The Checker Gate

# The Checker Gate

Something already told you it was done. It was lying. You shipped it.

Not a bug. A *claim*: a summary that said "tests pass, deployed, fixed" and that nobody opened.
The CHECKER kit is the fix: a second agent, with different instructions from the one that did the
work, that never writes anything, only checks, and refuses to take the first agent's word for
it.

## The four questions

Run all four on every "done" claim. Mechanical, not vibes. The first line of the verdict is always
PASS, FAIL, or PASS-WITH-FLAG.

1. Schema + range. Do the numbers, units and formats in the claim match what the file actually
holds? Unit drift (thousands vs millions, currency, seconds vs minutes) is an instant fail.

2. Confidence + escape hatch. Is there a hedge on a load-bearing claim (*should*, *probably*,
*seems*, *I think*), or a claim that traces to nothing? That's low confidence, and low confidence
forks. On outbound, money, credential or destructive work it's a FAIL ("unverifiable"). On
reversible, self-scoped work it's PASS-WITH-FLAG: it passes, and the unverified claim gets named in
one line. This fork matters. A gate that fails everything gets switched off.

3. Budget / truncation. Did the run stop mid-work (a claim trailing off, a file half-written)
while the summary kept going anyway?

4. Source-grounding. Does every factual claim trace to a file or a command output the checker
opened *itself*? Never to the worker's own prose. A prose-only claim is a fail.

There's one extra rule that isn't a question: a worker re-claiming an identical fix in identical
wording after a prior FAIL is a hallucination loop. Fail it and say so explicitly.

1. Unzip. You get verifier.md (the agent definition), CHECKLIST.md (the four questions,
printable), example-check.md (one real claim, worked through all three verdicts), and a README.
2. Copy verifier.md to ~/.claude/agents/verifier.md.
3. Start a new session — the verifier agent is now available.
4. If you'll use its OUTBOUND mode, fill in the six rule placeholders inside the file with your own
real rules first. An unfilled placeholder means "no rule configured yet," not "anything passes."

The system prompt inside verifier.md works with any tool that lets you define a named agent with
its own instructions. It is not Claude-Code-specific.

If the checker is a weaker model than the maker, its pass is only provisional. The kit says so on
the same page as the four questions, because a weaker model reading a stronger model's
work will accept reasoning it cannot follow. Half of the OUTBOUND mode also does nothing until you
write your own rules into those six placeholders.

And the gate can only check what it is handed. It never fires on its own.

Comment CHECKER on the reel and I'll send it over — or just grab it at
sharpendaily.co/kit/checker. No email, no DM, nothing
expires.