Update to Rohas Legal AI: Litigation

{
  "description": "Designs defensible document-review protocols for litigation, arbitration, investigations, or regulatory productions. Use when defining review population, responsiveness and issue codes, privilege and confidentiality treatment, document families, technology-assisted review, reviewer instructions, sampling, quality control, escalation, or production readiness.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 307
    }
  ],
  "name": "document-review-protocol-builder",
  "skill_md_contents": "---\nname: document-review-protocol-builder\ndescription: >-\n  Designs defensible document-review protocols for litigation, arbitration,\n  investigations, or regulatory productions. Use when defining review population,\n  responsiveness and issue codes, privilege and confidentiality treatment,\n  document families, technology-assisted review, reviewer instructions, sampling,\n  quality control, escalation, or production readiness.\n---\n\n# Document Review Protocol Builder\n\nCreate a repeatable protocol that different reviewers can apply consistently and\nthat preserves a review decision's source, reason, and quality-control history.\n\n## Intake\n\nObtain the mandate, jurisdiction and procedural orders, pleadings and issues,\ncollection map, review population and processing report, requested categories,\nsearch methodology, technology platform, document languages, confidentiality\nregime, privilege law and client structure, production specifications, team roles,\ndeadline, budget, and known high-risk custodians or subjects.\n\n## Method\n\n1. Define the review universe and exclusions. Reconcile collected, processed,\n   deduplicated, promoted, excluded, corrupted, encrypted, and unreviewable items.\n2. Translate the issues and requests into concise responsiveness, issue,\n   confidentiality, personal-data, hot-document, and technical-problem codes.\n   Give inclusion, exclusion, and boundary examples without inventing case facts.\n3. Define family treatment for emails and attachments, duplicates, near-duplicates,\n   threads, loose files, embedded objects, containers, versions, translations, and\n   structured data.\n4. State the applicable privilege categories and required facts. Create separate\n   paths for withheld documents, redactions, potentially privileged material,\n   privilege exceptions, common-interest or joint-client issues, and inadvertent\n   production. Do not infer privilege from lawyer involvement alone.\n5. Define escalation triggers for unclear scope, novel issues, personal or secret\n   data, technical failure, potential crime-fraud or equivalent exceptions,\n   inconsistent family coding, and material adverse documents.\n6. Specify reviewer training, calibration, decision notes, coding permissions,\n   batching, re-review, audit trail, productivity reporting, and conflict controls.\n7. Build quality control using reasoned samples: random and targeted checks,\n   confidence or error reporting where supported, senior review, disagreement\n   resolution, corrective action, and re-sampling. Do not claim statistical\n   assurance without a valid design and complete figures.\n8. If analytics or technology-assisted review is used, document objectives,\n   inputs, validation, sampling, stopping criteria, limitations, human oversight,\n   version changes, and reproducibility. Do not describe opaque scores as truth.\n9. Define production readiness: responsiveness, family completeness, privilege,\n   redaction, confidentiality, metadata, numbering, format, exception handling,\n   and final sign-off.\n\n## Output\n\nProvide the protocol, coding dictionary, decision tree, issue-and-request map,\nprivilege and redaction rules, reviewer escalation matrix, training examples,\nquality-control plan, exception log, production gate, and change-control record.\n\n## Guardrails\n\nDo not instruct reviewers to suppress adverse or inconvenient material, encode\nlegal conclusions unsupported by counsel, disclose privileged review notes, or\npermit automated tools to make unreviewed dispositive decisions. Protect reviewer\naccess, client confidentiality, personal data, source material, and audit logs.\n"
}