How this works — and how you can check it
Everything on this page is verifiable: the model's parameters, calibration rules and backtest data are versioned, and every live number traces to a logged assessment.
The theory
Wars of aggression rarely arrive from nowhere. FORESHOCK models the convergence of four forces toward a war-permissive configuration: a strategic culture that frames expansion or restoration as destiny; regime insecurity that makes external assertion useful at home; leaders' core beliefs that license force; and a permissive structure — the practical opening (adversary dependence, perceived Western unwillingness to act, usable military advantage). When these converge, a discrete shock — a government falling, a formal commitment shift, a status breach — can tip leadership into perceptual collapse: the moment alternatives to force stop being seen. The Rubicon Index measures distance from that point of no return: falling = danger rising.
The model
The engine is a deliberately small, transparent causal network — seven variables with weighted influences, every parameter in an open configuration file. It is not an AI black box. Collapse requires two gates open simultaneously: convergence above a calibrated threshold and a sufficient active shock. A huge shock alone does nothing when the structure isn't permissive — a property the model must and did demonstrate (Taiwan 1996, below).
Where the AI fits — and where it doesn't
A language model works as an analyst assistant, not the model: it reads collected sources and scores the sub-indicators, with two disciplines. The evidence stream scores facts from corroborated reporting only. The narrative stream measures state propaganda as signal, not noise — what a regime wants believed is data about its intentions. Guardrails: per-cycle change caps, confidence floors, cross-bloc corroboration before any shock is injected, and a verbatim audit log — the Analyst log on the monitor — so you never take a number on faith.
Validation: calibrate on one case, test blind on the rest
The collapse thresholds and probability layer were fitted on Crimea 2014 only, using rules written down before fitting. Seven further cases were then run blind against the locked parameters — four where war came, three where it didn't:
| Case | Role | Outcome |
|---|---|---|
| Russia–Ukraine (Crimea) 2014 | CALIBRATION | WARNING 15 months before annexation; collapse fires Feb 2014 |
| Russia–Georgia 2008 | BLIND | WARNING 4 months before the August war; collapse reproduced |
| Russia–Ukraine 2022 | BLIND | Collapse fires January 2022 — one month before the invasion |
| China–Taiwan 1996 | BLIND — negative control | Stays quiet through the missile crisis: no collapse, no false alarm |
| Iraq–Kuwait 1990 | BLIND | WARNING July 1990, collapse fires before the 2 August invasion — a fast-onset case, and the first outside the post-Soviet space |
| India–Pakistan 2001–02 | BLIND — negative control | Stays quiet through two war scares (Parliament attack, Kaluchak): no collapse, no false alarm |
| Argentina–UK (Falklands) 1982 | BLIND | WARNING 3 months out; collapse fires March 1982 — the month the junta actually decided to invade |
| US–North Korea 2017–18 | BLIND — negative control | Stays quiet through "fire and fury": no collapse, and the index visibly recovers through the 2018 détente |
On the blind cases the model's probabilities beat both a base-rate-only benchmark and a simple statistical baseline (Brier score). The full backtest data, calibration rules and report are versioned in the project repository.
Robustness: did we just get lucky with the numbers?
A sensitivity analysis (versioned in the repository) perturbs every model parameter around its calibrated value and re-runs the full blind evaluation each time. The convergence gate is the parameter that matters most: it tolerates about ±0.03 before some validation criterion fails — set it lower and the negative controls start firing false alarms, higher and the probability layer stops beating naive baselines. The shock floor is far less sensitive (−0.15/+0.20). Across a broad grid of alternative gate pairs, 19% pass every criterion, and the calibrated pair sits in the centre of that passing band, not at its edge — the pre-registered calibration rule did not land on a knife-edge. The parameter-sensitive numbers are the short warning leads of the fast-onset cases (Georgia 2008, Kuwait 1990); Ukraine 2022's long lead is stable across every sweep. Robust to its parameters is not the same as right about the world — see below.
What we won't pretend
Honesty is the product, so the caveats are load-bearing:
— The model was built in hindsight on a small number of cases, and its
historical inputs are author judgments (sourced and documented, but judgments).
— The probability layer was revised once after failing a blind test (it
originally let raw shocks inflate risk without convergence — Taiwan 1996 caught it). The fix
aligned it with the theory, but that revision partially burned the blind set: the real
out-of-sample test is the live track record, which builds in public on the
monitor, Brier-scored, misses included.
— Backfilled history is shown as a labelled hindcast — assessed
retrospectively, with unavoidable knowledge of outcomes — rendered distinctly and never
scored into the track record.
— Probabilities are probabilities. A 10% event happens one time in ten. This is an
early-warning risk indicator, not a prediction machine, and nothing here is
investment or policy advice.