[
  {
    "section": "Coverage",
    "items": [
      "Which languages and frameworks are supported, and is there a public case study on code like mine (ZK circuit, MPC library, TLS stack, Solidity)?",
      "Does the tool understand cryptographic specifications (field arithmetic, constraint systems, pairing checks), or only general code patterns?",
      "Can I supply a threat model or scope file, and does the vendor show it measurably reducing noise?"
    ]
  },
  {
    "section": "Evidence quality",
    "items": [
      "Do findings ship with a reproducible proof of concept or executable test, or only prose?",
      "What are precision and recall on a benchmark that post-dates the model's training cutoff, and can I see the full findings list rather than only the hits?",
      "Are severities calibrated by a named human, and does the report say who?"
    ]
  },
  {
    "section": "Validation and accountability",
    "items": [
      "Who validates findings before I see them, and what false-positive rate has been observed on real engagements?",
      "Is the report signed by a named reviewer (AI-assisted audit) or delivered as machine output (AI audit)?",
      "What is the disclosure path if the tool finds bugs in my dependencies?"
    ]
  },
  {
    "section": "Security of the pipeline",
    "items": [
      "Where does my code run, which models see it, and is it used for training?",
      "How is repository content isolated from agent instructions (prompt injection), and what credentials does the agent hold?",
      "Are results reproducible run to run, and are runs logged with model versions and dates?"
    ]
  },
  {
    "section": "Economics",
    "items": [
      "How are re-scans triggered, priced and deduplicated as models change?",
      "Does pricing reward validated findings or raw volume?",
      "What human triage time should I budget per scan, based on the vendor's own case studies?"
    ]
  },
  {
    "section": "Red flags",
    "items": [
      "Accuracy claims such as '100 percent' or '92 percent detection' with no dataset or methodology.",
      "Findings without file and line references, or severities passed through from the model unchanged.",
      "Benchmarks built only from public audit reports the model may have trained on."
    ]
  }
]