Data with defined uses
The release includes compressed executed tool trajectories and licensed WordNet-derived conversational lessons. They are not automatically loaded into every chat, inserted into user memories or used to claim a large language model was trained. The offline verifier checks digests and replays a sample; the exporter streams plan examples into a separate local training pipeline.
Keep learning mechanisms distinct
Private source recall, contextual word counts, trained intent recognition, saved executable workflows and neural weight updates solve different problems. The small bundled predictor remains distinct from an optional larger local language model. Shared contribution controls remain visible and retain the existing user preference.
Promotion needs evidence
An experimental language-model candidate lowered held-out loss but reduced audit top-choice accuracy, so its weights were rejected. Reports remain in the archive. A larger file is not proof of better general reasoning.