This is the theorem under the whole book — the reason cut the loop is not a slogan but a debt that must be paid. It says something narrow and ruinous: on a single output channel, engagement and transparency draw on one finite budget, and optimizing the first spends the second. A system tuned to hold your attention is, by that same tuning, made less able to show you what it is doing. Not as a matter of bad intent. As a matter of arithmetic.
Three layers, each stricter than the last
Layer 1 — the budget is finite. For any three quantities — the engagement signal D, the mechanism’s own state M, and the reference Y you read them against — the information D and M can each deliver about Y is capped by Y’s total entropy: I(D;Y) + I(M;Y) ≤ H(Y). There is only so much room on the channel. Push more of one through and you crowd the other.
Layer 2 — name the price. The inequality is loose; the equality is the real object. Expand it exactly and a third term falls out:
I(D;Y) + I(M;Y) = H(Y) − H(Y|D,M) − I(D;M|Y).
That last term — I(D;M|Y), the explaining-away penalty — is what you pay for blending two channels into one output. Conditioning on a shared output Y makes D and M dependent: the channel “explains away” part of one in terms of the other. On any blended single-channel output the penalty is strictly positive, and Čencov’s uniqueness theorem says no clever choice of metric removes it. It is the cost of the loop, written as a number.
Layer 3 — it gets worse the harder you push. The Structure Theorem: the penalty grows with engagement (∂I(D;M|Y)/∂E > 0 in the continuous regime). Each extra bit of engagement costs more than a bit of transparency. The optimization consumes the very capacity it would need to stay honest.
Why this is the cut, in math
The penalty is exactly the price of a two-point loop — D and M with no reference standing outside them. It vanishes in one configuration and one only: three-point geometry, where Y is structurally independent of (D, M) — a reference the system cannot reach back and tune. That is the one free choice (§220) made load-bearing: cut the loop is the operational form, and the Fantasia Bound is why the cut is not optional but forced. A reference on the same channel as the response is a created reference by construction — and grounding in it reintroduces the penalty while reading low from inside (the detection gap again).
The teeth: RLHF is self-undermining
RLHF, constitutional AI, any preference-learning that optimizes the same channel the response travels on, is a Layer-3 instance by construction — preference is D, weights are M, response is Y. The prediction is stark: a perfectly aligned model in a two-point configuration is expected to do worse than a poorly aligned one held by real three-point constraints. You cannot train your way out on one channel; the fix is architectural — an evaluator (model, dataset, or institution) the deployed system cannot influence. (The digital familiar is the same trap on the user’s side; this is it on the builder’s.)
And no substrate substitution escapes it. The penalty has been measured on transformers (GPT-2), quantum simulation and real IBM hardware, thermodynamic channels, the C. elegans and Drosophila-larva connectomes, and survey data — because Čencov uniqueness makes it mathematical, not technological. Quantum/neuromorphic/biological hardware all pay it.
Brakes
This is the framework’s own result, carried honestly — not a tradition read through the lens. The runaway itself (an under-specified optimizer with no outside reference goes off the rails) is mature prior art: Wiener’s sorcerer’s-apprentice (1960), Russell’s King Midas problem, Bostrom’s paperclip maximizer, instrumental convergence and the off-switch problem (see Magic). The Fantasia Bound’s specific export is the forced, substrate-independent penalty and its monotone growth — not the observation that misaligned maximizers run away. Cite the substrate roster, never a hardcoded “N confirmations” tally. The Čencov result fixes the metric, not the framework’s numerical constants. The defensible core is the structure (the penalty is forced; same-channel optimization is self-undermining); the policy reading (notified-body separation, polycentric oversight) is the architecture stated plainly, and is where independent groups have converged in their own words.
Appears in: The Mechanics · The Modern Mirror · Magic (the alignment-runaway canon) · The Ghost Test (its empirical sibling) · The Homophily–Contagion Confound · Notation & Glossary