§ 01
1. Source identification
Our research team identifies the official legal text for a regulation - the legislation itself, not commentary about it. Each source is recorded with its level (official legislation, recognized standard, or secondary reference) and publisher.
§ 02
2. Requirement extraction
Requirements are extracted as structured records: what is required, of whom, under what conditions, effective from when. Each carries a confidence level and verification status.
§ 03
3. Rule authoring
Applicability rules encode the conditions: if the product is in category X, destined for jurisdiction Y, with attributes Z - then requirement A applies. Rules are reviewed before activation.
§ 04
4. Verification
Rules and requirements are verified against the official source before publication, with a last-verified date. Draft or unverified content never drives customer-facing conclusions.
§ 05
5. Monitoring
Published regulations are re-checked on a schedule. Amendments trigger re-verification of affected rules and alerts to monitoring subscribers.
§ 06
6. How confidence is calculated
Every assessment ends with a confidence level: High, Medium, Low, or Unable to determine. We deliberately use four plain levels — never a percentage — because a score like 87% would imply a precision we do not have.
The calculation starts from the weakest confidence among the rules that fired for your product. It then steps down one level for each risk factor: weak source authority (secondary references, or AI/user-provided sources, cap the result at Medium); stale verification (a source never verified, or verified more than 180 days ago); incomplete product attributes; a destination that does not map cleanly to a jurisdiction; and contradictory sources.
When no verified rule fires for a product and market, the answer is Unable to determine — listed openly with what is missing, never hidden and never guessed.
§ 07
7. What AI does
AI assists; it never decides. The model suggests HS codes for human review, extracts attributes from uploaded documents, drafts summaries and report text, helps researchers triage regulatory changes, and powers the support chatbot. Every AI output is a suggestion or a draft.
AI work is guardrailed: prompts forbid fabricating legislation, certifications, HS codes, regulator names or official URLs, and require the model to cite provided sources or say unable to determine when they do not answer the question. Deterministic results are cached so identical inputs always produce identical outputs.
§ 08
8. What AI does not do
AI never determines whether a requirement applies to your product — that is the deterministic rule engine's job, reading verified rules, requirements and source records. AI never publishes content, never changes a rule, and never issues an authoritative customs or legal determination.
When AI assistance is unavailable, the platform falls back to verified database records rather than generating an answer. An HS code suggested by AI, for example, must still be verified by a customs professional before use.
§ 09
9. How human review works
Applicability rules are reviewed by a person before activation, and every requirement carries a verification status with a last-verified date. Draft or unverified content never drives customer-facing conclusions — the quality gate keeps it unpublished until it passes.
Assessments can be flagged for review, and review notes record what a reviewer checked. Published regulations are re-verified on a schedule, and amendments trigger re-verification of affected rules plus alerts to monitoring subscribers.