Investigation Review JSON
Core Workflow
- Read the user prompt,
input/payloads/answer_template.json, and any task-local context payloads before querying the hub. - Treat the running Investigation Review Hub and task-local payloads as the only business evidence. Do not inspect environment source files, database files, generated manifests, answer files, or evaluator files.
- Identify the matter ID, hub base URL, required top-level keys, enum choices, field types, and ordering rules from the prompt and template.
- Fetch matter-scoped hub evidence from schema, matters, categories, productions, custodian sources, documents, privilege log, QC findings, retention events, and remediation actions. Prefer the bundled snapshot helper:
python3 skill/scripts/hub_snapshot.py --base-url "$TASK_ENV_BASE_URL" --matter-id "$MATTER_ID" --api-key review-key-017 > /tmp/hub_snapshot.json
If the task gives a different API key or says no key is needed, pass the task value or omit --api-key.
- Build the answer directly from the template. Include every required key, use only allowed enum values, fill non-applicable counts with
0, fill unavailable strings withnullwhen the schema allows it, and return exactly one JSON object with no prose.
Evidence Review
Read review_patterns.md when mapping hub records into legal review findings, category rollups, metrics, or actions.
When reviewing the snapshot, inspect these record families:
retention_events: post-hold losses, pre-hold policy destructions, missing required records, active system losses, auto-purge windows, available archives, record volumes, hold dates, and affected categories.custodian_sources: destroyed, lost, partial, not-collected, personal, board, archive, shared-drive, laptop, email, messaging, Teams, and offsite sources.review_documentsandqc_findings: responsive documents miscoded nonresponsive, zero-claim contradictions, privileged documents miscoded nonprivileged, missing required records, and production coding defects.privilege_entries: withheld counts, logged counts, unlogged deltas, third-party waiver indicators, over-designation, privilege miscoding, and category impacts.production_stats: category-level production state, zero-claim reasons, responsive counts, withheld counts, and nonresponsive counts.remediation_actions: stable action IDs, owners, priorities, target references, due days, and action types. Use these when they exist and match the template.
Answer Assembly
- Anchor each material issue to the most specific stable hub ID available, such as a retention event, source, privilege entry, QC finding, or document.
- Keep supporting reference lists sorted ascending unless the template says otherwise.
- Include only material non-ready categories or categories with open gaps unless the template explicitly asks for complete/no-gap categories.
- Aggregate multiple blockers for the same category into one category status when the schema has a mixed or multiple-blocker enum.
- Calculate metrics from the same selected issue records used in the answer body. Counts must reconcile with the listed risks, corrections, sources, or categories.
- Prioritize critical post-hold source loss and government disclosure before collection, recoding, privilege-log supplementation, waiver assessment, archive search, and monitor/no-action items, unless hub action ranks say otherwise.
- Validate the final object against the template manually before answering: required keys present, no extra enum labels, all required item fields present, numeric fields are integers, booleans are booleans, arrays sorted as directed.