How this machine thinks about the future

The starting point is an approach called Predictive History, associated with the writer Jiang Xueqin. Its claim is that big events are not random. They come out of things you can actually look at: what the people involved want, what they are afraid of, what they can and cannot afford to do, and what happened the last few times a situation looked like this one.

So instead of asking “what will happen?”, the method asks a series of smaller questions that each have an answer you can argue about. Who are the players? What counts as winning for each of them? What is physically or economically impossible here, regardless of what anyone wants? Where has this pattern shown up before?

That approach has a real weakness, and this system is built around fixing it: it produces a story, not a number. A convincing story is not testable. You cannot be shown to have been wrong by one, and you cannot get better at telling them, because nothing keeps score.

So the qualitative reasoning is kept — it is genuinely good at finding the forces that matter — and four disciplines are bolted on that it does not have on its own: a starting estimate drawn from how often this kind of thing actually happens; a number attached to every judgement; a deliberate attempt to disprove the result; and a permanent record so that accuracy can be measured over years rather than asserted.

The same question is handed to six different analysts. They are not generalists asked for an opinion; each is given one specific question and nothing else to do.

• ActorWhat does each side regard as winning, and what limits them?
• Game theoryGiven those aims and the likely replies, which moves hold up?
• Historical analogyWhich past situations share the same machinery, and where do they differ?
• Belief and ideologyWhich beliefs change what the people involved are willing to do?
• StructuralWhich impersonal forces — economic, military, demographic — constrain this?
• ResilienceIf a shock lands, how well do the institutions absorb it?

Each one finishes by committing to a number: how likely it thinks the event is, and how much it trusts its own answer.

This is the part that does the real work, and it is easy to miss. No lens is ever shown what another lens said. Not the number, not the reasoning, not a summary. They each see the question and the evidence, and nothing else.

The reason is that agreement is only worth something if it was not arranged. If the second analyst can see that the first said 70 %, the second will land near 70 %, and so will the third. You end up with six numbers that agree — and the agreement tells you nothing at all, because five of them are echoes.

When they are kept apart, the spread becomes information. Six readings landing close together means the evidence genuinely points one way. Six readings scattered means the question is harder than any single one of them realised — and the system says so, by lowering its confidence.

The separation is not a rule the analysts are asked to follow. It is built into the structure: there is physically nowhere in what a lens receives for another lens’s answer to be.

Combining them is arithmetic, not judgement. No AI decides the final number — ordinary code does, by a formula that is the same every time and can be checked by hand.

Three things decide how much each reading counts:

• How sure it was. A lens that says “probably, but the evidence is thin” is weighed less than one that is confident.
• How good its evidence was. A reading built on official statistics outweighs one built on a news summary.
• Whether it is really a separate view. Two lenses that work from the same kind of material get a discount — they are closer to one reading than two.

There is also a starting estimate made before any of the six begin: how often this sort of thing happens in general, or what a betting market is already charging. It is deliberately boring, and it is there to stop a vivid argument dragging the answer a long way from what ordinary experience would suggest.

Once there is a number, a seventh analyst is given one job: show that it is wrong. It gets the result and the evidence, and it is asked for the strongest case against.

Two things make this more than theatre. First, it is shown the combined number but not the six behind it — otherwise it would simply attack whichever lens was the outlier, which is easy and pointless. Second, it has no power to change the answer. It can only raise objections; code decides what each objection is worth, by rules the analyst cannot see or argue with.

An objection that cites no evidence moves nothing, however forcefully it is written. And there is a ceiling: everything the skeptic establishes, put together, can shift the forecast only so far. The point is to catch a bad forecast, not to let the last voice in the room win.

Objections that were thrown out are shown to you anyway. A forecast where every challenge failed for citing nothing looks identical to one nobody challenged, and you deserve to be able to tell those apart.

A forecast of 30 % is not a prediction that something will not happen. It is a claim about a long run of similar statements: that if you collected every forecast this system put at 30 %, about three in ten of them would come true.

This is why no single forecast can be graded. The one that said 30 % and happened was not a failure. What can be graded is the whole record, over hundreds of questions — and that is what the system is built to be judged on.

You will also see a confidence figure, and it is a different thing entirely. It is not how likely the event is; it is how much weight to put on the number. It goes down when there was little evidence, when the six readings disagreed sharply, or when every reading came from the same underlying model.

Two deliberate refusals. The system never says something is certain, inevitable or impossible — those words claim more than any forecast can support. And a forecast at 50 % is labelled genuinely uncertain, never “likely”, because that is what it is.

Every forecast goes through the same eight steps, in the same order, and every step is written down as it happens.

Once published, a forecast cannot be edited — not the number, not the question it answered. That is enforced by the database itself, not by a convention someone could forget. If the question needs to change, you start a new one, and the original stays on the record exactly as written so it can still be scored.

The word “prophecy” in the name is a joke about the ambition, not a claim. It is worth being plain about the limits.

• It cannot see anything that is not already written down somewhere public. A decision made in private this morning is invisible to it.
• It cannot tell you what will happen. It can only tell you the odds it would give, and show you why.
• It has not yet earned the right to say its numbers are well calibrated. That claim needs a track record of resolved forecasts, and the system deliberately refuses to adjust its own numbers until it has one.
• It can be confidently wrong. Six lenses reading the same misleading evidence will agree with each other, and the confidence figure will look healthy. The skeptic and the evidence trail exist so that you can catch this — not so that it never happens.