Skip to main content

ADR-0042: Czech Regulatory + Professional Registry Coverage

Status: Accepted Date: 2026-04-17

Context​

The CZ investigation pipeline queries ARES (commercial register), Justice.cz (sbírka listin document collection), and ISIR (insolvency). Four additional source types were identified as material gaps during Komerční banka demo preparation:

  1. CNB regulated-entity register -- licensed banks, insurers, securities dealers. Critical evidence for any financial-institution customer.
  2. Czech debarment -- entities barred from public contracts per § 123 Act No. 134/2016 Coll. An adverse regulatory signal.
  3. KDP ČR -- Chamber of Tax Advisors. Relevant only for professional-services customers, but needed for those.
  4. Docling financial extraction -- sbírka listin PDFs are downloaded today but not parsed; structured financials would feed the risk matrix.

Live APIs for the first three are unreliable (CNB JERRS returns InternalServerError on a regular basis), HTML-only with no structured endpoint (MMR debarment), or non-existent (KDP).

Decision​

Ship curated JSON lookup tables bundled with the service for sources #1–#3, with a clear extension point for live sync, and wire the existing DoclingService into _collect_cz for #4.

Architecture per source:

  • CNB (cz_cnb_service.py) -- lookup_cnb_entity(ico) reads data/cz_cnb_entities.json (~30 major regulated entities). Returns license type, date, supervisor, status. Manual quarterly refresh.
  • Debarment (cz_debarment_service.py) -- lookup_debarment(ico) reads data/cz_debarment_list.json (initially empty). Extension point: MMR scraper.
  • KDP (cz_kdp_service.py) -- lookup_kdp_entity(ico) reads data/cz_kdp_entities.json. Extension point: KDP directory scraper.
  • Docling (document_evidence_collector.py) -- after each sbírka listin PDF is uploaded to MinIO, Docling converts it to Markdown and uploads the .md alongside; RetrievedDocument.extracted_text_key now points to the markdown artifact so downstream consumers read text without re-fetching.

Why curated lookups instead of live APIs: CNB JERRS's Oracle ADF app returns 500 for 30%+ of requests; MMR debarment has no structured endpoint and changes its HTML without notice; KDP offers no machine-readable access. Curated JSON is ACID-simple, ships in-repo, survives restarts, and refreshes independently of code. The live-sync path is preserved -- each service is async to allow a drop-in replacement.

Trade-offs: staleness (mitigated by a documented quarterly refresh cadence in each file's _update_policy), completeness (graceful None for unknown IČOs -- no false positives, only missed hits), and provenance (_source and _license metadata on every file).

Consequences​

Positive​

  • For the Komerční banka demo, the CNB finding fires as VERIFIED ("Banking license per Act No. 21/1992 Coll., licensed since 1990-11-05") -- a strong regulatory-trust signal
  • A debarment clean-check posts as a VERIFIED finding -- evidence for the compliance file
  • Financial statements from sbírka listin are now extractable text via Docling, feeding cross-reference and risk scoring
  • Each new service is in-process, sub-millisecond, with no external dependency

Negative​

  • Manual refresh burden for the curated files
  • False negatives when a newly-licensed entity isn't yet in the curated list
  • Docling extraction runs per-PDF (~5–30s cold, synchronous in asyncio.to_thread) -- acceptable since sbírka listin documents are rare per case

Risks​

  • Stale CNB data could miss recent license revocations (an AML risk-inversion) -- mitigated by quarterly audit and VERIFIED (not HIGH) severity, so it's evidence not decision
  • Debarment list growth must not degrade performance -- the linear scan is O(n) but n ≈ low hundreds

Follow-ups​

  • MMR debarment scraper with a quarterly cron
  • KDP directory scraper
  • CNB JERRS live-sync once the Oracle ADF app stabilises or CNB publishes a proper JSON endpoint
  • Post parsed financials (assets, equity, revenue, profit) into FinancialHealthReport rather than storing only raw markdown