Investigation Review Hub Gap Analysis

Investigation Review Hub — gap & remediation analysis

You are outside-counsel legal-operations support. Each task gives you one matter (MTR-*) and asks for a single JSON object that inventories the material production/retention/privilege/QC gaps, maps them to the subpoena request categories, reports numeric metrics, and lays out a prioritized remediation plan — in the exact shape of that task's answer_template.json.

The hub is deliberately seeded with distractor ("noise") records that look like issues but are not. The entire difficulty of this family is separating the small set of material exception records from the noise, then translating them into the template's controlled vocabulary. Get that right and the numbers follow.

Inputs to read first (every run)

1. input/prompt.txt — the matter, agency, the specific review being asked for, and which dimensions to emphasize (retention, privilege, personal sources, deleted channels, coding quality, etc.).
2. input/payloads/answer_template.json — the output contract. Read it in full: required_top_level_keys, every enums list, ordering_rules, numeric_precision, and each field's per-item required keys. Your answer must match this template exactly; templates differ between tasks.
3. Any other input/payloads/*.json (e.g. request_context.json, review_scope.json, matter_context.json) — client-facing context and category labels only. Treat category titles here as labels; the evidence (events, sources, counts, statuses) must come from the hub, not the payload.
4. environment_access.md (repo root) — the hub base URL, the required credential header, and the allowed endpoint list for this run.

Source-of-record rule (stated in every task): use only the running hub endpoints for business evidence. Do not read local env files, database files, seeds, generated manifests, hidden notes, or any answer/evaluation file. Payload files are allowed only for the client-facing context above.

Network access

Read the base URL and credential from environment_access.md (and/or the task payload) at runtime — do not hardcode values that may change between runs. Send the credential header on every request (e.g. X-API-Key: <value>). Full endpoint/table map and query patterns: reference/hub_api.md.

Minimum data pull, always filtered by ?matter_id=<MTR-...>: /api/matters, /api/subpoena-categories, /api/productions, /api/custodian-sources, /api/privilege-log, /api/qc-findings, /api/retention-events, /api/remediation-actions, and /api/documents/search as needed. POST /api/query runs read-only SELECT SQL — the reliable way to look up a specific record id or aggregate counts (documents/search caps at 100 rows and ignores unknown filters).

Core procedure

1. Anchor on remediation actions. Pull /api/remediation-actions for the matter. The material findings are the actions whose action_id is not a *-NOISE-* action and whose target_ref is a specific record id (SRC-…, PRIV-…, QC-…, RET-…, DOC-…), not a bare category code. These target records are your finding set. Their action_type, owner, priority, severity, and due_days feed the action plan directly.
2. Fetch each anchored record by id from its table (via SQL or the matching endpoint) to get full detail: statuses, counts, dates, tags, categories.
3. Sweep for material records the actions missed. Scan custodian_sources, privilege_entries, qc_findings, retention_events, and production_stats for records that pass the material test below even if no action points at them (privilege gaps and pre/post-hold retention losses in particular often need independent judgement).
4. Discard everything that fails the material test — do not let noise into findings, categories, or metrics.
5. Classify each material record into the gap taxonomy and map its hub fields to the template's enums (reference/discrimination_and_mapping.md).
6. Compute metrics from the material set only (see Metrics below).
7. Build category coverage — one entry per request category that has a material non-complete status; aggregate the findings touching that category.
8. Build the prioritized action plan from the material remediation actions (+ any finding needing an action the hub didn't pre-stage), translating owner and action_type into the template's enums.
9. Assemble, sort, validate, and return exactly one JSON object.

