LLM Judgment Layer

Seasoning, not the meal. The app is 100% deterministic and works offline; the model answers a few narrow questions per time-bucket and code keeps control of everything else.

HARD-GATE

• The model never decides policy. It ranks, scores, or answers yes/no; code sorts, clamps, thresholds, and chooses.
• The UI never waits on the network. Render the cached/default plan first; update in place if the model answers.
• Every failure path lands silently on the deterministic fallback. No error copy, no retry storms, no dead UI.
• The API key never ships in the client. Server-side proxy adds it; unset proxy URL means the layer is inert by design.

Core loop