Most analysis of world events starts with stated intentions. Governments announce why they did something, companies issue a press release, and the explanation gets repeated as though the statement were evidence.
Stated intentions are evidence, but they are not privileged evidence. Revealed interests, prior movement, institutional constraints and the sequence of decisions may tell a different story.
The old Roman question is cui bono? Who benefits?
Asked loosely, it is dangerous. Somebody benefits from almost everything. Follow enough funding, ownership and personal connections and you can build an impressive network around any conclusion you started with.
The engine is a set of brakes designed to stop that happening.
What the process does
Every analysis follows seven stages.
1. Establish what happened. Build a timeline and attach every factual claim to its source. Primary means the actor’s own document, direct statement or original data. It describes provenance, not truth. Derivative reporting is retained and labelled.
2. Map the actors. Start with the people and institutions directly involved, then expand through owners, funders, customers, rivals and regulators. Media and ideological networks come later, if at all.
Every expansion must pass two tests. The event has to matter at that actor’s scale, and researching the connection has to offer a reasonable chance of distinguishing between explanations. If the same connection fits every possible story, it adds atmosphere rather than information.
3. State the interests and mechanisms. Who benefits, who pays, who moved early, who shaped the available options, and what would we expect to observe if an alleged mechanism were actually operating? Benefit alone is not evidence of causation.
4. Define the questions and explanations. Every set of probabilities must answer one question. If explanations can coexist, they are represented as combined states or tested in separate questions. Overlapping motives are never forced into one probability pie.
Starting probabilities are labelled either empirical or judgmental. An empirical prior needs a real sample, numerator, denominator and classification rule. A few historical examples recalled from memory are not a measured base rate. Judgmental priors are allowed, but the report must show how the answer changes under different starting views.
5. Weigh the evidence. Related claims are grouped before scoring. Forty articles repeating one anonymous source are one chain, not forty confirmations. Independent sources can still be dependent if they all report the consequences of one underlying event.
For each evidence group, the question is: how expected is this observation under one explanation compared with the same observation under a stated reference explanation, given everything already counted?
The calculator combines those declared judgments. It does not decide whether the judgments are sensible.
6. Separate causal analysis from forecasting. A causal explanation can be plausible without producing a unique future event, and a forecast can come true under several different motives. The engine therefore keeps them apart.
Forecasts are registered before the outcome with a probability, deadline, resolution rule, sources and a fixed baseline. Later they are scored with a proper scoring rule. An unresolved event stays unresolved. It does not quietly become false because the deadline passed without a public document.
7. Publish the conclusion and the working. A report can finish with mixed causes, an unstable ranking or insufficient evidence. The method does not require a winner.
How the numbers work
Suppose a report compares explanations A, B and C. One becomes the common reference. For each grouped observation, the analyst states how likely that same observation is under A and B relative to C.
A factor of 4 means the observation is judged four times as expected under that explanation as under the reference. A factor of 0.25 means it is one-quarter as expected. A factor of 1 means the observation provides no update.
The factors modify the starting odds. A tested script performs the arithmetic and retains the whole distribution.
The arithmetic can be correct while the result is wrong. The question may be badly framed. The explanations may overlap. A source may be unreliable. An evidence factor may be a poor judgment. That is why each report includes the exact questions, dependency grouping, reasoning and sensitivity tests.
What sensitivity means
The result must be rerun under at least:
- a different set of starting probabilities;
- weaker evidence weights;
- removal of each load-bearing evidence group.
If the leader changes, that belongs in the report. If a conclusion holds only when one analyst-supplied number is treated as strong, that belongs in the report too.
The resulting range is not a confidence interval. It shows how the model behaves under named alternative assumptions.
What gets measured
Historical replays test whether the process is coherent and whether old conclusions were artifacts of the model. They do not demonstrate forecasting skill.
Forecast performance is measured only on questions registered before the outcome. The model is compared with a fixed baseline on the same resolved cases using Brier scores. Related forecasts are grouped so that five versions of one event do not masquerade as five independent successes.
There is not yet enough prospective evidence to claim that the engine is calibrated or better than its baseline. Reproducible arithmetic is a starting condition, not validation.
Who does what
The research and first-pass analysis are AI-assisted. An AI agent builds timelines, follows sources, proposes actor maps, drafts hypotheses, identifies dependencies and suggests numerical judgments.
I review the framing, sources, priors, likelihoods and publication. The calculator determines what those declared inputs imply. It cannot convert bad inputs into a good conclusion.
The original source record, model inputs, outputs and correction history are retained so that a reader can identify the disputed judgment rather than merely disagree with the headline.
What this is not
It is not journalism. I do not have private sources and I do not conduct interviews for these reports. The analysis uses publicly available material.
It is not proof of motive. Public evidence often cannot establish a private counterfactual such as whether a leader would have made the same choice without one particular incentive.
It is not a machine for finding conspiracies. Coordination is one possible mechanism among several. Institutional drift, contingency, error and mixed motives remain live wherever the evidence allows them.
The aim is modest: make the question precise, keep compatible causes compatible, expose the important judgments, and make it possible to correct the result without rewriting history.