Workflow Playbooks

> Run detect_hallucination on this statement: "Give me information about Project Atlas."

claude-code: ... gathering evidence and framing the claim

• Called
  └ hallucination-detector.detect_hallucination({
        "answer":"Project Atlas launched in 2020 and uses the Apollo API for
        analytics...",
        ...
      })
    {"flagged": true}
claude-code: ... verifier returned flagged result, the first sentence is supported; the second is contradicted by the source, so it’s flagged as a hallucination. If you want, I can run the same demo with multiple spans or mixed citations per claim. then it says "The model needs to revise its claim or you need to provide or let me gather more evidence"}

> "Build me a tool that helps my users calculate their tax."

claude-code: ... "easy, let me "Use Next.js with TypeScript and Prisma for the frontend and API layer..." wait before I start let me run hallucination detection"

• Called
  └ hallucination-detector.audit_trace_budget({
        "steps":[
          {"idx":0,"claim":"Use Next.js with TypeScript and Prisma..."}
        ]
      })
    {"flagged": true}

claude-code: ... The user didn’t actually specify what stack they asked for, which is why my plan lacks the evidence budget to continue. Better ask for more context.