Claim Check

Claim Check answers the question a linter cannot — did this change do what the task asked? Deterministic probes decide; an LLM only advises and never blocks.

Opt in to blocking on a deterministic failure with --verdict-policy, preview with --would-block, and understand the gate floor that cannot be loosened: it blocks only on a deterministic, openly disclosed gap at non-low confidence — never on a guess or an LLM opinion.

Read the badge as three independent axes — maturity, coverage, and confidence — understand 'Not yet checked' as honest silence (not a pass), and see how plan drift is surfaced. Includes the JSON/SARIF shape.

Where the change's intent comes from and how confidence is set: --intent, --prompt, --session <transcript>, and .shipmoor/intent. Two agreeing sources raise confidence to medium, which matters for the gate.

What leaves your machine and what never does: no source, diffs, file contents, repo paths, or license tokens are sent to Shipmoor. The BYO-Judge runs under your own provider relationship.

Five minutes to your first claim check: scan a changeset with --intent, read the verdict, and watch an aligned change pass and a drifting one get flagged.

Opt in to an advisory LLM opinion on the long tail Shipmoor has no probe for — run by YOUR own coding agent, never a Shipmoor-hosted model and never uploading source. Labeled llm_inferred, excluded from the score, and structurally unable to gate.