← Back to the chapter · Previous detail

Technical detail · IoT Machine Learning Pipeline

What each pipeline handoff proves

A passing handoff supports a limited claim. Read its evidence alongside what it does not prove, then use the retest trigger to decide which part of the release needs another review.

Handoff
What It Proves
What It Does Not Prove
Retest Trigger
Sensor to dataset
Readings, timestamps, units, context, missingness, and collection conditions are represented in training data.
That labels are correct, features are leak-free, or deployment inputs will stay similar.
Sensor, placement, firmware, sampling, gateway, storage schema, or collection condition changes.
Dataset to features
Cleaning, alignment, scaling, windowing, and feature code can be repeated for training and inference.
That the model generalizes, the split is valid, or feature meaning remains stable over time.
Feature code, time window, unit, normalization, missing-data rule, context field, or leakage check changes.
Features to model
The model artifact was trained and evaluated with a defined feature set, label target, split, baseline, threshold, and metric.
That deployment resource limits, data drift, feedback loops, or human response are adequate.
Model family, threshold, training set, label source, metric, class mix, or evaluation population changes.
Model to operations
The deployed system can run the artifact, handle unsupported inputs, monitor drift, capture feedback, and roll back.
That future data, devices, owners, alert policies, and retraining choices remain within the approved claim.
Deployment target, resource budget, alert policy, owner, feedback source, monitoring signal, or retraining gate changes.

← Back to the chapter · Previous detail