Approval Trail Audit Log Builder

Goal

Produce a traceable, end-to-end approval trail for every active matter by extracting, normalizing, and validating evidence from onboarding documents, signed forms, and approval emails. Deliver a structured audit log plus an exceptions report that highlights missing signatures, missing approvals, unreadable files, or conflicting instructions.

Inputs to gather

• Active matters list (IDs, names, owners, types/status — defines the reconciliation universe).
• Approval policy/matrix (required approver roles, thresholds, acceptable evidence types — drives validation rules).
• Source locations or files for onboarding documents, signed forms, and approval emails (where to read evidence).
• Time window and timezone preference (ensures consistent timestamp handling; default ISO 8601 with timezone).
• ID patterns or naming conventions in filenames/subjects (improves matter matching accuracy). Before doing any work, ask the user for these inputs in ONE message. Skip anything they already provided. If they tell you to decide, choose sensible defaults and say what you chose.

Workflow

1. Confirm scope
   a. Verify the active matters list, approval matrix, time window, timezone, output format (CSV and/or JSONL), and any redaction requirements.

2. Inventory sources
   a. Request or read any manifest of files; if none exists, ask the user for one.
   b. Use Read to open representative samples of each source type to confirm accessibility; note any image-only, encrypted, or unreadable items for escalation.

3. Normalize and fingerprint each source
   a. Extract full text and metadata (title, author, creation/modification dates).
   b. Generate a source fingerprint (filename, path/URI, byte size, checksum if available).
   c. If OCR is required but unavailable, flag the item as “needs OCR” in the exceptions report and pause further extraction for that file.

4. Parse approval evidence
   a. Onboarding documents: pull matter ID, requester, owner, start date, embedded approval sections or sign-off fields.
   b. Signed forms: detect each signer, role, signature presence, signature timestamp, and decision (approve/deny/conditional). Record whether the signature is digital or handwritten and include any certificate block metadata.
   c. Approval emails: for every message, capture sender, recipients, subject, message date/time, explicit approval or denial language, message-id, and threading data. Exclude FYIs or tentative statements.

5. Map evidence to matters
   a. First match by explicit matter ID; if absent, use normalized matter name, requester, owner, and timeframe heuristics.
   b. If a confident match cannot be made, tag the evidence “unmapped” and include it in the exceptions report. Do not guess.

6. Create structured approval events
   For every approval-related signal, create an event object with: matter_id, matter_name, matter_type, event_type (requested|approved|denied|revoked|acknowledged|conditional), approver_name, approver_role, actor_email (if email), decision_note, event_timestamp (ISO 8601), evidence_type (signed-form|email|onboarding-doc|other), source_reference, source_fingerprint, extraction_confidence (high|medium|low). If conditions exist, set event_type to conditional and include the text.

7. Reconcile events per matter
   a. Sort chronologically.
   b. Deduplicate near-identical events generated from forwards or duplicate forms, keeping the earliest authoritative record.
   c. Detect conflicts (e.g., denial after approval, contradictory directives) or approvals by the wrong role and mark them for escalation.

8. Validate against the approval matrix
   a. Compare observed events to required approver roles and thresholds.
   b. Identify missing approvals or signatures with exact role/step details.
   c. Note any unclear policy items for human clarification.

9. Generate outputs with Edit
   a. audit_log.csv or audit_log.jsonl: one row/object per event (schema from step 6).
   b. exceptions_report.md: per-matter list of missing approvals/signatures, unmapped evidence, conflicting instructions, unreadable/needs-OCR sources, and ambiguous policy items; include source_reference and a concise next action.
   c. reconciliation_summary.csv: per matter, status (complete|incomplete|conflicted), last approval date, outstanding items count.

10. Review and handoff
    a. Provide an executive summary (total matters, percent complete, exception counts).
    b. Ask the user to supply missing information or clarifications for exceptions; integrate fixes and regenerate outputs as needed.
    c. Archive outputs with a run timestamp to ensure traceability.

Throughout, never fabricate data or timestamps. Treat tentative language as non-final. Preserve original timezones per event, redact sensitive PII if requested, and keep a changelog for reruns.

Output

1. audit_log.csv or audit_log.jsonl with the full event schema.
2. exceptions_report.md listing all issues and next actions.
3. reconciliation_summary.csv summarizing completeness per matter.
4. Brief executive summary in-chat highlighting key statistics and any items requiring user action.