How it works

From a consultation to a note you have actually read.

The clinical workflow is deliberately unglamorous: capture, inspect, edit, approve, export. The interesting decisions are in what the system refuses to do at each step.

Step by step

Five stages, and the refusal at each one.

  1. Start a consultation

    A consultation is keyed to a hospital record number, not a patient name. Names are not used as identifiers anywhere in the workflow.

    Refuses to make the patient's name the primary key.

  2. Capture — or type

    Before recording, the screen states which transcription route is used and how long audio is kept. Two input paths exist: audio capture, and a typed transcript for settings where recording is not appropriate.

    They are not equivalent The audio path runs a normalisation layer on recognition output; the typed path does not. In English this rarely matters. In code-switched Malayalam it currently changes the result.

  3. Inspect the evidence

    The transcript is shown with speaker attribution and a parallel fact panel. Each fact carries a status, a subject, and the span it came from. Selecting a claim in the note highlights the words that support it.

    Refuses to show a claim with no span behind it.

  4. Review and edit the draft

    The note is organised under fourteen standard sections. Sections without evidence say “Not documented” rather than being filled with plausible content. Any clinician edit moves the note to reviewed — never to approved.

    Refuses to treat generation as review, or review as approval.

  5. Approve and export

    Approval is a separate, credentialed action by a named clinician. Only then can the note be exported, and only then does the audio retention clock start. An approved note cannot be silently edited afterwards.

    Refuses to export an unapproved draft.

The review screen

Where the clinician actually spends time.

Everything before this screen is preparation. The review screen is the product: a draft on one side, the evidence behind each line on the other.

Clinoria review screen on a synthetic ultrasound case, showing a validated fact graph and the rendered report sections with laterality and measurement detail.
Synthetic Prototype review screen. The header reads “Fact graph validated”, which means the validator found every rendered claim supported — not that the note is clinically correct.

Reading

Facts before prose

The fact panel is shown alongside the note, so disagreement between what was said and what was written is visible rather than inferred.

Editing

Edits move the state

Any change transitions the note to reviewed and records that a human changed it. Approval is a deliberate second action.

Signature

Approval is the retention event

Approving is what makes the note exportable and starts the 24-hour audio deletion clock. Nothing else in the workflow does.

The note

Fourteen sections, including the empty ones.

Clinoria renders into a fixed documentation structure rather than free prose. The shape is intentional: a clinician reviewing a note needs to know where to look, and a missing section is itself clinical information.

  • Presenting complaint and history of presenting illness.
  • Relevant positives and pertinent negatives, kept distinct.
  • Temporal course: onset, duration, resolution.
  • Family and social history, attributed to the right subject.
  • Examination findings and measurements, preserved verbatim.
  • Impression sections that state only what the transcript supports.

Where the conversation was silent, the section says so. An assessment or plan the patient never discussed is not invented to make the note look complete.

What the note is not

  • Not a diagnosis.
  • Not a clinical decision.
  • Not a substitute for examining the patient.
  • Not validated against clinician-authored references at scale.

The honest summary

A faithful, structured, evidence-linked draft that is often linguistically plainer than a generated note, and occasionally wrong in ways the validator does not catch. That last clause is the research programme.

What we demonstrate

Text and audio are shown separately, on purpose.

A demonstration is not a highlight reel. It runs both input paths, states which one is being used at each stage, and includes cases the engine gets wrong.

Path A

Typed transcript

A synthetic consultation entered as text. Exercises fact extraction, rendering, validation and the approval gate. Fast, deterministic, reproducible.

Path B

Audio capture

A synthetic recording through speech recognition, then the same downstream stages. Slower, non-deterministic at the transcription boundary, and configured locally rather than through a third-party service.

Say it plainly Anyone demonstrating Clinoria is expected to state which path is running. Presenting a typed transcript as if it were transcribed audio is the most common way clinical AI demos mislead.

See it

Bring a case you think will break it.

We would rather find the failure with you in the room than after publication.

Request a research demonstration Read the defect register