Forecasting
Predictive Runs
Short-horizon probabilistic forecasts (2–3s) that surface runs, overlaps, and pressing triggers before they happen.
What it is
If a representation truly understands movement, it should be able to continue it. Predictive runs is our research direction for short-horizon forecasting: given the last seconds of skeletal motion, roll the model forward two to three seconds and read out likely continuations — runs, overlaps, pressing triggers — as probabilities, not certainties.
This page is labeled Research on purpose. Nothing here is shipped, and forecasting quality is a claim we will make with evaluation results — or not at all.
How it works
The mechanism under investigation is generative rollout: the same foundation model that reconstructs masked movement can, in principle, be sampled forward in time. Today the simulation stack is an evaluation workbench — reconstruction, denoising, counterfactual probing — and forward rollouts are the research frontier built on top of it.
What it would unlock
| Parameter | Value | Notes |
|---|---|---|
| Horizon | 2–3 s | Short enough to learn, long enough to matter. |
| Output | distribution over continuations | Probabilities, never scripted predictions. |
| Reading | analyst-facing review | Decision support in review, not in-game automation. |
| Gate | evaluation results first | Ships when forecast quality survives the eval suite. |
Status & roadmap
Research. Short-horizon forecasting is a direction, not a feature: it builds on validated reconstruction capability and inherits every caveat of proof-of-concept data scale. Progress will be published through the evaluation docs first.