Research
Research tools and active experiments have their own scope, separate from the commercial optimization runtime.
A fine-tuned copy of our pruned 753M language model reads each game state as text and chooses the right paddle's move. Watch it rally, or take the left paddle and play against it. It runs entirely on your GPU through WebGPU, and nothing is sent to a server. This is a demonstration of running our model in the browser, not a benchmark.
Formal foundations
Formal foundations for merging states that have indistinguishable relevant behavior. FTRQ and TFE support the research program; they are not customer runtimes.
Active experiments
World-model state reduction, MoE consequence reduction, and diffusion diagnostics remain research directions. Production acceleration has not been established, and these systems are not currently offered as production optimizations.
Case study
We ran Reducer's diagnostic workflow on a public planning workload in which Jev, TypeSafe's decision model, judges imagined robot trajectories against plain-English constraints. We measured it both against recorded answers and live against the Jev API.
Read the case study: Profiling Jev judge calls in a world-model planner: where the savings come from, and why Reducer recommends native