Automated actions are gated on model confidence or plausibility

rel-python-ml-003

Intent

Prevent raw model mistakes from directly causing costly or unsafe automated side effects.

Applicability

Applies when model outputs trigger money movement, user state changes, or other consequential automated actions.

What to inspect

Confidence thresholds, plausibility checks, and human or manual fallback paths.

Pass criteria

Consequential automated actions are gated by confidence or plausibility checks and have a fallback path when the prediction is uncertain.

Fail criteria

The diff executes consequential actions directly from raw model output with no visible confidence or plausibility gate.

Do not flag

Pure analytics, offline scoring, and advisory-only predictions.

Confidence guidance

HIGH when the side effect and missing gate are directly visible. MEDIUM when confidence may be enforced elsewhere. LOW when action criticality is unclear.

Remediation

Add thresholding, reasonableness checks, or human review before irreversible or sensitive automated actions.

Pass example

if score >= 0.95:
    approve()

Fail example

approve()

Sources

  • Machine Learning Engineering — Andriy Burkov book