Leak Record Classifier

Leak Record Classifier

Shared scaffold for the recurring "client hands us a leaked-database CSV, find the records that belong to our subject" task. Consolidated 2026-07-21 after two cases (case-001-example, case-001-example) independently rewrote the same ~250-line Maltego-parsing + classification + Excel-writing scaffold from scratch, differing only in each subject's identity anchors.

Prerequisite -- ingest first (enforced by a hook, not optional)

The source CSV must already be registered as an EV-NNNN item and deterministically extracted before you write or run a classification script against it:

bash skills/osint/scripts/ingest-client-document.sh <file> <slug> document \
  --case <case-folder> --provided-by "<who>"

python3 skills/osint/scripts/extract-intake.py --case <case-folder>

skills/osint/scripts/evidence-pipeline-guard.py (wired as a PostToolUse hook in .claude/settings.json) BLOCKS writing or running a script under a case's investigation/evidence/scripts/ if either step above is missing for any file in that case's investigation/intake/. This exists because a prior case skipped both tools entirely (see .claude/rules/evidence-capture-protocol.md's scar note).

Usage

1. Copy templates/case_classifier_template.py to the case's investigation/evidence/scripts/reformat_csv.py.
2. Do a recon pass on the extracted CSV first (investigation/evidence/extracted/<stem>/text.md) to find the subject's real SSN/DOB/address/email anchors -- do not guess, verify each one against actual rows, the way both origin cases did (grep for the SSN digit string, confirm DOB co-occurrence, check which addresses cluster together).
3. Fill in TARGET_SSN_DIGITS, TARGET_DOB, CONFIRMED_ADDR_TOKENS, CONFIRMED_EMAILS, and classify_row() in the copied file. Define your own tier vocabulary in TIER_CONFIG / CLEANED_TIERS / FILTERED_OUT_TIERS -- the two origin cases used different tiers (one had a "household" tier for a blended family, the other didn't need one) because every case's data is different. Don't force a case into tiers it doesn't need.
4. Precedence lesson (read this before writing classify_row): case-039 shipped an early version where an SSN-matches-but-DOB-conflicts row got silently absorbed into "confirmed" because a weaker address-match branch ran first and returned before the conflict check was reached. Check for genuine identifier conflicts FIRST, before any weaker corroborating signal can short-circuit past them. Leak databases (National Public Data especially) have real, documented record-matching quality issues.
5. Run it: python3 investigation/evidence/scripts/reformat_csv.py --case <case-folder>

What's shared vs. case-specific

Report generation (the docx synopsis) stays entirely case-specific and is NOT part of this skill -- every subject's findings, employment, household, and narrative are genuinely different and don't benefit from a shared template beyond what the report structure already gives (see any case's output/briefs/generate-report.js for that pattern).

Origin cases (not migrated -- already delivered, left as-is)

• investigations/case-001-example/investigation/evidence/scripts/reformat_csv.py
• investigations/case-001-example/investigation/evidence/scripts/reformat_csv.py

Verification

lib/leak_csv.py was verified by reimplementing case-039's exact classify_row() on top of it and confirming an identical classification result against the real 170-row source CSV (37 confirmed / 8 flagged / 50 household / 6 excluded / 69 unrelated -- exact match, not approximate).

Known follow-up (not part of this skill, flagged separately)

A second, larger duplication exists: extract_intake.py / extract_all.py variants live in 5 case folders (case-010, case-013, case-020, case-025, case-028) alongside the canonical skills/osint/scripts/extract-intake.py, all added in the same 2026-06-14 commit and never pruned. Some have real per-case additions (e.g. case-020's ocr_hd.py, case-028's ocr_pass2.py), so this needs a careful case-by-case review before consolidating, not a blanket delete. Out of scope here.