DRAFT — NOT LEGAL ADVICE. Clinical and attorney review required before use.

TEMPLATE. Draft for ViaVela Solutions LLC describing how ViaVela Medeor's AI features behave. Intended to be given to any clinician or organization evaluating Medeor, and to be attached to the Evaluation Agreement. Have both a licensed attorney and a licensed clinician review it before you rely on it. Technical claims checked against the Medeor codebase at c1b1d61 on 2026-10-25; first drafted 2026-08-18. If the AI behaviour changes, this document must change with it — its whole value is that it is accurate.

ViaVela Medeor — Clinical AI Disclosure

Last updated: [DATE — fill before publishing]

1. The one-sentence version

Medeor's AI reads what has been recorded about a visit and offers a licensed clinician an advisory urgency signal and a list of things to evaluate or rule out. Its outputs are advisory. The licensed clinician makes the assessment and decisions; AI output is not automatically copied into the clinician's note, diagnosis codes, prescriptions or orders.

2. What the AI produces

• An advisory urgency signal — one of emergent, urgent, routine, or unassessed.
• Considerations — things for the clinician to evaluate or rule out, phrased as prompts rather than conclusions.
• A pre-consult brief — a summary for the clinician before the encounter.
• Separate red-flag items — model-suggested findings for the clinician to check before the other considerations.

AI output fields are designed for decision support rather than diagnosis or verdicts. Their text can still contain an inappropriate conclusion or advice; the field names and filters do not guarantee safe or accurate content. ICD codes and the SOAP assessment are entered by the human clinician.

3. Software checks and their limits

The software applies the following checks. They do not validate the model's medical reasoning or replace review of every output:

• Patient-facing responses omit AI material. A patient's own intake response and record export omit AI assessments and pre-consult briefs. This does not prevent a person from sharing material outside those paths.
• A bounded phrase filter withholds some certainty and order language. Triage and pre-consult considerations and red-flag items are checked against a fixed English phrase list. Matching lines without a recognized hedge are withheld and a notice is added. Unlisted wording, other languages and lines containing a recognized hedge can pass. The pre-consult summary and relevant-history text are not checked by this phrase filter. Clinicians must review those fields as well as the filtered lists.
• Thin input removes unwarranted routine reassurance. Where the recorded intake carries no meaningful symptom narrative, a reassuring routine signal is downgraded to unassessed and the thin-evidence state is stated. Escalations are never suppressed this way — hedging must not silence an alarm.
• Structured red flags can raise advisory urgency. A nonempty red-flag list remaining after phrase filtering raises a routine or unassessed signal to urgent; an emergent signal is retained. This does not establish that the listed findings are present or that the model found every relevant warning sign. A red-flag item withheld by the phrase filter does not trigger this escalation. If all red flags are withheld, a routine signal can remain, or become unassessed on thin input; the withheld-items notice is not an urgency assessment. Clinicians must review that notice and the recorded intake.
• Grounding is not guaranteed. The model receives recorded intake and history, but the summary and relevant-history text are not validated against source records. The model can invent or misstate details. The clinician must check them against the record and appropriate references.
• AI output has no automatic action path. The generated material is stored as an advisory assessment or brief. It does not automatically enter a clinician note, diagnosis code, prescription or order.

4. Where the processing happens

The default provider runs locally, on the clinic's own hardware, so clinical text used for AI features does not leave the premises. An optional cloud provider exists, is off by default, and when enabled sends content to that third party — which is a decision the organization must make knowingly. When neither is reachable, the system returns a clearly labelled placeholder rather than fabricating output.

5. Known limits — state these to clinicians

• The AI can be wrong, including confidently wrong, and can miss a presentation a clinician would catch.
• It reasons only from what has been recorded. An incomplete intake produces a weaker signal, and the system cannot know what was never entered.
• It has no access to vitals trends over time, imaging, or laboratory results beyond what appears in the recorded text it is given.
• It is not validated against clinical outcomes. No study supports its accuracy, sensitivity, or specificity, and none is claimed.
• Long encounters can exceed the model's context; the system does not yet summarize or trim history automatically.

6. What this is not

Medeor is not a medical device, is not FDA-cleared, is not certified or accredited by any body, and is not a substitute for clinical judgment or for the standard of care. ViaVela claims no HIPAA, SOC 2, or other compliance certification. Medeor is currently not cleared for real patient data.

7. The clinician's responsibility

The licensed clinician remains responsible for the assessment, the plan, and every decision. AI output is one input among many and carries no authority. Nothing in Medeor should be relied on to rule out an emergency.

8. Contact

ViaVela Solutions LLC — legal@viavelahq.com.

---

© 2026 ViaVela Solutions LLC. All rights reserved.