Rättskällor med officiella primärkällor

Utskrivet ·

Skip to main content
Skip to the answer

United States regulatory decision agents

Ten deterministic agents, one per federal or state core rule that reaches nearly every company operating in the United States. Each agent reads its register row, tests the statutory elements in a hand written tree and returns a machine outcome with a source chain, a hash and a read date. No language model writes a field.

Read 2026-08-15 · 10 · deterministic, no model in the answer path

About the US agent layer

The agents are deterministic rule trees over US federal regulation. Each agent carries jurisdiction scope, risk dimensions, coverage score, change impact and tier, and every conclusion traces to the section it rests on.

Beneath each agent sit section nodes, one address per cited CFR or USC section, so an answer can be quoted at section level rather than only at regime level.

  • Read the answer block first, the rule tree after it.
  • Section nodes give the citable address for each provision.
  • The machine paths under each agent return the same content as JSON.

How the agent reaches the outcome

Inputs are closed fields, never free text. The engine tests the rules in order and stops at the first match. The outcome has three parts that always travel together: a verdict, an audit code and a neutral statement of what the rule text requires, with a paragraph reference. If no rule matches, the agent abstains rather than guesses. Every run yields a hash over the inputs and the tree, so the same question can be shown to give the same answer afterwards.

Machine surface

  • GET /api/public/v1/us-agents
  • GET /api/public/v1/us-agents/{id}/manifest.json
  • GET /api/public/v1/us-agents/{id}/run
  • GET /api/public/v1/us-agents/{id}/monitor.json

The verdict is a machine classification of the outcome, not legal advice and not a compliance decision. Responsibility position · Register of United States rules · Example: California consumer privacy

Verifiable trust signals

  • Six fixed blocks, one source per line
  • No sentence written by a language model
  • Engine version and read date on every answer
  • No customer data, no documents, no advice
  • Model card and audit published under the EU AI Act

Model cardAudit