs beside the parts around it.
Why Claims Drift Upward
A builder who has the whole session in their head feels that a partial check is complete. The reader only has the sentence, and the sentence does not say which check was run. The fix is to make the sentence carry its own evidence, so the reader does not have to take the feeling on trust.
The same drift happens to a claim that is restated to fit the check. A claim rewritten to be easier to prove is a different claim, and the grade then describes the wrong thing. Verify grades the claim as it was made.
What Evidence Looks Like
Evidence has a shape. It is a test run with its output pasted, a row count from the live table, a screenshot of the rendered page, or an exit code with the exact line that changed. The words "should", "probably" and "seems to" mean nobody ran the check, and a celebration written before the check is the same signal.
| The claim | The tier it usually reaches | What is missing |
|---|---|---|
| "Checks pass" | 1 or 3 | The full output of a fresh run on this version |
| "Bug fixed" | 2 or 3 | The original symptom shown fixed, and the test that failed before the fix |
| "Work complete" | 2 or 3 | A line by line check against the request |
| "Helper finished" | 1 | The change record, since the report says only what the helper believes |
| "Panel agreed" | 1 | The filled matrix, with both columns on every row |
| "Reviewed" | 1 | The review file and a verdict on each finding |
Why Verify Audits The Other Four
Four other habits produce claims that drift. A search can be shallow and reported as thorough. A debate can end in "the panel agreed" with no matrix behind it. A fix can name a root cause with no test. A review can be skimmed and relayed as agreement. Each one is a tier 1 claim dressed as tier 4, and Verify is the habit that refuses the dressing.
A review from the same family of model that built the work is weak evidence too, because it carries the builder's assumptions. A delegated result is graded from its change record, never from the helper's report.
Hard Rules
The check runs on the version that will ship, with no debug flags switched on. The full output is read, including the skips and warnings under a green summary line. Counts are never rounded up, so three failures out of sixteen are written as "3 of 16 failed" with the three names listed. A check that cannot run is written as not run, with its reason.
When A Check Cannot Run
Say so in the same message as the claim. Grade what can be shown, name the exact command that would close the gap, and write the honest sentence in full: "Built, and not yet verified, because [reason]." A passing run on a sibling project or a copy proves that sibling or copy, and an earlier run proves only the version it ran against.
The One Question
Before a finished claim leaves the room, ask which tier was actually checked. If the honest answer is "I read it," the claim is tier 3 at most, and the message should say so.