The world-state predictor,
with its own scorecard

Every sense, and how well it actually does

Every sample is tagged by source — a real device's numbers are never averaged into the simulated ones. Read straight from /api/predict/metrics — the same board the engine writes to itself. “Beats baseline” is the margin over the naive alternative; negative means the naive alternative is still winning, and it stays on the page.

Drift, measured

Short horizons are strong and long horizons drift — that is true of every world-model, neural or not. So it gets measured on live traffic at several horizons at once, rather than quietly avoided.

One coordinate system for every sense

Each sense's state encodes into one 4D space and snaps to the nearest of 48 computed lattice points (the 24-cell and its dual). That shared alphabet is what lets separate senses be compared, and gives every state a stable canonical code. A learned transition chain rides on top of it — and the drift table above shows where the snap pays and where it does not.

Three rails

Every number above is one curl away

# loading…

The whole package in one call: /api/predict/world. The public fleet is a high-fidelity behavioural simulation; real Blackbox units feed the identical pipeline.