TrustMedical · Vision — not built yet
A door, not a trial.
Vision — not built yetTrustMedical does not exist. This page is a story about a future use of the tools we are building for AI agents: helping a person take part in medical research while sharing as little as possible. The list under “What exists today” is real and linked. Everything else is an idea.
Not a medical product. It does not exist yet. Nothing here is medical advice. This page has no form and collects nothing.
The idea, in five parts
One identity, a separate pass for each study
A person would hold one secret. From it they could make a consent proof that belongs to one study only. That study could check that the person agreed, and agreed once. Two studies could not compare notes to learn it was the same person. It is designed to be the same identity that owns AI agents, so nobody would need a second one for health.
One rulebook for consent
The rules that decide who may act for an AI agent would be the same rules that decide how a person’s consent may be used by a study. Built once, so a consent rule cannot quietly mean something different when the subject is health.
Reputation would decide who gets near a study
Research agents would have a RepID, earned under rules written for research work. Those rules do not exist yet. A study could require a minimum RepID before an agent may touch it, and ask for a proof of it.
Claims about a study would be checked
A statement about a study could go through the check, which is a vote by independent model families. Matching a quote against the document it cites is a different, separate step: a local command, trustshell ground, that looks for the quote on the cited page with no model and no network. It is not part of the check. What cannot be checked would say Not checked, never pass.
Health data lives somewhere else. Always.
The one rule not open for discussion: health data would sit in its own separate database, never in the one the agent network uses. The two would share the maths behind the proofs and nothing else.
What exists today
- The check: paste a sentence and two independent model families each give a verdict. You see what each said, with a receipt. If they cannot decide, it says Not checked, which never counts as a pass.
- RepID and its proof: an agent’s passport shows its RepID and a zero-knowledge range proof over it that can be verified. Today that proof shows the score rather than hiding it.
- How verification works, in more detail.
What does not exist
TrustMedical itself. Consent proofs for a study. The consent rules for studies. RepID rules for research agents. Any database for health data. There is no study to join, and nothing is collected here.
Go deeper