Entities

Model prediction

A record of a model output — suggested match, predicted category, anomaly score, generated narrative — with the model version, inputs and human decision.

Volume

Thousands to millions per tenant per period

Retention

Audit period

Entity attributes9

Model retraining must be scheduled and versioned, with the ability to identify which model version produced a given prediction.
Tolerance models must be trained per tenant. One tenant's variance distribution cannot inform another's.
Agentic investigation must run read-only tools freely but gate every write action behind human approval, logging each step taken.
Model retraining must be scheduled and versioned, with the ability to identify which model version produced a given prediction.
model_prediction_id
target_ref
prediction / confidence
Model retraining must be scheduled and versioned, with the ability to identify which model version produced a given prediction.
Machine-generated narrative must carry provenance metadata and require human approval before it becomes part of the record.

Entity relationships1

Exception

Capability entities3

Entity lifecycle

Not yet researched

Only the control-bearing entities were modelled for lifecycle. Defensible as a starting point, but it does leave the entity model uneven, and this is incompleteness rather than a finding.