Invented cases · research, not medicine
One next check, for a clinician to review
Pick a made-up rare-disease case. Fixed rules on this page check first whether the person may be acutely unwell. If so, the case goes to a person and no AI is asked. Otherwise an AI reads the numbered facts inside verified hardware and returns a brief: the leading explanation, the gaps in the evidence, and one next check, each tied to the facts.
-
1
Pick a made-up case
Each case shows one path: a well-supported lead, too little evidence, or an urgent presentation.
-
2
The facts, numbered, with their sources
Only these facts are sent to the AI, encrypted. The case's title and the path it shows stay on this page.
Pick a case first.
-
3
Is it urgent? Fixed rules decide, before any AI
The AI never judges urgency. These rules read the observations with NEWS2 (adults), the lab's own critical flags, and a clinician's confirmation that the person is not acutely unwell.
-
4
Check the hardware
Your browser asks NEAR AI Cloud for a fresh hardware report on the model and checks it here: Intel's signature over the TDX quote, the model key inside it, the software it started with, and NVIDIA's verdict on the GPUs. The same checks as DNA Confidential.
-
5
Encrypt, send, and check the brief
The facts are encrypted on this page to the attested model key. Our server relays ciphertext only. The reply comes back encrypted to a key this page made, signed by the attested key, and is checked here.
How it works
- Facts first. A case is a list of numbered facts, each with its source. The brief must cite them by number, and the page checks every citation.
- Urgency by fixed rules, not by the AI. A result the lab flagged as critical, or a NEWS2 score with a 3 in any parameter or a total of 5 or more (Royal College of Physicians, 2017), stops the case: a person first, no AI. A case goes on only when a complete adult NEWS2 score is low or a named clinician confirmed the person is not acutely unwell. Otherwise it is held.
- The hardware is checked before anything is sent. The browser verifies that the model key belongs to
z-ai/glm-5.3-flashrunning in an Intel TDX confidential VM with its NVIDIA GPUs in confidential mode, bound to a fresh nonce. If any check fails, nothing is sent. - Only ciphertext on the way. The facts are encrypted to that key with NEAR AI's end-to-end encryption (version 3). The relay adds the API key and checks the request's fixed shape; it cannot read it.
- The brief is checked on the page. One JSON brief; exactly one next check, with why and what it cannot show; every point citing this case's facts; a lead that states its confidence; and no variant classification or phase written by the model, because the page shows those from the lab report. Flags point at wording about doses, treatment or a settled diagnosis. Then the reply's signature is verified against the attested key.
What this does not show
- That the advice is right. The checks cover form and traceability. Whether the next check is the right one is the clinician's call; a study with clinicians would have to measure that.
- That the urgency rules are complete. They check only what they read. NEWS2 is for adults and reads observations, not blood results; a child goes on only with a clinician's confirmation.
- Anything about real patients. The cases are invented. Real use would need clinical governance, consent and a reviewed data path.
- The hardware makers and the model's software. The attestation proves which hardware and measured software hold the key. It trusts Intel's and NVIDIA's keys, and it does not prove that the software has no bugs.