Build Legal Manifest

For each result that looks relevant, use WebFetch to confirm the page exists and contains useful content. Discard 404s and pages without substantive legal text.

Classify each source:

• document_type: statute | legal_guide
• authority: binding_legal_authority | practical_self_help | official_interpretive
  • binding_legal_authority: statutes, regulations (nysenate.gov, amlegal.com, justia codes)
  • official_interpretive: government agency guides (hcr.ny.gov, nyc.gov, nycourts.gov)
  • practical_self_help: tenant advocacy orgs (metcouncilonhousing.org, legalaidnyc.org)
• jurisdiction: NYC | New York State

Target: 6–15 statutes + guides total.

Phase 2: Search Justia for case law

Use build_manifest.py --justia-search for 3 targeted searches. The topic determines the keywords.

For each search, run:

uv run python -m tenant_legal_guidance.scripts.build_manifest \
  --justia-search \
  --keywords {keyword1} {keyword2} "New York" \
  --max-results 50 \
  --output data/manifests/_tmp_{slug}_{n}.jsonl \
  --filter-relevance

Choose 3 keyword sets that cover different angles of the topic (e.g. legal standard, evidence type, remedy sought). See references/keyword-sets.md for examples by topic.

After all 3 searches, deduplicate by locator across the temp files:

# Print all locators to check for dups — no code needed, just read the files

Review the deduplicated results. Keep cases that have:

• A clear plaintiff/defendant and outcome (tenant win or loss)
• Evidence-to-outcome chain mentioned (not purely procedural)
• Relevance to the topic

Target: ~25 keeper cases. Discard the rest.

Phase 3: Write manifest files

Statutes + guides → data/manifests/{slug}_statutes.jsonl Cases → data/manifests/{slug}_cases.jsonl

Where {slug} is a snake_case identifier for the topic (e.g. habitability, harassment_destabilization).

Each line must be valid JSON in this exact format. See references/manifest-format.md for the complete spec and field rules.

Statute/guide entry:

{"locator": "https://...", "kind": "url", "title": "Full title with section number", "document_type": "statute", "authority": "binding_legal_authority", "jurisdiction": "NYC", "organization": "City of New York", "tags": ["topic_tag", "subtopic_tag"]}

Case entry:

{"locator": "https://law.justia.com/cases/...", "kind": "url", "title": "Party A v Party B", "document_type": "court_opinion", "authority": "binding_legal_authority", "jurisdiction": "New York", "tags": ["housing_court", "topic_tag"], "metadata": {"court": "Housing Court", "decision_date": "YYYY-MM-DD", "citation": "..."}}

Rules:

• document_type MUST NOT be "unknown" — ingestion will reject it
• title must be specific enough to identify the source (include section numbers for statutes)
• tags should include the topic slug and any relevant subtopics
• Delete the temp files after writing the final manifests

Phase 4: Validate and optionally ingest

Validate both manifests:

uv run python -m tenant_legal_guidance.scripts.validate_manifest \
  data/manifests/{slug}_statutes.jsonl

uv run python -m tenant_legal_guidance.scripts.validate_manifest \
  data/manifests/{slug}_cases.jsonl

If validate_manifest doesn't exist, spot-check by reading the files and confirming valid JSON on each line.

Ask the user before ingesting:

"Manifests ready: {n_statutes} statutes/guides and {n_cases} cases. Ingest now? (requires make services-up)"

If yes, ingest:

make services-up  # remind user to run this first if not already up

uv run python -m tenant_legal_guidance.scripts.ingest_all_manifests \
  --skip-existing \
  --concurrency 3 \
  --report data/ingestion_report_{slug}.json

Check the report for failures:

# Read data/ingestion_report_{slug}.json and summarize: how many succeeded, failed, skipped

Update ROADMAP.md: mark the relevant M2/M3 session as complete.

Output Summary

Report to the user:

• Files created: data/manifests/{slug}_statutes.jsonl (N entries), data/manifests/{slug}_cases.jsonl (N entries)
• What was discarded and why (404s, procedural-only cases, etc.)
• Ingestion result if run (N ingested, N failed, N skipped)
• Any manual follow-up needed (e.g. cases that need closer review)