Trust

A senior consultant should not have to take the model on faith. Show the work.

LUMEN does not ask the consultant to trust a black box. Every clinical statement traces to a citation. Every drug dose traces to the formulary. Every guideline traces to the source publication with date. The disagreement between agents is visible. The attending signs at the end of every decision. The audit chain holds the entire reasoning chain end to end.

01

The model card

Kimi K2.6 base, fine-tuned on the hospital's own anonymized historical EMR plus formulary plus protocols. Evaluation methodology, evaluation set composition, test set size, and benchmark results published per fine-tune cycle. The card is the contract. If the metric drifts, the next fine-tune restores the floor before deployment to bedside.

1T
parameters, mixture-of-experts
per cycle
hallucination rate published
monthly
clinical advisory panel review
0
features ship without model card
02

Hallucination metrics

Per-category fabrication rate, measured against an evaluation harness with confidence intervals. Drug doses. Drug interactions. Clinical guidelines. Anatomy. Lab interpretation. Imaging interpretation. Each category has its own measured rate and its own evaluation set. The numbers are published in the model card per fine-tune cycle. Drift triggers re-evaluation before bedside deployment. Hospital admin sees the metric continuously; clinical staff sees it at deployment.

03

Citation discipline

Every clinical statement traces to a citation. Every drug dose traces to the formulary entry. Every guideline traces to the source publication with date. Every score traces to the validation cohort. The citation is one click away in the side drawer; the source publication opens in the same drawer, never in a new tab, never via a third-party redirect. No clinical statement reaches the chart without a citation behind it.

04

The disagreement surface

When agents in Counsel disagree, the disagreement is visible. The attending sees the argument before the synthesis assembles. Pharmacology and Infectious Diseases on the meropenem dose: the senior consultant reads both arguments, both citations, both cohort data, then decides. The model never overrides the attending. The synthesis is a human-confirmed output, never an autonomous answer.

05

The countersign

No clinical decision is final without a human countersignature. The attending signs. The countersignature is recorded on the audit chain. The audit chain is on the hospital's own server. The signature is the contract between the model and the chart: nothing is final without a human commitment.

06

Continuous validation

Every month, a sample of decisions is reviewed by the clinical advisory panel against ground truth. The hallucination rate is published to hospital admin per category. Drift is caught early. The validation methodology is co-designed with the hospital's quality and safety committee. The model is never static; the validation is never optional.

07

FDA and regulator alignment

FDA Software as a Medical Device (SaMD) pathway alignment at the architecture layer. Specific clearance pathway per use case. EU MDR Class IIa CE marking pathway, subject to EU regulatory counsel confirmation. Saudi MoH alignment per ARDIN framework. Canadian Health Canada SaMD pathway. The compliance posture is aligned at the architecture layer, not retro-fitted to a settings page.

Trust comes from showing the work, not asking for faith.

Model card, citations, countersign, audit chain. End to end.

Talk to procurement