Most analysis of world events runs on stated intentions. Governments announce why they did the thing, companies issue a press release, and the story gets written down as though the announcement were evidence.
Stated intentions are the least reliable data available. Revealed interests are better. Who benefits, who pays, who was already moving before the news broke, and who is working hardest to make sure you understand it a particular way.
That question, cui bono, is old and it is not mine. What is mine is the machinery around it, because the question on its own is dangerous. Run it loosely and you can prove anything. Somebody always benefits from everything, and if you follow the money far enough you will always find a connection. That is how people end up on YouTube with a whiteboard and string.
So the engine is built mostly out of brakes. Here is how it works, and here is exactly where you should distrust it.
The problem the design is solving
If you go looking for who benefited from an event, you will find someone. If you then expand outward to who funds them, who owns them, who they went to school with, you will find more. The network never runs out. Every step feels like progress and none of it is.
The engine’s answer is a stop rule with three parts, applied before anyone is allowed into the analysis.
Does it matter at their scale? An outcome has to move something the actor actually cares about: a meaningful chunk of their revenue, their power, their security, their survival odds. “BlackRock owns 7% of a company involved” fails this immediately. At BlackRock’s size that event is a rounding error, and a rounding error is not a motive.
Would looking there actually tell us anything? Before adding an actor, we have to say what evidence that would surface and which competing explanations it would help us tell apart. If the honest answer is “whatever we find will fit every theory equally well”, we do not look. This is the test that kills most conspiracy reasoning, because most conspiracy reasoning produces facts that are consistent with everything.
A hard budget on how far out we go. We start with the people directly involved and expand outward in rings: funders and owners, then competitors, then regulators and states. Going past that, out to media amplifiers and ideological allies, requires a written reason logged in the report. It is a decision, never a drift.
You will see all three of these applied by name in the working of every report, including the actors we considered and threw out.
The seven stages
1. What actually happened. A timeline, with every item marked primary or derivative. Primary means the actor’s own document, a direct statement, the original data. Derivative means reporting on reporting. Most public argument runs entirely on derivative sources, which is why most public argument is worthless.
2. Who is in this. The actor map, built with the stop rule above.
3. Six questions, asked of every actor. Who benefits, and in what currency: money, power, territory, information, attention, law? Who pays? Who moved early, before the event, through trades or hires or legislation or troop movements or PR seeding? Who is amplifying the story, and who is quietly ignoring it? What does each actor want you to believe? And last, always last, always asked: what is the cheapest explanation? Incompetence, coincidence, or things just drifting.
4. Rival explanations, with starting odds. We write down three to six competing stories about what caused the event. Two are compulsory. One is the boring explanation: nobody planned anything, it was a cock-up or a coincidence or structural forces grinding along. The other is the official version, which enters as one contender among several rather than as the frame everything else has to argue against.
Then each gets starting odds, before we look at any of the case-specific evidence. Those odds have to come from a reference class: in events of this type, how often does this kind of explanation turn out to be the right one? “Feels like 70%” is rejected. Historically incompetence and drift beat coordination, so the boring explanation usually starts in front.
5. Weighing the evidence. Every fact gets scored against every explanation. Details below, because this is the part that does the work.
6. What happens next. For explanations still standing, we run the actor map forward one turn: who moves next, what is their best play, what breaks second. Scenarios get odds bands, never a single number.
7. The report. Plain English on top, the full working underneath.
The scoring method in stage 5 is not mine either. It is a variation on Analysis of Competing Hypotheses, developed by Richards Heuer at the CIA and published in Psychology of Intelligence Analysis in 1999. Heuer’s insight was that analysts do not fail because they lack information, they fail because they settle on an explanation early and then read everything as confirmation. His fix was to force every piece of evidence to be scored against every rival explanation at once. The change here is that we score with numbers rather than ticks, and the arithmetic is done by a script rather than in someone’s head.
How a fact earns its weight
For every piece of evidence we ask one question: how expected is this fact if the explanation is true, compared with how expected it is if the explanation is false?
