Published 13 September 2026

Newcomb’s problem

A predictor has already decided what to put in an opaque box. If it predicted that you would take only that box, it put a large prize inside. If it predicted that you would take both available boxes, it left the opaque box empty. The other, transparent box contains a visible smaller prize. You can now take only the opaque box, or both.

Two arguments pull in different directions. People who take one box may be strongly associated with receiving the large prize. Yet if the past contents are held fixed, taking both always adds the smaller prize. Newcomb’s problem asks how a decision should use this combination of prediction and dependence.

Specify what each expectation conditions on

The large prize here is one thousand tokens. Let s be the smaller prize. The observational model explicitly assumes P(full | one-box choice) = q and P(full | two-box choice) = 1 − q. This specifies reliability within each eventual choice group. Overall prediction accuracy alone would not determine these two conditional probabilities without additional information about the groups.

Using those conditional probabilities, expected tokens for one box equal 1000q. For both boxes they equal s + 1000(1 − q). At q = 0.99 and s = 10, the values are 990 and 20 tokens. The association-based comparison favors one box.

Set the two expectations equal to find the crossover: q = (1 + s/1000)/2. With a ten-token smaller prize, the threshold is 0.505. At q = 0.5 the opaque contents are equally likely in both choice groups, so the conditional comparison also favors taking the extra ten tokens.

Hold the past contents fixed instead

The second calculation treats the action as an intervention that changes which boxes are collected while leaving their already determined contents unchanged. Let F be the full-box probability held fixed under this comparison. One box then has expectation 1000F, while both have expectation 1000F + s.

For each fixed state, taking both adds s: if the opaque box is empty the outcomes are zero and s; if it is full they are one thousand and one thousand plus s. A positive s therefore gives a strict dominance argument under the stated intervention model, regardless of F.

The F control is separate from q because the two observational conditional probabilities do not by themselves determine an intervention distribution. No mixture over the already fixed states changes the extra-s result. Conversely, that state-by-state result does not make the two observational groups receive the same opaque prizes.

The disagreement is about the counterfactual

Conditioning on an action selects cases in which that action occurred. Intervening on an action in a causal model asks what happens when that action is changed while specified causes remain fixed. When actions and states are dependent, the two operations need not assign the same probabilities to states.

The experiment makes the difference explicit instead of applying one probability table to both questions. It illustrates an evidential comparison and a fixed-past causal comparison. Richer formulations can model the predictor’s relation to a decision procedure, the timing of commitments, or what a hypothetical change of decision is allowed to change. Their recommendations depend on those further premises.

Saying that the predictor is accurate does not, by itself, explain that causal structure. Nor does saying that the boxes are already filled settle every question about evaluating a policy before the prediction is made. A policy-selection problem before prediction and an action-selection problem afterward should not be silently treated as identical.

Boundaries that expose the assumptions

With a perfect conditional predictor, one-box cases receive the full prize and two-box cases receive only s. The fixed-content intervention comparison still adds s, because it is answering a different hypothetical question. At s = 0, taking the transparent box adds nothing, so the intervention comparison is tied. The conditional comparison can still differ if q differs from one half.

Nozick’s presentation sets out the conflict between principles of choice. The lesson preserves that conflict rather than presenting either calculation as a universal resolution. Continue to Berkson’s paradox for another case where conditioning changes associations.