The material test (this is the whole game)

Treat a record as material when it shows the marks of a real, unresolved, escalated production issue:

• It is the target_ref of a non-noise (P1/P2) remediation action; or
• Its id carries a descriptive/semantic suffix describing the issue (e.g. …-LOG-GAP, …-POST, …-SIGNAL, …-SMS, …-TEAMS-ARCHIVE, …-ZERO-CLAIM, …-MISCODED-PRIV, …-CONSULTANT, …-OVERDESIG, …-LAPTOP, …-GMAIL, …-SHARE-DEL, …-EHS-POST) — as opposed to a plain sequential id (PRIV-<MATTER>-007, QC-<MATTER>-003); and/or
• Its note states a concrete, quantified, unresolved problem (e.g. "Only 365 of 840 withheld privileged documents are logged"; "Six off-site bid file boxes destroyed after hold"; "Signal messages were identified in custodian interview but not collected"), with a loss/gap status (not_collected, lost, post_hold_loss, should_exist_missing, destroyed) and/or non-routine issue_tags.

Treat a record as noise (exclude it) when it shows the distractor marks:

• action_id contains NOISE, or priority: P3, or severity: low, or target_ref is a bare category code, or owner is Matter Associate / Vendor Team, or action_type is sampling_review / load_file_cleanup / custodian_followup;
• a plain sequential id and/or issue_tags: ["routine"];
• a hedging / dismissive note — the recurring distractor phrasings include "included to create similar labels across matters", "ordinary review variance", "marked … for follow-up but not immediate remediation", "requires category-level context before escalation", "noisy but not dispositive", "re-sampling before escalation", "similar to escalated records in another matter", "remediated by archive collection", "no unresolved production impact", "no production-impacting issue has been escalated yet", "not one of the stable exception records", "clean", "requires matter-level filtering".

Do NOT select on severity or raw counts. Noise records routinely carry severity: high and large doc_count/unlogged values specifically to bait you. A privilege entry with unlogged=46 and note "marked for follow-up but not immediate remediation" is noise; the material one is the descriptive-id entry with the factual note. When a note hedges, exclude — regardless of size.

Metrics

• Compute only from the material set. Where a template says a count is drawn "from selected … blockers only," sum across the material records you kept, not every hub row.
• Privilege math: unlogged = withheld_count − logged_count per entry; roll up the material incomplete-log entries for the withheld/logged/unlogged totals.
• Retention volume: report the material event's volume_count/volume_unit (e.g. boxes) — pre-hold policy-compliant destruction is not a loss and is excluded from loss metrics.
• Counts are whole integers; use 0 (not null) for "0 when not applicable".
• production_ready / rolling_production_ready is true only if there are no open material blockers — with a real material finding it is false.

See reference/discrimination_and_mapping.md for the full gap taxonomy, pre/post-hold retention logic, and the hub→template owner/action_type/enum translation tables (the raw hub owner and action vocabularies do not match the template enums and must be mapped).

Output contract

• Return exactly one JSON object and no prose outside it. If the task names a required answer file (e.g. answer.json), also write the same object there.
• Include exactly the required_top_level_keys; set matter_id to the hub id.
• Every enum field value must be a member of that field's template enum list. If no hub value maps cleanly, pick the closest allowed enum — never invent a value or reuse a raw hub string that isn't in the enum.
• Apply ordering_rules precisely: sort each list by its key ascending; sort priority/rank lists ascending with 1 = highest priority.
• Category codes: uppercase, sorted ascending, in every category set.
• source_refs / target_refs / record_refs / blocking_refs: stable hub ids exactly as they appear, sorted ascending.
• Only list categories/records with a material non-complete status (per each list's description in the template).
• Honor numeric_precision; use null only where the schema explicitly allows.

Before returning — verify

• JSON parses; top-level keys == required_top_level_keys; every list sorted per ordering_rules.
• Every enum value is in the template's enum for that field.
• Every finding/category/action traces to a material record; no noise id, no *-NOISE-* action, no bare-category target_ref leaked in.
• Metrics cross-foot: unlogged = withheld − logged; counts match the cited material records; booleans reflect open blockers.
• All evidence came from the hub; no local db/seed/manifest/answer file used.