Meana Petro

Ciclo de vida del modelo

Modo sombra, puertas de validación, monitoreo de deriva y reentrenamiento.

01

Shadow mode

Every model runs against live data, recording predictions and changing nothing, for one to three months before anyone acts on it. Results are scored on the real pipeline with the real sensor failures included, not on a retrospectively cleaned dataset.

  • Validation gate and target metric written before the shadow period starts
  • Predictions logged with the model version, input snapshot and configuration
  • Scored against the outcome that actually happened, not against a backtest
  • A model that misses its gate does not ship, and the gate is not renegotiated afterwards
02

Registry and reproducibility

Model versions are immutable. Every prediction stored in the platform can be reconstructed eighteen months later with the same weights, the same inputs and the same configuration — which is what an audit or an incident investigation actually requires.

  • Immutable model versions with the training dataset snapshot recorded
  • Every prediction carries the version that produced it
  • Promotion between environments is an approval, logged with a name against it
  • Rollback to any previous version without a rebuild
03

Drift monitoring

Input distributions and prediction quality are monitored continuously. Drift is expected — reservoir pressure declines, instruments get recalibrated, operating practice changes — and the point is to detect it before the field does.

  • Input distribution monitoring per feature with alerting thresholds
  • Prediction quality tracked against outcomes where labels arrive
  • Retraining triggered by measured drift rather than by a calendar
  • Degradation is reported to the operator rather than quietly retrained away
04

Running this air-gapped

The whole lifecycle runs inside the operator’s network. Training, registry, deployment and monitoring have no external dependency, and updates arrive as signed offline bundles verified against a local key.

  • No external model registry, artifact store or package fetch at runtime
  • Signed offline update bundles, verified locally
  • Diagnostics exportable as a single file for offline support
  • Capacity is sized up front, because it cannot be added elastically