game · 3 min
Context Cargo
Pilot a retrieval ship through a field of plausible-looking chunks and deliver a context window worth reading. Every chunk you grab adds mass and burns tokens; overfill the window and the oldest one gets silently truncated. You're competing against your own urge to collect everything.
Three queries, eleven floating chunks apiece. A few are exactly what the query needs; the rest are distractors from the same domain with the wrong intent — the sql join syntax doc that shows up for every Spark shuffle question. The cards don’t reveal which is which. Reading the label and thinking is the whole game.
Your ship steers toward your pointer, and it starts out nimble. Every chunk you collect adds mass — by six chunks you’re piloting a barge, which is roughly what your p99 latency feels like when you stuff the prompt. The context window bar fills as you hoard; overflow it and the oldest chunk is evicted with a quiet puff. Silently truncated, exactly the way real context windows lose things: no error, no apology.
Scoring pays for precision (+120 per relevant chunk), taxes recall-maxxing (−80 per distractor, −1 per 25 tokens carried), and tips +100 for a clean window. Cousin of Semantic Gravity and Chunk It Right — this one is about what you do when retrieval returns too much. Best total saved to localStorage. Single HTML file, no dependencies.