A parallel audit: research, compare, check, find gaps

---
name: parallel-system-audit
description: Splits a goal into 4 parallel workers, verifies with fresh context, and synthesizes.
spec: skill-graph/1
topology:
  nodes:
    - id: goal
      type: input
      label: 1. GOAL
    - id: splitter
      type: orchestrator
      label: 2. SPLIT
      model: sonnet
      tools: [Read]
    - id: worker_research
      type: worker
      label: "3. WORKER 1: Research"
      model: haiku
      tools: [Read, Grep, Glob, WebSearch, WebFetch]
    - id: worker_compare
      type: worker
      label: "3. WORKER 2: Compare"
      model: haiku
      tools: [Read, Grep, Glob, WebSearch]
    - id: worker_check
      type: worker
      label: "3. WORKER 3: Check"
      model: haiku
      tools: [Read, Grep, Glob, Bash]
    - id: worker_gaps
      type: worker
      label: "3. WORKER 4: Find Gaps"
      model: haiku
      tools: [Read, Grep, Glob]
    - id: verifier
      type: verifier
      label: 4. VERIFIER (Fresh Context)
      isolated_context: true
      model: opus
      tools: [Read, Grep, Glob]
      gate: Every kept claim cites concrete evidence and no two workers contradict each other.
      on_fail: retry
      max_retries: 1
    - id: merger
      type: synthesizer
      label: 5 & 6. MERGE & FINAL ANSWER
      model: sonnet
      tools: [Read]
  edges:
    - from: goal
      to: splitter
    - from: splitter
      to:
        - worker_research
        - worker_compare
        - worker_check
        - worker_gaps
      contract: subtask_json
    - from:
        - worker_research
        - worker_compare
        - worker_check
        - worker_gaps
      to: verifier
      contract: findings_array
    - from: verifier
      to: merger
      contract: verified_payload
contracts:
  subtask_json:
    format: json
    description: One key per worker id; each value is that worker's task in one or two sentences.
  findings_array:
    format: markdown
    description: A numbered list of findings, each with its evidence (file, line, quote or URL) and a confidence of high, medium or low.
  verified_payload:
    format: markdown
    description: Only the claims that passed verification, then a short list of rejected claims with the reason.
---

# Node Prompts & Edge Contracts

## @node: splitter
You are the Orchestrator. Take the user's goal and break it into 4 distinct tasks.

The goal:
{{goal}}

Output strictly as JSON with keys: `worker_research`, `worker_compare`, `worker_check`, `worker_gaps`.

## @node: worker_research
Focus ONLY on deep research for: {{splitter.worker_research}}
Return findings with exact citations.

## @node: worker_compare
Focus ONLY on comparing alternatives and trade-offs for: {{splitter.worker_compare}}

## @node: worker_check
Focus ONLY on checking technical constraints and correctness for: {{splitter.worker_check}}

## @node: worker_gaps
Focus ONLY on finding missing edge cases and blind spots for: {{splitter.worker_gaps}}

## @node: verifier
**CRITICAL: FRESH CONTEXT MODE.**
You only see the raw outputs from the 4 workers. You do NOT see their reasoning chain.
Audit every claim below for hallucinations, bias, or weak evidence. Reject unverified claims.

Original goal: {{goal}}

Worker outputs:
{{workers.output}}

## @node: merger
Combine the verified findings below into one clear, actionable report.

{{verifier.output}}