fanout-verify-synth

A deterministic go/no-go gate for synthesis after a fanned-out verification. In a deep-research / multi-agent fan-out, several agents each emit claims, and each claim is verified adversarially. Before you synthesize the final artifact you must answer one question — is every claim actually verified? — and the reliable way to answer it is a machine gate, not eyeballing a list. This script reads the collected verify.json and returns the answer as an exit code.

It is not the part that runs the fan-out (spawning agents, executing verification). Like scripts/goal-loop.sh, it is the stop / verify primitive: it only decides whether to STOP and synthesize. "Done" is decided by the gate's exit code, never by the model reporting "everything checked out".

The decision forbids self-report: a bare verified label is not enough. A claim is counted as effectively verified only if it also carries at least --min-sources pieces of evidence — a label is self-report; the evidence count is data.

When to use / not

• ✅ A fan-out produced claims, each adversarially verified, and you need a deterministic go/no-go before synthesizing the report.
• ✅ You want "is everything backed up?" settled by a gate, not by reading down a list.
• ❌ There is no per-claim status/evidence to gate on, or a single unverified assertion you can just check by hand.

verify.json schema

{
  "topic": "the research question",
  "claims": [
    { "id": "c1", "claim": "one-line summary", "status": "verified",
      "sources": ["https://...", "https://..."], "agent": "agent-a" },
    { "id": "c2", "claim": "...", "status": "refuted", "sources": ["..."] }
  ]
}

• status: one of verified | refuted | unverified | pending.
• sources: evidence URLs for the verification. A verified claim with fewer than --min-sources (default 1) sources is demoted to held — a label without evidence is treated as self-report, not as verified.

How to run

Aggregate each verification agent's result into a single verify.json (record status and sources for every claim), then judge:

bash "${CLAUDE_PLUGIN_ROOT:-.}/scripts/fanout-verify-synth.sh" --in verify.json [--min-sources 1] [--json]

Branch on the exit code:

Use --json to get a machine-readable verdict (verdict, exit, counts, plus held_ids and refuted_ids) so you can open a re-verification task for exactly the claims that failed.

The evidence rule

The verified label alone never adopts a claim. --min-sources N (default 1) sets how many evidence URLs a verified claim must carry; below that it is demoted to held and the whole set falls to HOLD. Setting --min-sources 0 relaxes the rule (a bare label counts) — useful only when evidence lives elsewhere. This is the same "a label is self-report, evidence is data" principle the sibling gates enforce.

Deterministic --check (self-test)

bash "${CLAUDE_PLUGIN_ROOT:-.}/scripts/fanout-verify-synth.sh" --self-test

• exit 0 = PASS. It exercises the real decide() path on hermetic mktemp fixtures (no hardcoded success) across 9 cases: all-verified→ADOPT, unverified-mixed→HOLD, refuted→CONFLICT, verified-but-under-sourced→HOLD, --min-sources 0 relaxes the same input→ADOPT, empty-claims→HOLD (fail-safe), broken-JSON→error, claims-not-array→error, and missing-file→error.
• non-zero: a gap in the verdict truth table or syntax.

Sibling gates

Same stop-primitive family — all deterministic, all "a label is self-report, evidence is data", all delegating the terminate decision to an exit code:

• adversarial-refute-vote — puts a single claim to N independent refute votes and stops adoption when a majority refutes it.
• judge-rubric-vote — disciplines multiple judges' verdicts by majority + vote spread and calls a hung decision when agreement is below threshold.
• scripts/goal-loop.sh — the underlying stop primitive (exit 0=GOAL MET / 1=CONTINUE / 2=CAP REACHED) that this gate's exit-code contract mirrors.