Before you dive in
What you’ll find in the original
- The direct-readout hypothesis is supported by TypeSafe’s description and by latency that did not scale with serialized “output token” counts or option count.
- Moving a secret from a sibling question into shared state changed whether another question could recover it, supporting question isolation with shared-state visibility.
- Latency and request limits fit a shared-prefix, batched-branch design; a sparse mixture-of-experts backbone remains the least certain inference.
Worth knowing
This is a careful but speculative black-box reconstruction, not disclosed architecture. The author distinguishes published, observed, and inferred evidence, and the tests do not uniquely identify an implementation.