The player's placement and goal are the inputs; the task cards, the cost model and the milestone cutter are shared; each structure decides what the deterministic code keeps as its primary object and what the AI may change. Press the events on a structure's tab to see how it behaves when a player races ahead, struggles with a warm-up, fails a check or skips two days. Real data: the catalog's levels and finish rates, Für Elise's 19 scanned phrases with their per-hand features, and the harmony the Orchestra Conductor's analyzer reads off each score.
The cost model is four numbers (a hand alone is 70% of a phrase's load; a check at target T costs 1 + ⌈(T − S)/3⌉ sessions; a clean pass lifts S by 60% of the gap; a cut fits when its drill days are at least 0.8 of the horizon) and is the first thing to fit against real phrase_plays. The "AI's pick" of a bridge piece is a stand-in: the code hands the model the candidate band and the player's taste, and the model chooses. Nothing on this page touches the game. Recipe: node temp/pw/shot-plan-lab.js.