cross-verify

Cross-Verify — Independent Dual-Voice Review

The "independent eyes" gate. Mirrors gstack's codex skill pattern: read the produced artifact in fresh context (no session pollution from the producing chain), critique adversarially. Mark every claim as Substantiated / Speculative / Hallucinated / Missing-Source.

Why Independent Context

The specialists, writer, and reviewer all share session context. Patterns of reasoning leak between them. Cross-verify deliberately reads ONLY the artifacts — ./rfp-parsed.md (the original RFP) and ./rfp-draft.md (the response) — so it can spot claims that look plausible to insiders but lack grounding to a fresh reader.

Workflow

1. Spawn isolated context. Use the Agent tool with subagent_type=general-purpose. Pass it ONLY:

   • ./rfp-parsed.md (full text)
   • ./rfp-draft.md (full text)
   • ./rfp-eval-criteria.md if present
   • ./eligibility-decision.md if present (so it knows what was decided)

   Do not pass session history, specialist drafts, or the review report. The point is independence.

2. The sub-agent's prompt (used by this skill when it dispatches):

   You are a critical reviewer of an RFP response. You have NOT seen the drafting process. You have ONLY the RFP and the response.

   Read both. For every claim in the response, classify:

   • Substantiated — directly grounded in the RFP, common knowledge, or self-evident.
   • Speculative — plausible but not provable from what you can see; would benefit from a citation.
   • Hallucinated — appears specific (numbers, dates, named past wins, technology versions) but you have no way to verify; high suspicion of fabrication.
   • Missing source — language like "industry-leading" or "best-in-class" with no substantiation.

   Also identify:

   • Unanswered RFP requirements — anything the RFP asks for that the response doesn't address.
   • Weasel words — language that softens commitments inappropriately ("we strive to", "we believe we can", "our goal is").
   • Tone mismatches — places where the response's voice clashes with the RFP's evaluator profile (technical evaluators want different language than executive committees).

   Output a structured critique with line-references where possible.

3. Optional: Codex dual-voice. If CONSULTADD_USE_CODEX=true is set and the codex CLI is on PATH:

   if [[ "$CONSULTADD_USE_CODEX" == "true" ]] && command -v codex >/dev/null; then
     # Pass the same artifacts to Codex with a similar adversarial prompt.
     # Capture Codex's output to ./codex-critique.md
   fi
   

   Run Codex in parallel with the Claude sub-agent. Two independent voices catch different failure modes.

4. Merge findings. Combine the Claude sub-agent's critique and (if used) Codex's into ./cross-verify-report.md:

   # Cross-Verify Report — <RFP Number>
   
   ## Substantiated Claims
   <count + sample>
   
   ## Speculative Claims
   - <list with line refs>
   
   ## Hallucination Suspicions
   - <list with line refs — HIGH severity, surface to Taste Gate>
   
   ## Missing-Source Language
   - <list — superlatives without substantiation>
   
   ## Unanswered RFP Requirements
   - <list with RFP section refs — HIGH severity>
   
   ## Weasel Words
   - <list>
   
   ## Tone Concerns
   - <list>
   
   ## Reviewers
   - Claude sub-agent (independent context): <ran | failed>
   - Codex CLI: <ran | not-available | disabled>
   

5. Severity routing.

   • Hallucination suspicions and unanswered requirements → Challenge Gate.
   • Speculative claims and missing-source superlatives → Taste Gate.
   • Weasel words and tone concerns → Mechanical (writer can fix on next pass) or Taste depending on context.

6. Emit telemetry.

   (Telemetry is out of scope for v1. When the telemetry MCP Connector is live in v1.5, this skill will call its emit tool here.)

Constraints

• Independence is the whole point. Sub-agent gets ONLY the two markdown files — not session history, not specialist drafts, not the review report. If you slip context in, you've just made a second internal reviewer.
• Don't auto-fix. This skill only critiques. Fixes go through writer → reviewer on a re-pass, or are surfaced at the Taste / Challenge gate.
• Codex is optional. If unavailable, run Claude-only and note in the report.

Output Files

• ./cross-verify-report.md — merged critique
• ./codex-critique.md — Codex output if used (not deleted; keeps audit trail)