Weather prediction worked because the atmosphere has conservation laws, weak coupling at the forecast scale, and indifference to the forecast. Society has partial, local, conditional analogues of each — and the engine reports its own scope as a first-class output.
This site is split into a guided tutorial, a worked-out account of the mathematics, and the full empirical record of every test run against real data. Pick a door.
Psychohistory from scratch: the weather analogy, the three structural conditions, the eight mechanism layers, and one cascade (GameStop, Jan 2021) walked down every layer.
Every equation explained in plain language and rendered properly: conserved-measure transport, near-decomposability and Neff, Kuramoto criticality, mean-field-game fixed points, the skill horizon, and assimilation. Each tied to a slider you can move.
The empirical program, run against Reddit, Wikipedia and GitHub. The structural gear carries; all four powered dynamical claims deflate — the split the theory predicts. Every verdict with its real numbers and figure.
The program's first module about intervention. One new axiom (Ashby: only variety can absorb variety), two premises the program already holds, and out of them a survival theorem: societies survive technology in proportion to the control they exert over its speed. From the first campfire to the Bronze Age pileup to the IAEA to frontier AI — one hazard law, six historical end-states, tested on a century of public data, with an instrument whose bottom band is hours-for-dinner, per tribe.
The full manuscript, 79 pages: the framework, the modern instantiation as a distributed Multivac, the governance trilemma, and the complete first empirical contact. The site is the paper made interactive.
Numerical weather prediction is the existence proof that a system of comparable apparent difficulty can be forecast — not because of compute, but because of three conditions. The conditions, not the compute, are the theory.
Mass, energy, momentum are conserved; trajectories are confined to low-dimensional manifolds. Social analogue: attention is a conserved normalized measure — allocated, never created. Belief is a drift field riding on top, not a stock.
Ensemble statistics over quasi-independent regions are meaningful. Social analogue: Simon near-decomposability — dense within communities, sparse between. The statistical unit is the block; Neff ≈ K, not the population.
The atmosphere does not read the forecast. Social analogue: reflexivity is dissolved at mean-field-game fixed points — a forecast is publishable iff it reproduces itself once everyone acts on it.
Attention concentrated in days; the forecast of failure caused the failure; the run propagated within one professional block before jumping; a single deposit-guarantee announcement at peak sensitivity halted the cascade. Every component appears in one week.
| Objection | Repair | Mechanism |
|---|---|---|
| No conservation — belief appears freely | Attention conserved; belief = drift | Total media time is ~flat; composition churns. Continuity equation ∂ρ/∂t + ∇·J = s; belief is a direction on transport, not a sub-measure. ▸ the math |
| Reflexivity — the forecast is an input | Dissolves at MFG fixed points | Lasry–Lions equilibrium ρ* reproduces itself under best-response. Publishable set = equilibrium set; deposit insurance is the engineered fixed point. ▸ the math |
| Correlated agents — small effective N | Simon blocks, Neff ≈ K | Near-decomposable hierarchies; model communities, not persons. Neff read from the macro variance-ratio, never Pearson-of-fluctuations. ▸ the math |
The framework is a literal, bounded reading of the Foundation machinery. Each canonical concept maps to a piece of mean-field / control theory.
| Asimov concept | Real-math analogue | Lives in |
|---|
Seldon inverts secrecy: he published the Plan's existence (the fixed-point components) and hid its contents. Information release is itself a control input. ▸ the full tutorial