reality-check

A SIGNAL is not an OUTCOME. Before you act on "X is down/dead/blocked/broken",
run the gate:

scripts/dev/reality-check.sh "<belief>" [--detector <kind>] [--halt-class]

What it does (the 5-step gate):

1. Names the belief + the signal that triggered it.
2. States the outcome the belief predicts (fleet stops shipping / trunk red).
3. Checks ground truth — recent merges (is the fleet actually shipping?) + trunk status.
4. Checks signal reliability — is there an OPEN gap marking this detector a false-positive?
5. Verdict — exit 1 REFUTED (stand down), 0 CONFIRMED (proceed), 2 UNVERIFIED.

Rules:

• Never broadcast or act on an alarm-class belief until reality-check returns CONFIRMED.
• A --halt-class belief (declare outage / stop the fleet / page the operator) additionally
  requires a fresh-eyes (or peer) confirm — a single session may not act on it alone.
• If the detector has an open false-positive gap, fix the detector, don't act on its output.

Full procedure + case study: docs/process/REALITY_CHECK.md.

Examples that should trigger this skill: "is the fleet really down?", "before I declare CI broken",
"reality-check the auth-dead alert", "should I stop the loop / page the operator", "verify this outage
is real".