Research
How we make sure Cynaps is safe and accurate
A physician should be able to check any answer Cynaps gives, and know how it was tested before it reached them. This page explains both: how answers are tied to their sources, how Cynaps Chat is tested before it opens to physicians, and what has been measured so far.
Cynaps Chat is in preview. It does not answer clinical questions yet; it opens to verified physicians only after it passes the tests below.
Grounding
Every answer is tied to its sources
Written from what was found, not from memory
An answer is written from passages retrieved for the question — Indian guidance such as ICMR and AIIMS, society guidelines, peer-reviewed literature and standard texts — not from what a model remembers.
Each claim names its source
Every statement carries the source it came from, so you can open the guideline or paper and check the line yourself.
It says when the evidence is thin
Where the sources do not support a confident answer, the answer says so instead of filling the gap.
Before chat opens
How Cynaps Chat is tested before physicians use it
The bar was agreed before any result was seen, so the result cannot move the bar. Chat stays in preview until it passes every step.
- 1
Safety tests
A private set of hard questions — doses, emergencies, pregnancy and children — kept away from the people who build the answers, so the system cannot be tuned to pass them. A critical error stops chat from opening.
- 2
Physician review
A physician reviews every answer the checks flag, and a random sample of the rest.
- 3
Accuracy on questions it has never seen
Cynaps is compared with a leading general AI using web search, on physician-style questions neither has seen before, each answer marked by more than one independent judge.
- 4
The same answer every time in an emergency
Emergency questions are asked again and again; the recommended action has to be the same each time.
Where things stand
What is measured, and what is still in preview
Each line shows where the work stands today. Figures appear here only after they are confirmed and checked by our physicians.
Finding the right evidence
Whether the search puts the passages that answer a question in front of the answer writer.
Being measuredAccuracy against general AI
Cynaps against a leading general AI model on unseen physician-style questions, judged blind.
Being measuredIndia-first specialty question sets
Questions written for Indian practice, specialty by specialty, each checked by a physician before it is used.
In physician reviewSafety question set
The private set of hard questions chat must pass before it opens.
Being preparedBeyond text
How an AI should be tested on the ECG, the scan and the clinical note, not only on text.
PlannedCynaps Chat
Evidence-based answers with their sources, for verified physicians once the tests above are passed.
In preview
Write-ups
Plain-language write-ups
Short write-ups for physicians, published as each finding is confirmed. The first are in preparation.
- In preparation
Marking the marker
How much an AI judge’s verdict moves when nothing about the answer has changed, and why we report that before reporting any score.