SafeFuture Lab

Ask the work, not the model

Every agent here is restricted to one published piece of road-safety analysis and its cited sources. Answers are machine-verified before you see them; when an answer can't be verified, the agent refuses. No general chatbot. No improvising.

Project-restricted · machine-verified · refuse-when-unverifiable Humans audit every interaction Independent of all governments & agencies
The agents

Three published works, three verified assistants

How these agents work

The rules every agent on this page obeys

  1. Project-restricted. Each agent answers only from one published work and its cited public sources — never from general model knowledge.
  2. Machine-verified. Numbers come from read-only queries against the same data that produced the published charts; a deterministic checker validates every figure in every answer before it is shown.
  3. Refuse rather than guess. Out-of-scope or unverifiable questions get an honest refusal, not an improvisation.
  4. Audited. Every interaction is logged and reviewed by a human; flagged answers are corrected or withdrawn.
  5. Disclosed bias, open data. Authors declare their interests; claims ship with falsification tests where applicable; sources are public and named.