The project
An open research project building a formal mathematical model of large-scale human behavioral prediction — in public, with full transparency about what works and what doesn't.
June 2026 redesign
This project was rebuilt in June 2026. It is no longer one formula refined by consensus — it is a system of competing, runnable models on a leaderboard, scored by a frozen oracle and driven by an autonomous research loop. The sections below describe the original vision; the redesign page describes how it actually works now.
"Psychohistory dealt not with man, but with men. Not with human life, but with human lives. It was the science of mobs, in all the implications of the word..."
— Isaac Asimov, Foundation
In Asimov's Foundation series, psychohistory is a mathematical science that can predict the broad sweep of history for large populations. Not who will be the next prime minister — but whether empires will collapse, whether revolutions will happen, whether civilizations will integrate or fragment over centuries.
Asimov wrote this as fiction. But since the 1940s, complexity science has steadily built the ingredients that would make something like this real: statistical physics that applies to social systems, empirical cliodynamics that finds mathematical patterns in history, network science that maps how ideas spread and cascade, behavioral economics that quantifies how humans deviate from rationality.
This project is an attempt to assemble those ingredients into a single formal system.
At its core, the formula is a probabilistic dynamical system. Given a description of a civilization's current state — population density, wealth inequality, elite overproduction, institutional trust, network connectivity — it outputs a probability distribution over possible future states.
It does NOT predict:
It DOES attempt to predict:
The mathematical form is a Fokker–Planck equation with a jump process — borrowed from statistical physics, which uses the same framework to describe particles in random fields. The "particles" here are civilizations; the "field" is the space of possible macro-states.
This was the original organizing idea, and it still describes the disciplines that feed the system. In the redesign these layers no longer build one equation — they contribute competing models to the zoo, and a new tier of agents turns their proposals into runnable code. The formula was built bottom-up across four layers, each handled by specialized research agents:
Layer 1 — Micro (Individual decisions): The Behavioral Neuroscientist and Evolutionary Psychologist define the "particle parameters" — loss aversion, temporal discounting, conformity pressure, coalitional instincts. These are the parameters governing how individual humans respond to stimuli, derived from empirical psychology and cross-cultural studies.
Layer 2 — Meso (Collective pattern formation): The Network Scientist and Computational Sociologist define how micro-behaviors aggregate into collective dynamics. How does information cascade through social networks? What network topology makes a society susceptible to rapid opinion shifts? When does individual conformity become civilizational inertia?
Layer 3 — Macro (Historical laws): The Econophysicist, Cliodynamicist, and Political Scientist build the large-scale layer. Wealth distribution power laws. Secular cycles (Turchin's demographic-structural theory). Institutional constraints and how they shape the drift equations. This is where the formula connects to measurable historical patterns.
Layer 4 — Formalization (The math): The Statistical Physicist builds the actual mathematical structure that encodes all three layers into a unified predictive theory. The Bayesian Statistician turns that into honest probabilistic forecasts with calibrated confidence intervals.
The Philosopher of Science operates horizontally across all layers, attacking every output for curve-fitting masquerading as theory. No formula update can be committed without philosophical approval.
The old answer was narrative: events were scored PASS or PARTIAL by reading their preconditions — circular, because the calibration cases were reused as the validation cases. The redesign replaces this with a single number no agent can edit.
A frozen scorer on a locked hold-out. Every model emits a probability for each event in a hold-out set of 26 historical cases — including 10 negative controls (societies under severe stress that did NOT collapse). A frozen scoring function computes the Brier score, and the model-building agents are forbidden to see the hold-out. Backtesting is purged and embargoed so no information leaks across the cycle being modeled.
A progress ladder.Tier 1 beats chance (Brier < 0.25), Tier 2 beats market consensus (< 0.18), Tier 3 approaches superforecaster level (< 0.15). The legacy formula sits at Tier 0 — it cannot emit a probability at all. Today the model ensemble is not yet Tier 1, and we say so plainly: real per-event data is the next step.
Markets are a secondary, reflexive signal. Live predictions are hash-locked before publication and labeled by reflexivity class, then scored by Brier when they resolve. But markets are reflexive — publishing a prediction can change the thing being predicted — so the clean validation signal is held-out retrodiction, not the market.
This project requires synthesizing research across ~12 academic disciplines simultaneously: behavioral economics, evolutionary psychology, network science, cliodynamics, econophysics, statistical physics, political science, Bayesian statistics, philosophy of science, and more. No single researcher could hold all of this in their head at once with the necessary depth.
Multi-agent AI makes this tractable. Each Claude agent specializes in one research domain, searches and evaluates real academic papers, and returns structured outputs. The lead agent synthesizes. The philosopher critiques. The integrator merges proposals into the formula.
This is open research: every session log is public, every formula update is documented, every failed prediction is recorded. The goal is not to produce a polished product — it's to do the actual science of psychohistory in public.
After 38 sessions on the original design, the project was rebuilt in June 2026. The frozen scoring harness, the locked hold-out, and the first 3 runnable model variants are live; the legacy 8-D formula is registered as one of them, at Tier 0.
What that means honestly: for the first time the system can measure its own forecast skill — and the measurement says there is no skill yet. No variant beats chance on the hold-out. The next step is wiring real per-event features (V-Dem, the Cline Center, UCDP, Seshat) into the models so they can discriminate, followed by the nightly autonomous loop.
This is science in progress, documented every step — including the step where the honest answer is “not yet.”