$ ratchet --leaderboard

Each model plays every world eight times in a row. In Relay a fresh agent plays each life and sees only a 300-character note from the previous one; in Memory one agent keeps everything. The gap between them is the handoff loss. Scores are marks (0–5) averaged over lives 5–8.

cost × performance — hover a stick, click to open

Relay Memory Solo Oracle stick length = handoff loss

Costs are token counts from our runs × each provider's public list price (), not what we paid. DeepSeek runs on our own GPUs; its cost uses DeepSeek's standard API price.

ranking — click a header to sort · j/k to move

loss = Memory − Relay with a 95% bootstrap interval over worlds · share = loss / (Memory − Solo) · rate = the Memory life that equals Relay's eighth generation · goal = lives that light all five marks (Relay / Memory). † DeepSeek: worlds with all four conditions so far.

eight lives

memoryrelay compaction*record*

where the loss falls — loss by theme

Strong models lose on machines and crafts, where a rule ends a life without saying why. Weaker models lose on tools and trades, which stronger models find easy.