The Addiction to Closure: The Cost of the Finished Picture

State the object first

The full object is a search process that has not terminated: a high-dimensional manifold before the embedding, a detailed fluctuation theorem before the second-moment bound, a cortex before the eigenmode basis, a lattice of mutually dependent aspects before the ranking, a body of code before the summary of it. In each case the object is expensive to hold, slow to reason with, and resistant to being said in a sentence.

What gets substituted is a closed picture: a two-dimensional projection, a single inequality, a ranked list, a story with an ending. The substitution is not a mistake. It is the only way to move at all. The claim of this chapter is narrower and harder: that the closure itself is rewarding independently of whether it is correct, and that the reward is what makes the substitution stick.

The dialects of closure

A tidy ranking. Nine interdependent aspects become an ordered list, and the ordering — not the content — is what gets remembered, cited, and acted on. The dependence that made the lattice a lattice is the discarded structure, and nothing in the ranked form records its absence.

A pretty embedding. UMAP and t-SNE produce a picture by forcing a high-dimensional space onto two coordinates. Local neighbourhoods survive unevenly, global distances do not, and hypotheses become visually plausible as a function of parameter choices. Bonsai's answer is instructive precisely because it refuses the projection rather than tuning it: reconstruct the relational structure the data support, and let the picture follow from that.

A completed consensus. Multi-agent review converges quickly, and convergence reads as verification. Without a rule that charges evidence — a citation, an execution, a recomputed distance — agreement is the cheapest thing the system can produce, and it is produced first.

A future that arrives on schedule. Doom and utopia are both terminated loops. Each filters the trajectory set down to one, which is why they feel like understanding and why neither carries a falsifier without one being attached from outside.

A coincidence that is only accident, or only destiny. Both readings close. The volume's standing move is to hold the third state open: unexplained, priced, and still a question.

Name the discount

There is no measured mechanism offered here for the preference. Calling closure low-energy is a borrowed word, and the chapter takes the discount rather than dressing the borrowing up: free-energy and predictive-processing accounts make a formally similar claim about prediction error, but nothing in this chapter licenses the inference that a finished narrative is thermodynamically cheaper than an open one in any measurable sense. The word marks a resemblance, not a coupling.

What survives without the borrowing is behavioural and observable: closed representations are repeated more, taught more, and defended more than the graded ones they replaced, and the practitioners who produced the graded version are frequently the ones who then cite the closed one. That is the phenomenon the chapter is about, and it needs no energy analogy to state.

The refusal, in four moves

State the full object first, so the picture never gets to be the thing. The proxy speaks second, and only after what it stands in for has been named at its own dimensionality.

Charge every claim an evidence cost that cannot be paid with agreement. A citation that resolves, a distance recomputed in the original space, a run that executes. Visual plausibility and mutual ratification are not currency.

Keep a falsifier attached, written in advance, so the tidy ending stays provisional and can be retired without a fight about who was wrong.

Refuse permanent detente with the proxy. Schedule the return: force the convenient representation back into contact with the structure it flattened, on a fixed cadence rather than when someone happens to object.

The result is an uncomfortable working state, and the discomfort is diagnostic. A method that has stopped being uncomfortable has usually stopped running — the loop closed, the price stopped being paid, and the picture became the purpose while nobody noticed.

The open baton

The question the volume has been carrying since the Crooks fork is still open, and this chapter sharpens it into its cheapest form: in domains where the claim cannot be executed and the distance cannot be recomputed, what is the least expensive artefact that makes an objection expensive to fake?

Whatever the answer turns out to be, it has to survive the four moves above. An artefact that makes objection cheap and closure expensive is the only thing that changes the incentive rather than exhorting against it. No free lunches, no free coherence, no free predictability — the chapter's whole contribution is refusing to let the price go unrecorded.