Autocomplete Tennis

How Autocomplete Tennis Works

You are rallying word-associations against a fully self-contained "language model" — a weighted graph walk that lives entirely in game.js. No network, no API, no GPUs were harmed (it deeply wishes they had been). Each turn you must reply with a word genuinely associated with the previous one, checked against the bundled association graph, before the volley timer runs out.

1. The model serves a word (it always opens near its beloved compute cluster).
2. Reply with a valid association: tap a suggested chip, press 1-6, or type any word (fuzzy-matched to the bank, typos forgiven).
3. The model volleys back its own association — and it's deterministically biased away from your secret target.
4. Each rally hands you a secret target word. Land on it (or anything one hop away) for +3 bonus while the model drifts elsewhere.
5. Whiff — no valid association, a repeat, or too slow — and the run ends. Corner the model into a dead end and the rally point is yours.

Scoring

Every completed volley is a point. Reaching the secret target is +3. We track your longest rally and targets reached, and persist your best run locally. No repeats per rally — the model refuses to enable your mode collapse.