Third Friday of every month, beginning 18 September.

Every desk has a story about a flow they saw coming.

The index rebalance you positioned for a week early. The month-end hedging programme that arrives like a commuter train. The fund that has to buy into a rising market at the close because its own documents give it no choice in the matter. The trend follower you know is a few points from turning, because you worked out roughly where its trigger sits and you were patient.

None of that flow has a view. It is not expressing a thesis about the world. It is a rule being executed by someone who would often rather not be executing it, and the rule is written down somewhere — in a prospectus, a methodology document, or a research paper from 1993 that a surprising portion of the industry still runs on.

That fact is the quiet foundation of an enormous amount of what happens in markets. Somebody wrote a sentence. When this happens, buy this much. The sentence existed before the code did, and because it exists, the behaviour can be anticipated — by the desk running it, and by anyone outside willing to do the arithmetic.

A good deal of what we call skill on a trading desk is having read the sentence before the other person did.

Machine learning takes the sentence away.

What is actually lost

Less than the headlines suggest, and in a more specific place.

A firm running a machine-learned strategy still chooses the objective. It still sets the risk limits, picks the universe, decides how positions are sized and when to halve the book and go home. All of that is written down. Most of the apparatus is a rulebook, and it is as legible as it ever was.

What is not written down is one link in the chain: the step that turns this morning's market data into a view on what to own. That is the only piece that has changed, and it happens to be the piece that determines direction. Everything else governs how much and how carefully.

Which gives us a clean test. Hand me every document the firm has written — mandate, limits, universe, sizing methodology, execution policy, the lot. Can I tell you what they buy on Thursday?

For a fund that must maintain constant exposure, yes. The prospectus is sufficient. I can compute the size and the timing before the close.

For this firm, no. I can bound the size and predict the rhythm. The direction sits in the one component nobody wrote.

The ladder

Flow does not divide neatly into readable and unreadable. It sits on a ladder, and machine learning moves it down a rung.

Computable. The rule is published, so you calculate the flow in advance. Index rebalances, scheduled hedging programmes, funds that must maintain constant exposure. Specific, dated, near-certain — the flow people build businesses around.

Inferable. Nothing published, but the signal family is well known and the research is public, so trigger levels can be estimated. Trend following lives here. Approximate, occasionally embarrassing, but a genuine forward statement made from public information.

Estimable. Nothing to read and nothing to reverse-engineer. But stable behaviour leaves a statistical footprint, and footprints can be regressed. On an ordinary Tuesday this works well enough to trade on.

Notice what survives the descent. At the top you can say this vehicle sells this quantity into Thursday's close. At the bottom you can say systematic funds are probably long and would probably sell through the level. Both are useful. Only the first is an event you can position against.

And the bottom rung has a defect worth more attention than it gets. An estimate fitted on the last six months keeps producing confident answers on the morning the relationship stops holding. It does not degrade gracefully; it stays fluent and becomes wrong. A published rule never does this — it is exactly as precise in a regime it has never encountered, because it was never inferred from a regime in the first place.

The estimate is at its most convincing shortly before it fails.

Three people who can no longer read the rulebook

The trader outside it. (AI as a market participant.) When a market moves oddly, you can no longer attribute it — which matters more than it sounds, because mechanical selling that exhausts by Thursday and genuine repositioning that has further to run call for opposite responses. Unreadable flow does not announce itself. It leaves a move nobody can account for.

The desk running it. (AI as an analytical tool.) Risk asks what the book does if yields jump twenty basis points overnight. With a written rule, someone reads the rule and answers. With a fitted model, the only available answer is what the book did on the twelve historical days that vaguely resembled it — a range drawn from the past rather than a statement about the future.

The supervisor above it. (AI as a systemic risk.) You cannot inspect a rule that was never written, and requiring a firm to explain its model asks for something that does not exist in explainable form. Meanwhile a great many firms are fitting similar models to the same few decades of history, so behaviour that used to be independent begins to converge — while the opacity makes the crowding almost impossible to see forming.

Why here

Pandemonium has spent its published life on flows whose rules you can read. Leveraged funds that must rebalance whether or not anyone wants them to. A capital structure whose convexity you can trace instrument by instrument. Hedging demand you can size from a forward book. The argument was always the same: read the plumbing, and the behaviour stops being mysterious.

This section is that subject with the plumbing removed. Same question, harder version.

Mechanics first, from a desk perspective, sources tagged, working shown rather than asserted. No forecasts about what AI does to civilisation, no opinions on whether the machines are conscious. Just a careful look at what happens to a market when a growing share of the flow inside it can no longer explain itself.

First piece: The Rule Nobody Wrote, 18 September.