ML monitoring

Open the model at /ml/{id} after the first call, or open Usage (/usage) → Data Intelligence for the account.

Each POST /predict/ records one inference. A batch call records one inference per row that succeeded. The channel is Heimdall UI when the call comes from the app, and API otherwise.

Drift is a percentage stored with the call when that model has drift_stats from training. Models without those stats record 0. There is no separate switch to turn drift on.

The request log fields the service stores are endpoint, inference count, response time, channel, user agent, and drift percentage. Auth failures return 404 before a successful inference is counted.

Account-wide charts, including Read, Vision, and Resolve, are on Monitoring and usage.