That ratio is all a piece of evidence is worth. It gets one of five settings, and only five.
| Weight | What it means |
|---|---|
| 1 | Useless. Fits every explanation equally. Moves nothing. |
| 2 | Weak. A mild lean. |
| 4 | Moderate. Real, not decisive. |
| 10 | Strong. Hard to explain if this story is false. |
| 30 | Near decisive. Barely reconcilable with the alternatives. |
Evidence that argues against an explanation gets the same numbers upside down. Five settings, no in-between, because the difference between a 6 and a 7 is a fantasy and pretending otherwise is how false confidence gets built.
Those weights multiply against the starting odds and produce final odds for each explanation. A script does that arithmetic, not a person and not a language model, because the entire value of the exercise is that the numbers you put in determine the answer whether you like the answer or not.
Four rules govern the scoring, and they do more work than anything else on this page.
Benefiting is not evidence. This is the trap the cui bono question sets for people who ask it casually. Defence contractors profit from a war, so a war must have been arranged for defence contractors. But defence contractors profit from a war that started by accident too. The fact is equally expected either way, so its weight is 1, and it moves nothing. Benefit plus documented prior positioning is a different matter, and that is what earns a real score.
Ten reports are one report. If forty outlets carry a story that traces back to one anonymous source, that is one piece of evidence, not forty. This is the quietest way to destroy an analysis, because volume feels like corroboration. Every report on this site shows the merge explicitly: you can see which stories were collapsed into one and why.
The boring explanation gets a head start. It has the highest starting odds most of the time, and it has to be beaten with real evidence rather than removed for being unsatisfying. Being boring is not a strike against an explanation.
No false precision. Odds are given as ranges. Nothing here is calculated to a decimal place, and if it looks like it is, distrust it.
If you want the technical terms
Nothing on this page needs them, but the working pages behind each report use the formal language, so here is the translation.
| What I call it here | What it is called in the literature |
|---|---|
| Starting odds | Prior probability |
| Evidence weight | Likelihood ratio |
| Final odds | Posterior probability |
| The boring explanation | The null hypothesis |
| Scoring facts against every rival at once | Analysis of Competing Hypotheses |
Betting in advance
Here is the part that separates this from opinion.
At the end of every report we name specific things that might happen next, and we commit in advance to what each one would be worth. If this document surfaces by March, that explanation’s odds get multiplied by four. If the meeting does not happen by June, this one gets halved.
Committing in advance is the whole point. Once you know how something turned out it is trivially easy to explain why it was always the likely outcome, and everyone does it, and nobody notices themselves doing it. Writing the weight down beforehand makes the update mechanical. When the thing resolves, the number gets applied whether or not it flatters the call we made.
Every one of those bets lands on a public scorecard along with what actually happened. Over enough of them it will show whether the odds coming out of this thing mean anything at all. If it turns out that things I called strong evidence pan out at the rate of weak evidence, that will be visible to you, on the site, and the weights get adjusted.
Analysis without a scorecard is opinion with extra steps.
Who does what
The research is AI-assisted and I am not coy about it. An AI agent runs the pipeline: builds the timeline, chases the sourcing, proposes the actor map, drafts the rival explanations, suggests the reference classes and the starting odds, and proposes a weight for each piece of evidence with its reasoning attached.
I sign off on the explanations, the starting odds and the weights, and I decide what gets published. The AI proposes, I dispose, and a script does the arithmetic so that neither of us can quietly round the answer toward the one we prefer.
The pipeline itself is a folder of plain text files, not a product. Any capable model can run it, which is deliberate: I do not want the method welded to whichever AI company is winning this year.
What this is not
It is not journalism. I do not have sources, I do not make calls, and everything here is built from material anyone can look up. The reports show you exactly what was used.
It is not proof of anything. It is a disciplined way of holding several explanations at once and letting evidence move the odds between them, which is a different and more modest thing than knowing what happened.
And it can be wrong in a way you can catch. Bad starting odds produce confident bad answers, and the machinery will not save you from them. So every report shows its starting odds, its reference classes, its evidence with sourcing flags, and the weight assigned to each item with the reasoning. If you think a number is wrong, you can find it and say which one and why.
That is the deal. You get the conclusion in plain English, and if you do not believe it, you get the working.