How this dashboard works

1. 01
   Discovery. A curated regulatory matrix per jurisdiction is fed to a large language model with live web-search grounding, which identifies the statutes, regulations, circulars and guidance notes worth tracking. Output: 1,267 candidate norms across 23 jurisdictions.
2. 02
   Scrape. Each norm is fetched from the official source (national gazettes, regulator portals, statutory databases). When a primary source is unreachable — strict firewalls, login walls, JavaScript-rendered single pages — the pipeline falls back to archive snapshots or to a headless browser, with the chosen path logged per norm.
3. 03
   Translate. Non-English texts are auto-translated; the original language is preserved alongside so legal verification can quote either side.
4. 04
   Analyse. A reasoning LLM reads each body and extracts thirteen structured signals: regime, status_regulatorio, principal deadline + deadline type, seven exige_* service triggers (audit, proof-of-reserves, pentest, AML/KYT, custody, formal verification, independent certification), free-text escopo, and gap_ou_ambiguidade. A second deterministic pass — pure algorithm, no LLM — then re-validates every claim: substring presence of the quote in the body, imperative-verb test for triggers, temporal-anchor test for deadlines, regime-vocabulary test for regimes. Anything that fails the validator is stripped back to null. Phase 1 rule: every non-null value MUST come with a verbatim quote copied from the body. Better silence than a fabricated claim.
5. 05
   Aggregate & export. Per-country overviews are built from the underlying norms. Any-true on triggers, earliest non-past deadline, six text-grounded coverage dimensions (issuance / custody / market abuse / AML / taxation / consumer protection) that drive the maturity tag. A flat CSV plus a typed graph JSON are exported and consumed by this dashboard.

Total edges: 2,999 across nine typed relations. See src/typed_relations.py for the derivation rules.

Six dimensions form the spine of a comprehensive crypto framework: token issuance, custody of client assets, market abuse, AML / KYT / Travel Rule, taxation, and consumer protection. For every analysed norm we detect which dimensions it touches — via boolean triggers plus a multilingual word-boundary keyword scan. A jurisdiction's maturity then ranks the union of dimensions across all its norms.

The previous 40-norms / 3-regulators / anchor-before-2020 heuristic is removed; see src/coverage.py.

• 40% Urgency. Days to the next regulatory deadline. Past-due decays from 100 toward a 50 floor over 12 months. Missing deadline + known regime = 30. Missing deadline + unknown regime = 0.
• 40% Service intensity. Share of the 14 CertiK services triggered (audit, pentest, AML/KYT, proof of reserves, …).
• 20% Market maturity. The text-grounded six-dimension count from section 4.

Formula in web/lib/scoring.ts. Weight sensitivity has been swept; nine jurisdictions stay top-12 under every combination — see the calibration note below.

A retroactive audit (scripts/audit_deadlines.py) confirmed only 16 of 83 norm-level deadlines had body grounding; the remaining 67 — many of them publication dates, founding years or inferred end-of-period markers wrongly flagged as deadlines — were removed from the vault. After the Phase 1 rerun, deadlines re-appear if and only if the body quotes the date next to a temporal anchor ("by", "before", "até", "spätestens", "deadline", "in force") within 80 characters.

For every jurisdiction we publish a 0-100 data-confidence score and a tier badge. Four orthogonal components: how much of the country's corpus is LLM-analysed (35%), how many of the six dimensions are covered (25%), how many distinct regulators contribute (15%), and the share of extracted fields backed by a Phase 1 verbatim quote (25%). The badge shown on the home table, the country profile and the recommended-moves cards reflects this score.

The gold set is a stratified sample seeded with the current extraction; a human reviewer corrects each field and pastes a verbatim source quote, then flips reviewed: true. The comparator (scripts/gold.py report) reports per-field precision, recall and F1 against a saved baseline. CI fails any PR that drops a field by more than 5 percentage points.

Current measurement on a 13-row starter pack: support-weighted F1 = 0.80. The model is perfect on exige_certificacao_independente and escopo (1.00) and strong on regime (0.91). The two structural weaknesses are unchanged after the Phase 1 rerun: a bias toward TRUE on exige_seguranca_custodia (0.36) and exige_kyt_aml (0.50), where the source body uses descriptive prose rather than the imperative verbs the validator demands. Net effect for the decision-maker: those two triggers should be cross-checked manually before any commercial move; the other eleven fields are reliable.

• Two jurisdictions (IN, SE) have only stub norms — the analysis pass returned no usable signals. They are shown for completeness, not for ranking.
• competidores_ativos and forca_relacionamento_certik are intentionally empty — they require business-development input that the LLM cannot infer from public text.
• A handful of national gazettes are behind aggressive firewalls and were fetched from official archive snapshots; their content can be slightly stale until the next refresh cycle.
• The opportunity score is a heuristic — a high score is a candidate to investigate, not a closed sale.