The trigger article is ingested and its source graded for reliability — a grading that weights every downstream piece of evidence. Actors and countries are resolved against structured registries, and the question the article raises is stamped against a reference class: a counted set of comparable historical cases drawn from the field’s canonical datasets, with a reproducible derivation script and an audit trail.
Several independent analytical passes read the article: the surface narrative and what it obscures, the actors and the constraints binding them, the scenario space, the power and economic transmission channels. Each pass emits typed, structured claims — graded counted / sourced / inference — and none of them is allowed to set the headline number.
The probability starts at the reference class’s counted base rate. Article evidence is weighed in log-odds with source-reliability multipliers, then deliberately dampened against model overconfidence. The scenario partition is reconciled against the calibrated posterior, and where the analyst layer disagrees with the calibrated number, the disagreement is published beside it as a registered position — scored at resolution, never blended in.
A falsifiability register assembles the case against the analysis — counter-evidence from the counted record, the analytical inversion, dated observables that would prove it wrong. An adversarial peer review scores the run and its warnings are published, not scrubbed. A compliance screen enforces analytical-not-prescriptive language. The record is then written once, append-only, with pre-registered resolution criteria — and scored publicly against outcomes when its horizon arrives.