Predict a cache before the token moves. A wrong route stays visible as a miss and can be corrected.
Teaching scope: the requests and outcomes are illustrative. The game makes no universal latency, cost, GPU, or provider claim; production cache behavior depends on exact infrastructure, model, invalidation, and privacy constraints.
HIT / REUSE VERIFIED WORK
MISS / CONTINUE TO GENERATION
This local routing game presents four fixed reuse patterns: exact prompt, shared prefix, semantic paraphrase, and deterministic tool input. Correct answers advance through four requests; wrong answers stay recorded and can be corrected. Bars show four mastered patterns versus zero at one hundred twenty pixels per pattern. The fixture demonstrates categories, not a running cache, model call or similarity detector. One deliberately wrong answer followed by the four correct routes produces five attempts and four mastered requests. Both bars show counts at one hundred pixels per event. The wrong choice contributes one recorded miss and does not advance the request index; a correction does. Complete exports retain each chosen and expected cache plus its evidence. Counts are quiz outcomes, not measured cache hit rates. The animated correct route goes request to its chosen cache and then result, using three path points. A wrong route adds the generation node between cache and result, using four points. Bars show three versus four vertices at one hundred twenty pixels per vertex. These are the actual authored path constructions, not network hops or elapsed latency. Freshness and invalidation remain separate deployment requirements.