Before you dive in
What you’ll find in the original
- The proposed design encodes shared state once, isolates question branches, and reads probabilities without a token-by-token generation loop.
- The cited probes found sibling questions could not see one another while all questions could see shared state, and latency stayed nearly flat up to roughly 100 questions.
- Choice options appeared to interact: adding irrelevant options shifted odds, and reversing option order caused small probability changes that matter near automation thresholds.
Worth knowing
This is a Japanese digest of a speculative black-box analysis, not TypeSafe architecture disclosure; the sparse mixture-of-experts backbone is explicitly inferred.