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.
Trust, provenance, and limitations
How Pure Global AI regulatory agents work
Our agents are designed to 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.
Evidence extraction
User text is treated as untrusted evidence. Documents, facts, assumptions, unknowns, revisions, and contradictions are separated before assessment.
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
- 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 | OpenRouter using DeepSeek V4 Flash 0731, fixed in source code |
| 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We do not publish a universal accuracy percentage.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.