World model & causal reasoning — narrative

Verified structures and timers: world-model-and-causal.html.

The purpose

Between "what is happening" (live data) and "what was computed" (engines, kernels) sits the question the platform actually promises researchers: what happens next, and why? The world-model subsystem records causal structure (SQLite-backed causal graph with embedded entities and edges; a temporal KG v2 whose edges decay on a 72-hour time constant; write-through of chat interactions as graph edges), fuses forecasts from five weighted sources (physics, statistical, pattern, causal, LLM) with per-domain weight sets, and — critically — keeps a ledger: predictions are scored when they resolve (Brier, calibration) on a six-hour sweep.

Prediction hygiene over prediction marketing

The subsystem's most important property is the scoring loop: forecasts are persisted as falsifiable claims with horizons, and the validation report exists so that accuracy is measured, never asserted. The physics layer feeding it is explicit hand-coded relations (Gutenberg–Richter, Omori, vapour-pressure, Rothermel) with their constants in source, not hidden weights. Causal discovery through the optional DoWhy microservice is an external analysis path, off unless deployed.

What to avoid reading into it

"World model" here denotes a structured state store plus forecast/counterfactual machinery — not a learned simulator of the Earth. The earthgen latent module and dream/fork engines explore synthetic scenario space; they are research scaffolding with honest headers (one literally states the entropy mixer "is not a neural net") and their outputs are clearly separated from the kernels' physics in every UI surface.