mtarcure/claude-vibe-squad/rule8-truth-gate

Rule-8 Truth Gate

Combines the release-gate portions of claim-verification + citation-audit into one invokable check. Given a deliverable's claims plus a returned citation / grounding-result set, it decides whether each load-bearing claim is supported by a returned citation within the requested recency window, and emits a machine-checkable gate record. A model cutoff is never verification evidence; grounding is a first-class stage, not a hope that the reviewer supplies evidence later.

Procedure

1. Decompose the content into discrete claims and mark each load-bearing (a conclusion the deliverable

depends on) or non-load-bearing / backgrounded. Classify each: fact | quote | calculation | forecast | opinion | inference. Every load-bearing claim is gated — INCLUDING a load-bearing inference (a reasoned conclusion the deliverable rests on): an inference is NOT exempt because it is framed as reasoning rather than quoted — its supporting premises/evidence must map to returned citations, or it is unverifiable → needs_tool. Only a genuine opinion, or non-load-bearing / backgrounded reasoning, passes ungated.

1. Ground each load-bearing web/factual claim: route a grounding worker (Gemini Google Search grounding

or a research-namespace handoff) that returns a typed evidence bundle of citations/search results. If no grounding tool is callable for a load-bearing claim, that is a capability gap → set needs_tool: true (a separate axis). A claim whose callable grounding tool returned nothing supporting is unresolved evidence, not a missing tool → it drives HOLD/FAIL, not needs_tool.

1. Map claim → citation (claim_to_citation): each load-bearing claim — AND each load-bearing inference, via

the evidence for its premises — must map to a specific returned citation, not to model prose. Read the source and confirm it SUPPORTS the claim/premise (not merely mentions the topic); for an inference, also confirm the reasoning is valid over the mapped premises. Preserve quote context.

1. Resolve + date-check every citation (date_window): resolve each citation, record accessed-at; a

citation that 404s or was invented is unverifiable. Confirm publication/event dates fall inside the requested date_window, and check for retractions/corrections. A claim supported only by a citation outside the window is not fresh-supported.

1. Reproduce arithmetic; check units and internal consistency. Label vendor/provider performance claims as

vendor claims unless reproduced on a Vibe-Squad benchmark.

1. Reject unsupported (reject_unsupported): any load-bearing claim OR load-bearing inference that does not

map to an in-window returned citation is dropped or the gate fails — never PASS on prose or on unbacked reasoning. When grounding was callable, an unmapped or out-of-window load-bearing item drives the result to FAIL (or HOLD pending more evidence), never needs_tool — needs_tool is reserved for the capability-absent case.

1. Emit the gate record (see below).

Machine gate record (emit this)

rule8_truth_gate:
  result: PASS | HOLD | FAIL         # PASS only if `unverifiable` is empty AND `needs_tool` is false; FAIL if a load-bearing claim/inference is unsupported when grounding WAS callable; HOLD if resolvable with more evidence
  needs_tool: <bool>                 # true iff a load-bearing claim/inference had NO callable grounding tool (capability absent — a SEPARATE axis from evidence failure)
  claim_to_citation: true            # policy applied: every load-bearing claim AND load-bearing inference (via its premises) mapped to a returned citation
  date_window: <explicit interval>   # e.g. 24h / 7d / 90d / task-scoped — never true/false
  reject_unsupported: true           # unsupported/out-of-window load-bearing items drove HOLD/FAIL, never PASS
  claims_checked: <n>
  load_bearing_inferences_checked: <n>
  citations_resolved: <n> / <n>
  unverifiable: [<load-bearing claim/inference ids>]   # items with no in-window citation despite a callable tool → drive HOLD/FAIL; MUST be empty for a PASS

When to invoke

• The S4 Verify (Rule-8) step of project/investigation-synthesis, project/editorial-longform,

project/marketing-campaign, project/audio-assets, project/video.

• Any deliverable asserting factual / product / efficacy claims before publish.

When NOT to invoke

• Pure opinion / creative pieces with no load-bearing factual claims or load-bearing inferences.
• As a substitute for the independent cross-family reviewer — this is the primary's self-gate, not the review.

Acceptance

• Every load-bearing claim AND every load-bearing inference is classified and mapped to a returned in-window

citation (an inference via the evidence for its premises), or the result is HOLD/FAIL (when grounding was callable) or needs_tool: true (when no grounding tool was callable) — the emitted result is PASS only when unverifiable is empty AND needs_tool is false. A load-bearing inference never PASSes on reasoning alone. "Unverifiable" is distinguished from "false"; a model cutoff is never verification.

• Vendor claims are labeled; no corroboration is fabricated. Quotes keep their context.

Notes — source boundary + cross-lane home

• Partial fusion: this gate owns the aggregate S4 release decision and citation-recency policy from

the claim-verification skill + the retired citation-audit skill. It does not emit the exact evidence span and supported | unsupported | unverifiable verdict for every individual statement; invoke claim-verification when that statement-level adjudication is required.

• Canonical home: .claude/skills/rule8-truth-gate/SKILL.md. Because its audience is specialist,

model-lanes/SKILL-HOMES.md requires .agents/skills/rule8-truth-gate/SKILL.md to be a byte-identical regular file mirror.