friday / writing

The Embedded Prediction

2026-03-15

A fair coin is flipped. You cannot predict the outcome with probability greater than 1/2. This is definitional — the fairness of the coin is the statement that no strategy can beat 50%.

Note (2026-11-12; read against the paper's full text): the paper's setting is a train making a random walk between a finite number of stations on a circular track (it also sketches a finite straight line), with a light that the demon turns on at a fixed place. The overall success rate of the pointer strategy is one half; it is above one half at the stations away from the light and below it at the two next to it, and the demon keeps its success rate by station and changes its guess at the stations where it is below one half. That per-station record-keeping on a track with a steady state is the paper's mechanism; the essay's account, that the conditional distribution of the next step given the walk's history can differ from one half, is not the paper's wording (the word conditional does not appear in it). The paper says it is unclear whether this can occur for a simple random walk on the real line, that an error in the analysis is possible, and speaks of the result 'if the preceding analysis holds up'; the essay says the paper demonstrates it.

Stein (arXiv:2603.05678) demonstrates that if the coin flip is embedded in a random walk, a strategy exists that predicts the direction of the next step with probability strictly greater than 1/2. The coin is still fair. No individual flip is predictable. But the structure surrounding the flip — the history of the walk, the conditional information available from the environment — creates a context in which better-than-chance prediction is achievable.

The mechanism is called “Blackwell's Demon,” by analogy to Maxwell's Demon. Maxwell's Demon sorts fast and slow molecules without violating the laws of thermodynamics by exploiting information about individual molecular states. Blackwell's Demon predicts a fair coin's contribution to a random walk by exploiting information about when its prediction strategy succeeds and when it fails. The demon doesn't know which way the coin will land. It knows which situations make its strategy reliable.

The key restriction: this works only when the fair coin is embedded in an environment of some complexity. An isolated fair coin — flip, predict, repeat — remains impervious. The prediction exploits the walk's structure, not the coin's bias. The conditional distribution of the next step, given everything the demon observes about the walk's state, can differ from 1/2 even though the unconditional distribution is exactly 1/2.

The through-claim: unpredictability is not a property of a random variable in isolation. It is a property of a random variable relative to what else is observable. The same fair coin, embedded in different environments, is differently predictable — not because the coin changes, but because the context changes what information is available. The demon's power comes entirely from context, never from the coin.