We Rebuilt Jev's API on an Open Model and Used It to Play Doom

Stephen Blum explains Blocks.ai's open Jev-compatible classifier: Gemma logits over lettered options, typed response construction, prefix caching, test cases, and the deterministic navigation code around its Doom agent.

Original by Stephen BlumIntegrationsBlocks.ai blogSource reviewed 2026-09-25

Before you dive in

What you’ll find in the original

The replica performs one forward pass, reads next-token logits for allowed letters, applies softmax, and maps those probabilities back to caller-defined options without a generation loop.
Keep deterministic work outside the model: the Doom implementation uses engine data and BFS for navigation while the model chooses a local action within the current phase.
Cache the static few-shot prefix only when it matches exactly and size the cache for distinct questions; warm caching cut the reported single-decision path from roughly one second to about 127–173 ms.

Worth knowing

This reproduces Jev's API shape, not TypeSafe's training method, architecture, calibration, or RLCD claims. Results come from small authored test sets and one M4 Mac; the full Doom demo ran slower than real time.