Safety & limits
Why the clinician summary is deterministic
A summary you can trust because it's not interpreting anything.
Generative summaries fail silently. A model can paraphrase fine 99 times and confidently invent on the 100th. For a clinical record — even a patient-reported one — that's unacceptable.
Teloma's summary is built deterministically in services/summary.py from the structured record. It picks the right rows for the period, formats them, and restates them. There is no model call. The PDF is the same content rendered with reportlab. If you re-generate after editing a row, the summary changes the same way every time.
Note
When we add AI-assisted features (e.g. summarizing free-text notes), they will be additive — flagged as AI-generated, never used to interpret values, and never able to replace the deterministic summary.