Reading a certificate

Every certificate derives from exactly one ruling (PMK-CRT-001): punchmark certify emits one line of text, and with --json or --out the same content as a certificate/v1 document. Here is a real line from the synthetic quickstart roundtrip:

punchmark certificate pmk-c-9f159a6c9f021abb: producer identity of route
synth/route-a (task synthtask, window
2026-01-01T00:00:00+00:00..2026-01-01T01:00:00+00:00) HOLDS at false-alarm rate
0.01; minimum resolvable substituted fraction 0.02 against candidate set
pmk-cs-169b400fc8599df1 (3 routes) over 72 items in 12 clusters. This certifies
the route label as served within the named candidate set; it is not a statement
about model weights.
[detector chargram v1; model pmk-m-3f963322e6324fe5; ruling pmk-r-3c53036ea627aa37]

Those ids are deterministic: the quickstart's commands reproduce them exactly, since every seed is derived rather than sampled. They are not pinned by any test, so a detector or generator change would move them without failing CI. Read the shape of the line here rather than the exact hashes.

Field by field:

field	meaning
pmk-c-...	The certificate id: a content hash of the certificate body. Identical inputs reproduce identical ids.
route ...	The route label being certified: the label as served. The weights behind it are outside the claim (PMK-CRT-002).
task ..., window ...	The claim's unit: one (route, window) per ruling. The window comes verbatim from the caller-written sidecar. A ruling scored under a declared task alias additionally prints scored as <model task> (PMK-RUL-005). That marker means the calibrated false-alarm rate was measured on different content than this archive's.
HOLDS at false-alarm rate 0.01	The verdict clause at the operating point you declared with --far. The other clauses are DOES NOT HOLD ... statistic fell below the calibrated threshold ... and IS UNDETERMINED (reasons).
minimum resolvable substituted fraction 0.02	rho*: the smallest substituted fraction the archive's k and item count could have resolved for the worst candidate pair. Only a HOLDS line carries it.
against candidate set pmk-cs-... (3 routes)	The closed set the verdict is relative to, by content id and size.
over 72 items in 12 clusters	What the verdict rests on. The cluster count is the one that bounds the effective sample size, since items inside a cluster share program content, so a row count on its own would overstate the evidence (PMK-CAL-001).
the scope sentence	Fixed wording, carried by every certificate (PMK-CRT-002): "This certifies the route label as served within the named candidate set; it is not a statement about model weights."
[detector ...; model ...; ruling ...]	Provenance: detector id and version, fitted-model id, and the ruling id this certificate derives from. This is enough to re-derive the line from the rulings store.

The tri-state exit code

punchmark certify exits 0 on HOLDS, 1 on DOES NOT HOLD, 2 on UNDETERMINED or unevaluable input. Both non-zero codes are red. An undetermined certificate can never read as success, and any pipeline that special-cases exit 2 back to green has defeated the instrument. UNDETERMINED carries its reasons in the line: too few items or clusters, a stub share over the cap, no operating point calibrated at the requested false-alarm rate, or a power limit.

What SAME-PRODUCER means exactly

SAME-PRODUCER means: a substitution of fraction >= rho_target by any candidate in the set would have been flagged with the calibrated power, and none was (PMK-RUL-001). It is a statement about what the archive could have detected and did not detect. It does not positively identify what served the route. When some candidate pair is unresolvable at your --rho-target, a passing statistic is forced to UNDETERMINED: absence of an alarm is not evidence there.

The closed-set caveat

Every verdict is relative to the candidate set declared at fit time (PMK-CRT-003). A producer outside that set was never considered. If you score a route that is not in the set, punchmark refuses instead of answering. If the true producer is not among your candidates, no punchmark verdict, including SAME-PRODUCER, says anything about it. Declare the candidates you actually mean, and read every certificate with its set.