Research institute · India · Healthcare first
The control layer for AI that writes to the record.
Models generate findings. We research the gate that decides whether a finding becomes state other clinicians will treat as true. Pre-pilot. Research prototype. No production clinical writes.
Pre-pilot · 1 clinical environment · 3 tested modules · 0 production writes
1 and 2 are not 3.
Permission
A role may call update_note. That is access control. It answers whether this identity may use a tool. It does not answer whether this finding may enter the chart.
Confidence
A model may return 0.94. That is a score over tokens. It is not evidence. It is not authority.
Authorization
Authorization is a named person, bound to a hash of the exact payload, under a named policy version, for this write. If the bytes change after the click, the approval does not travel with them.
Permission to call a tool is not authorization. Model confidence is not authority. A screen click is not approval if the payload can change afterward.
Refuse. Hold. Invalidate.
The gate is easiest to see when it says no.
Refuse
REJECT
Model asserts a perforation. No exam source. The finding cannot be constructed.
Hold
HOLD
The clinician still sees the refusal. It is not dropped. It is not written as fact.
Invalidate
APPROVAL_HASH_MISMATCH
Wording A was approved. The agent swaps wording B.
Specified behaviour of the Clinical Action Record. Hash-binding is Proposed, not enforced.
Inference is not the record.
Model inference
Upstream · not this institute
Clinical Action Record
The gate · Tested in part, Proposed in part
Chart / note
Write · only on ALLOW + bound approval · 0 production writes
A wrong ENT line is later read as fact.
A finding written into an outpatient note is not a chat answer. The next clinician reads it as examination. Completing a silence — “left tympanic membrane normal” when the left ear was not seen — invents a clinical fact.
Flagging that line still leaves it in the chart. Refusing construction means the line does not exist as a finding. The refusal is shown. That is the research bet.
What this is not.
- Not a diagnostic model.
- Not an EMR.
- Not a hallucination detector.
- Not a foundation model.
- Not certified for production clinical use.
- Not a hospital vendor.
We research the write. We do not sell a clinic system.
- 1
- clinical environment
- 3
- tested modules
- 0
- production writes
- 4
- researchers
As of September 2026. Empty directories are listed on Work, not hidden.
If you run clinical AI that can change a record — talk to us.
We are four people. We read what we receive. Schema review and technical argument first. Not procurement.