The ART Tail Does Not Follow: A Correction to the Concentration Corollary of the Algorithmic Regulator Theorem

A published corollary to the Algorithmic Regulator Theorem claimed that good regulators must contain models of their worlds — this note shows that claim is simply wrong, and explains exactly why.

The Algorithmic Regulator Theorem (ART) is about how a "world" W and a "regulator" R interact. The regulator intervenes on the world, and the key quantity Δ measures how much the regulator compresses the world's output — a large Δ means the regulator dramatically simplifies what the world produces. ART's core result is a per-pair bound: for any specific named (W, R) pair, the probability that this pair explains the observed data is suppressed exponentially in Δ unless the pair shares a lot of internal structure (measured by M(W:R), the algorithmic mutual information — roughly, how much W and R "know about each other"). That bound is correct and untouched.

The problem is a corollary that tried to go further. It claimed that because each individual pair is suppressed, the event that M(W:R) is small must also be suppressed — i.e., that a large Δ forces the regulator to carry a model of the world. This is a classic summation error. A bound on each term of a sum tells you nothing about the sum if you don't also bound the number of terms. The paper constructs an explicit counterexample: take a simple one-line regulator (K(R) = O(1), essentially zero complexity) that clamps the world's output to zeros. There are ~2^b such worlds (one for each incompressible string the world would otherwise emit), each individually suppressed by 2^{-b} exactly as ART requires — but there are 2^b of them, so their total probability is Ω(1), not exponentially small. The fiber of low-model-content pairs carries constant mass while the corollary claimed it should be negligible.

There's also a more basic conceptual error: Δ is a property of the readouts (the world's outputs with and without the regulator), while the ceiling on M(W:R) is K(R), a property of the regulator itself. A cheap regulator can open an arbitrarily large gap — by unplugging the measurement instrument, say — without containing any model of the world at all. The corollary mislabeled Δ as a maximum of M(W:R) when it isn't.

What's the correct reading? ART says: model-free explanations are bad bets. It does not say the winning explanation must contain a model. The theorem people actually want — "to regulate a system you must either model it or pay thermodynamic work" — exists, but it's a thermodynamic result (the Sagawa-Ueda feedback second law, in Shannon information terms), not an algorithmic one. Whether an algorithmic version with M(W:R) replacing Shannon mutual information can be proven remains open. The note also makes a methodological point worth keeping: the machine-checked Lean formalization of ART never certified the corollary, and that silence was a signal nobody read. When a proof assistant declines to state something, that's information.