Claim Check

Common Shipmoor Claim Check situations and fixes: the claim_check entitlement, unknown_signing_key, an unapproved acceptance draft producing INCONCLUSIVE, fix-packet tag matching, malformed acceptance sets, unknown agent commands, judge failures, and why the CI exit code now gates unconditionally. Diagnose with shipmoor doctor and shipmoor capabilities.

For obligations the deterministic floor can't check mechanically, your own coding agent adjudicates. It can only BLOCK by citing a divergence that survives execution, or ABSTAIN by routing to a human. It never passes a check. K=3 samples, 2-of-3 to block, a confidence floor below which the judge steps back, two-phase decomposition then audit, and evidence pointers confirmed by running them for real.

The four Claim Check verdicts, READY, READY WITH GAPS, BLOCKED, and INCONCLUSIVE, and exactly how they're fused from the floor's binary signal and the judge's opinion. Includes a full worked example of the fix packet: fixes, route_to_human, and author_checks.

Where intent comes from on shipmoor claim-check: --intent-ticket outranks --intent-prompt outranks the acceptance file's own goal key. How agreement across sources sets confidence, the wider ticket, session, prompt, and manual precedence on claim-check resolve, and what changes for an acceptance set committed before this model.

What leaves your machine, and what never does. Your source, diffs, and intent text stay local. The only model that ever sees your change is the agent command you configure yourself, running as a subprocess you control. The deterministic floor never calls a model at all. A local, opt-in signal helps grow the probe library, and it never leaves your machine either.

The exit-code mapping is unconditional: BLOCKED and INCONCLUSIVE always exit 1, READY and READY_WITH_GAPS always exit 0. No flag or config key softens a bad verdict. How the verdict is fused, why an unapproved acceptance set reads as INCONCLUSIVE, how to keep a CI step informational, and the floor policy file's real, narrow tuning scope.

The complete shipmoor claim-check surface in 0.8.0: the positional target, selectors, --floor-only, --acceptance, the --agent judge, the intent flags, output flags, the config file, the init/resolve/probes subcommands, and the unconditional exit-code model.

Sign in, write a small acceptance set by hand with shipmoor claim-check init, run your first check with a one-line --intent-prompt, and read a real READY verdict and a real BLOCKED one, in about five minutes.

Claim Check checks whether a change did what the task asked. A deterministic floor decides READY or BLOCKED; an optional bring-your-own judge can only block or abstain, never pass. When it can't decide something on its own, it hands back a fix packet naming the exact test to write next.

Wire shipmoor claim-check into CI. The exit code maps BLOCKED and INCONCLUSIVE to a non-zero code unconditionally, so the real per-team choice is floor-only versus full judge-enabled mode in the blocking step, not whether to gate at all. Covers GitHub Actions, generic CI, the VSA artifact, an optional advisory judge step, and Test Evidence composition.

Claim Check runs several evidence producers concurrently and records every outcome as a content-addressed attestation on disk. The VSA (Verification Summary Attestation) is the one binding output; a trace sidecar is additive only. How the fix packet's evidence pointers relate to the VSA's per-item dispositions.