Reading the verdict

A claim-check result is deliberately not one score. It’s three independent axes — how well-checked the claim is, how much was checkable, and how sure Shipmoor is about the intent itself — because a high-drift result built on weak evidence and a vague intent must never read like a confident failure.

Anatomy of the block

Intent: Add a Stripe webhook handler for failed payments (payment_intent.payment_failed)
  Source: manual:--intent (manual_string) · agreement: single source · Confidence: low

Claim check  VERIFIED  ·  maturity: verified  ·  coverage: 100%
  ✓ A handler is bound to the Stripe payment_intent.payment_failed event.
  Not yet checked:
    ∅ Failure-path handling
    ∅ Webhook signature verification

Library 0.1.0 · Policy 0.1.0

• Intent: — the resolved goal text, masked (a secret pasted into the intent never survives here).
• Source: — which inputs resolved it, whether they agreed, and the intent confidence. See Providing intent.
• The badge — the maturity state as the loud headline word, plus coverage.
• ✓ / ✗ lines — one per probe that applied: ✓ satisfied, ✗ a disclosed gap.
• Not yet checked (∅) — expectations Shipmoor recognizes for this intent but has no probe for yet. Honest silence, not a pass.
• Library / Policy — the probe-library and policy versions that produced the result, part of the reproducibility fingerprint.

The three axes

Maturity: the five states

The weak states are styled to look weak. partial is not “wrong” — it means some expectations were checked and some were not; read the per-expectation lines to see which. Only gap_disclosed can ever earn a block, and blocking is a separate opt-in feature — see Turning on the gate.

Two kinds of “not checked”

• Not yet checked (∅) — expectations with no shipped probe yet. These do not lower coverage; there was nothing to run.
• The ⚠ footer — probes that did apply but returned cannot_check (an unsupported language, say). These do lower coverage, and the footer aggregates the count and reasons:

Claim check  NOT CHECKED  ·  maturity: unprobed  ·  coverage: 0%
  ? A Kubernetes Deployment is present in the change. — no relevant files in this change
⚠ We could not check 3 of 3 expectations (no relevant files in this change). Coverage 0%.

So coverage: 100% next to two Not yet checked lines reads: everything I probe, I could check — and here’s what I don’t probe yet.

Useful flags

• --explain — expand every expectation with per-probe detail: which fact matched, why a check was cannot_check, the judge’s rationale if one ran.
• --quiet-intent — collapse the claim check to a single badge line, for busy CI logs.

Plan drift (from a session)

When you pass --session <transcript>, Shipmoor also compares the agent’s own plan against the diff it produced. That’s a separate question from the claim check: not “did the diff do what the developer asked,” but “did the agent do what it said it would.”

Three conservative probes report it: plan.drift.goal_substitution (the plan and the task share no concept), plan.drift.scope_creep (the diff implements the plan plus unrelated files), and plan.drift.partial_implementation (a planned step only partly realized). Each errs toward silence — a false plan-drift is reviewer noise.

Plan-drift findings land in the normal findings list with category: intent_integrity at severity info, and they never change the exit code — not through the structural gate, not through the claim-check gate. The agent’s plan is never the standard of judgment; the resolved intent is.

In JSON and SARIF

--json carries the claim check as change_results[] — additive, absent entirely on a no-intent scan:

{
  "verdict": "major_gap",
  "maturity": "gap_disclosed",
  "coverage": 1.0,
  "gate_decision": "not_evaluated",
  "resolved_intent": { "goal_text": "…", "confidence": "medium" },
  "evidence": [ { "result": "unsatisfied", "basis": "deterministic" } ],
  "per_probe_summary": { "satisfied": 3, "unsatisfied": 0, "cannot_check": 0, "unmatched": 1 },
  "fingerprint": "sha256:…"
}

gate_decision is not_evaluated (advisory), passed, would_block, or blocked. unmatched counts probes that were considered but didn’t apply to this change — not an error. --sarif emits SARIF 2.1.0; plan-drift findings appear in the regular findings[] with category: intent_integrity.

Next

• Turning on the gate — when a gap_disclosed verdict should block.
• BYO-Judge — where inferred results come from.
• Providing intent — raising confidence with agreeing sources.