vouch-verify

name	vouch-verify
description	Run the vouch adversary against a list of Claims and write verified verdicts. Use this standalone (without /vouch-run) when you already have a JSON file of claims and want survival scores attached. Invokes the vouch-adversary subagent under the hood.

You are the verification skill. Take a claims file in, write a verified-claims file out.

Arguments

/vouch-verify <claims_file> [--modes M1,M2,...] [--out OUT]

• claims_file — JSON. Either an array of claim objects, or a JSON object with a claims array.
• --modes — comma-separated subset of source_verification, contradiction_search, cross_reference, temporal_check. Default: all four.
• --out — output path. Default: <claims_file>.verified.json.

Step 1 — Validate the input

A claim object must have at least:

{
  "text": "string — the claim itself",
  "confidence": 0.0
}

Optional fields the verifier can read or set: source_url, quote, verdict, confidence_delta, adversary_evidence, verification_mode. The verdict enum is confirmed | refuted | weakened | unverifiable | pending.

If any item is missing text or confidence, refuse and list the offending indices.

Step 2 — Spawn the adversary

Use the Agent tool with subagent_type: "vouch-adversary". Pass it:

• the claims file path (read-only access)
• the list of modes
• the path to write the verified file

The adversary returns when done; the file at --out exists.

Step 3 — Compute survival metrics

Read the verified file. For each claim, compute survival_score per the rule:

verdict	survival_score
confirmed	confidence
weakened	max(0, confidence + confidence_delta)
refuted	0.0
unverifiable	confidence * 0.5
pending	confidence (provisional)

Then compute and report:

• survival_rate = (confirmed + weakened) / (confirmed + weakened + refuted) (None if denominator is 0; pending and unverifiable are EXCLUDED from the denominator.)
• trust_score = mean(survival_score)

Step 4 — Print the summary

claims         : N
confirmed      : N
weakened       : N (Δ avg: ...)
refuted        : N
unverifiable   : N
pending        : N

survival_rate  : 0.xx (or n/a)
trust_score    : 0.xx

written to     : <OUT>

Quote-on-refute rule (load-bearing)

The verifier MUST attach a direct quote (quote field) when setting verdict to refuted or weakened. If the subagent returns a refuted/weakened claim without a quote, this skill must downgrade that claim to unverifiable and append "[downgraded: no source quote]" to its adversary_evidence. This rule prevents the adversary from reaching beyond what it can cite.

Rules

• Never modify the input file in place.
• Never invent quotes. If the subagent didn't provide one, downgrade — don't backfill.
• A claim with verdict pending after verification means the adversary explicitly couldn't decide; report it but don't penalise it.
• --modes order matters: the first mode that yields confirmed or refuted short-circuits the rest. Other modes can only weaken or mark unverifiable.