# ML and AI policy

This release does not claim to contain a trained classification or forecasting model. A device-identification interface must not return a category or confidence until there is a documented model behind it.

Every future model report must state its objective, data licence, training/validation/test split, preprocessing, baseline, evaluation metric, results, model version, and limitations. Low-confidence outputs must be communicated as uncertain and offer manual search. Models must run server-side when they require a secret or are too heavy for the browser.

The future 3D-learning analysis is a research opportunity, not a conclusion: outcomes can only be compared after sufficient privacy-conscious, anonymized event data exists.