By 22 June, the repository's central preparation loop had taken shape. Imported material could be converted into extracted text, summarized into a deterministic Preparation Brief, divided into passages, and searched with a local lexical scorer. Retrieved passages retained their source identity so the interface could show where an answer came from.
The answer path was deliberately placed after retrieval. If a question produced no supporting passage, the app returned an insufficient-support message and did not call a model. If evidence existed, the production provider used Apple Foundation Models with instructions to stay inside the supplied passages and avoid invented citations.
This separation made the app useful before and around the language model. The brief, search, source coverage, and citations remained concrete local artifacts. Model availability could therefore be treated as a capability state rather than hidden behind a fabricated fallback.
Automated tests were added for Capsule isolation, deterministic ranking, evidence gating, citation preservation, unavailable-model messages, and the absence of a production fallback provider. Physical-device model behavior remains unverified in the current repository notes.
← Back to Cato