From a patient message
to a care-setting decision.
The current demonstration uses one model call. The engineering work makes its inputs, outputs, failures and evaluation boundaries visible.
Synthetic message
Select a sample or edit its text. Only the message enters the inference request; case IDs and physician labels stay out.
One hosted call
One Mastra step sends the fixed three-bucket instruction to the recorded Fable configuration. No automatic retry or retrieval step.
Strict output parsing
Accept one exact bucket and a nonempty short rationale. An invalid or incomplete response is an error, never a self-care decision.
Decision and trace
Display the result, latency and provider usage. Editing the input clears the old result. Credentials and accounting stay on the server.
Evaluation happens after generation.
The saved viewer compares recorded responses with a single physician’s post-output reference on 50 familiar synthetic messages. It makes no model calls. Agreement measures a care-setting match; it does not establish clinical safety, patient outcomes or deployment readiness.
Three output buckets
SELF_CAREASYNC_PHYSICIANURGENT_ESCALATION
Discuss which missing facts could change a decision and what a real clinical workflow would still need to do.
Compare failure modes
Open C22, C47 and C49. Compare missed clinician assessment, missed urgent escalation and incomplete output. Historical V25 is a separate, more complex experiment.
Class discussion presented by Brandon Dent, MD. This independent project does not imply university endorsement. Continue with the student guide.