# Operational scheduling amendment

2026-09-20, before any scored world beyond seeds 17001 and 17002.

The initial runner queued two complete matched worlds at a time with at most eight concurrent requests. After the faster endpoints completed those worlds, six worker slots stood idle while Moonshot AI continued processing memory requests (roughly 1–2.5 minutes per request). This is a transport scheduling bottleneck.

The runner now queues up to eight worlds per cohort, keeping **the same maximum of eight concurrent API requests**. Each trajectory still receives the identical prompts, notebook budget, request deadline, output cap, nominal reasoning effort and event order. No generator, answer key, scorer, bootstrap method, planned seed allocation or dollar ceiling changes. We retain all first-cohort observations. The initial runner and its pre-inference hash remain in the release; the updated runner has its own hash.

This amendment was made for elapsed runtime, not to respond to accuracy or select favorable results. It is an operational deviation from the initial implementation and is disclosed rather than represented as part of the original freeze. More concurrent calls to an endpoint may affect latency, so the release makes no comparative latency claim. Capacity comparisons are still exploratory. The budget stopping rule and requirement for completed matched worlds remain in force.