Disclosure
Defects stay listed
A defect is removed from the register only when a test covers it, and it is then marked resolved with a date rather than deleted.
Responsible AI
Assistance, not autonomy — enforced in code.
Clinical safety principles are cheap to publish and expensive to implement. These are the ones we have actually built into the prototype, followed by the ones we have not earned yet.
Principles
Six commitments, each with a mechanism.
-
The clinician is the author of record
Generation is not review, and review is not approval. Draft, reviewed and approved are distinct server-side states; export is refused unless approved. The clinician's signature is a code path, not a policy statement.
-
Evidence outranks fluency
No language model writes note text. Every rendered claim is tied to facts, and every fact to a transcript span. Where the evidence is absent, the note says so instead of producing something plausible.
-
Failure is closed, not flagged
A draft that fails validation is withheld from review rather than displayed with a warning banner. A warning invites approval; a refusal does not.
-
Limitations are published, not discovered
Eight open defects are listed on the research page, including two that would be commercially embarrassing to publish. They stay listed while they are open.
-
Language claims follow the code
Malayalam is described as research because it is research. The system's documented failure on transliterated Malayalam is stated on the homepage, not buried.
-
No real patients in public material
Every screenshot, transcript and example on this site is synthetic. No public inference endpoint accepts free text, because a heuristic content filter is not a safeguard.
Claim discipline
What we do not say, and what we say instead.
This table governs every page, slide, email and conversation about Clinoria. It is a working rule, not a values statement.
Prohibited claims
Binding
| Never say | Because | Say instead |
|---|---|---|
| “Clinically validated” | No independent validation exists | “Pre-clinical; evaluation is the next step” |
| “Zero hallucinations” | A laterality was fabricated and passed validation | “Drafts are validated against the fact graph; that does not prevent wrong facts” |
| “100% accurate” | No measurement supports any percentage | “No accuracy figure is published” |
| “Supports Malayalam” | Transliterated Malayalam fails; one case inverts | “Malayalam is experimental research” |
| “HIPAA / DPDP / GDPR compliant” | No assessment or certification | “Designed with DPDP principles in mind; not certified” |
| “Diagnoses” or “detects” | It drafts documentation | “Extracts findings and drafts a note” |
Regulatory position
What this software is, legally.
Assistive documentation software. It drafts text for a credentialed clinician to review, edit and sign. It does not diagnose, does not recommend treatment, does not triage and does not act autonomously.
No regulatory clearance, approval or registration has been obtained in any jurisdiction. No claim of conformity with any medical device regulation is made.
If clinical deployment is ever contemplated, that is a separate regulatory question that has not been begun.
Clinician responsibility
A draft is not a record. Until a clinician reads, edits and approves it, it has no clinical standing. The system is designed to make that review meaningful — facts and provenance alongside the text — and it cannot make the review optional, automatic or safe to skip.
Governance
How these rules are kept.
Evidence
No number without a source
Any quantitative claim requires an evaluation with clinician-authored gold documentation. Until then, none is published anywhere.
Correction
Public corrections
If something on this site is wrong, it is corrected here with a note. Silent edits to published claims are not acceptable.
The uncomfortable part
These principles do not make the system safe. They make it honest. The extraction is still wrong in documented ways, the validator still accepts self-consistent fabrication, and no clinician has yet evaluated any of it at scale. Responsible AI in this project means publishing that, not resolving it rhetorically.
Hold us to it