Every model and agent is registered, classified and watched.
Governance is not a review meeting. It is a registry, a risk class, an approval state and a monitoring record for every model and agent in production.
Nothing runs unregistered.
Registry entry
Every model and agent carries an identity, owner, purpose, data scope and declared output type.
Risk classification
Assigned by intended use and consequence of error; determines the evidence required before release.
Agent passport
For each MSI agent: task, permitted data, output type, evaluation record and approval state.
Version & lineage
Weights hash, training lineage and the evidence snapshot in force at release.
Performance & drift
Continuous monitoring against the validation population, with alerting on divergence.
Approval state
Released, restricted, or withdrawn — enforced at the serving layer, not by convention.
How a model is allowed to change.
Version-controlled data, code and configuration. No production data in development without governed de-identification.
Pre-registered thresholds. Performance reported with confidence intervals and subgroup stratification; calibration assessed, not only discrimination.
Approved change record, registry entry with weights hash, rollback tested before release.
Frozen versions at launch. Every prediction records the model version that produced it.
Input drift, output distribution, calibration and reviewer rejection rate, continuously.
Historical versions retained for the full retention period so past outputs remain reproducible.
Adaptive components need permission in advance.
Any component intended to adapt in production operates under an approved plan specifying the modification protocol, the performance envelope within which change needs no new authorisation, the monitoring that verifies it, and the automatic rollback that triggers on breach. Without that plan, a continuously learning device is not lawfully marketable.
Which is why the platform launches with frozen models and the change-control machinery built but not exercised. Building it later would mean re-validating every model already in clinical use.
Review our governance framework.
Control mappings, registry structure and evaluation methodology shared under NDA.