How Pure Global AI regulatory agents work
Our agents produce bounded regulatory workpapers and editable drafts — not confident-sounding compliance verdicts. This page documents the common runtime and the boundary between automation and accountable human review.
The common agent pipeline
Every public agent runs the same four stages:
- Bounded intake — Each task defines exactly what the user should paste, a character limit, a worked example, and a fixed set of review areas or JSON fields.
- Verified internal context — Country-specific agents read one structured Pure Global context. Source-linked regulator facts, maintained certification rules, planning estimates, official-source indexes, and registry analogs remain separate evidence tiers.
- Focused current research — Tasks that need current information can search public sources, read the most important primary source, and query relevant public registration data.
- Structured validation — The workflow must submit a typed tool result. Official forms and the Product Master use strict JSON contracts before deterministic file writers run.
What the model may and may not do
The model may:
- Use supplied facts and explicitly label direct, derived, unknown, and conflicting information.
- Use focused public research when current requirements or guidance materially affect the task.
- Return not assessed or null when evidence is insufficient instead of inventing a result.
- Create a preliminary workpaper, JSON contract, or editable draft for human review.
Current runtime
| Orchestration | Durable Vercel Workflow runs with cancellable, token-protected status and progress endpoints |
|---|---|
| Agent harness | Pi through the Vercel AI SDK harness and an isolated Vercel Sandbox |
| Model route | Google Generative AI using Gemini 3.7 Flash, centralized in source code |
| Taxonomy contract | A versioned read-only API enumerates canonical jurisdictions, countries, domains, resources, device types, certifications, fees, timelines, regulations, and URLs for every upper layer |
| Internal context | One read-only regulatory-context layer separates source-verified facts, maintained certification rules, planning estimates, official-source indexes, and registry evidence |
| Research tools | Focused public web search, primary-source reading, registration-database search, and Regulation News search where relevant |
| Public protection | BotID in production, per-IP rate limits, payload limits, run tokens, timeouts, and bounded tool calls |
| File generation | Validated JSON is handed to deterministic DOCX, XLSX, XML, or report code; the model is not trusted to mutate packages directly |
Quality claims and testing status
We do not publish a universal accuracy percentage. Different tasks require different gold sets: field extraction, citation precision, unsupported-claim rate, false-support rate, false-gap rate, file integrity, and task completion are not interchangeable. Current release gates include schema validation, type and lint checks, deterministic document integrity checks, representative end-to-end workflow runs, route and sitemap checks, responsive UI review, and regression examples. A regulator-specific benchmark will be published only after qualified adjudication produces task-level measurements.
Keep reading: Data methodology → · Browse every tool →
