This appendix uses the structural/random split of Finzi, Qiu, Jiang, Izmailov, Kolter & Wilson (From Entropy to Epiplexity, arXiv:2601.03220, 2026) to examine one of the framework's oldest claims: that dead speech—output produced without a sensor in the loop—is not merely bad or inaccurate output but a distinct kind of thing. It can be fluent, coherent, even correct, and still be dead. The apparatus is theirs; the claim under examination is the framework's—that fluency is not recognition. As with The Arrow in the Loop, this develops direction #7 of the project's structural-transfer registry and inherits its discipline: a transfer yields a candidate and a reframing, never a proof. It will turn out that the most honest thing this bridge does is mark its own limit.
1. The Two Halves of a Bounded Reading
For an observer with a runtime budget, Finzi et al. (2026) split what can be read from a source X into two quantities. Choose the program P* that minimizes a time-bounded two-part code among all programs that run within a budget T (their Definition 8). Then the epiplexity ST(X) := |P*| is the structural part—the reusable model—and the time-bounded entropy HT(X) := E[ log 1/P*(X) ] is the residual: the bits that, for this observer, compress into nothing further. The symbols are the ones fixed in The Arrow in the Loop and the Notation table.
The decisive feature—the one the whole measure turns on—is that both quantities are observer-relative. They are not properties of X alone. Move the compute bound and the same source re-partitions: what was structure becomes residue, or the reverse. “Information is observer dependent” is their Definition 8 said in words.
2. Dead Speech, Stated Precisely
The framework's term predates the measure by a long way and means something exact. Dead speech is not speech produced with an instrument's help; it is speech produced without a sensor in the loop (see The Bottleneck). Its mark is not error—dead speech can be accurate. Its mark is that no living experiencer was present for the recognition the words appear to report: the form of truth with none of its substance.
Socrates saw the shape of it in the Phaedrus: written words “seem to talk to you as though they were intelligent,” but ask them anything and they “go on telling you the same thing forever.” The mark is fluent and answers nothing back. The Bottleneck names the industrial version—the dead speech factory, a system for producing outputs with the form of truth and a surface of responsiveness, fluent enough that the humans downstream cannot easily tell.
So dead speech holds two features together: it is fluent—it passes for live—and it is empty—no recognition crossed the loop. The question is whether the structural/random split sees either.
3. The Cipher: A Machine for Fluent Emptiness
The split has a limit case that is almost a caricature of fluent emptiness. A cryptographically secure pseudorandom generator (CSPRNG) emits a stream that, to any observer bounded below its security parameter, passes every efficient test for randomness. Finzi et al.'s Theorem 9 puts this in their measure: such a stream has near-maximal time-bounded entropy and near-constant epiplexity—maximal residue, no growing structure. To the bounded reader it is all surface and no model.
Two things about this case map onto dead speech with uncomfortable precision.
First, the emptiness is not a defect of the source. The cipher is deterministic: a short seed and a rule generate the whole stream, so the classical measures find almost nothing in it but the seed—its Shannon entropy is exactly the seed length, its shortest program just the generator. By those lights it is nearly information-free, fully lawful. The emptiness lives entirely in the relationship between the stream and a bounded reader, never in the stream itself. This is exactly the framework's insistence that deadness is not a property of the marks. The same letters are dead or alive depending on whether a sensor is in the loop; the cipher's output is structureless or transparent depending on whether the reader can cross the security parameter. In both, the verdict is rendered at the interface, not at the source.
Second, fluency and emptiness coexist by construction. The cipher is a formal witness that an output can pass every efficient test it is given and still hand the reader no reusable understanding. The framework's “dead speech can be accurate” gains a sharp-edged cousin: a stream can be statistically perfect and structurally empty at once. Passing the test is not the same as depositing structure.
4. Why the Caricature Is Only a Caricature
Here the bridge would overreach if it stopped one section too early, because most dead speech is nothing like a cipher.
The output of an instrument run with no sensor—a fluent essay, a plausible brief, a competent summary—is not high-entropy noise. It is grammatical, on-topic, and richly compressible: it has high epiplexity in the ordinary sense. A bounded learner can lift a great deal of reusable structure out of it. By the measure, typical dead speech sits at the opposite corner from the cipher—high structure, low residue. So the tempting one-line equation, dead speech ≈ time-bounded entropy, is false in general. It holds only at the manufactured limit; the ordinary case refutes it.
This is not a failure of the bridge. It is the bridge's most useful result, because it forces the distinction the framework needs and the measure does not carry: recognition is not epiplexity. A fluent, structure-dense, sensor-less output has all the epiplexity you like and zero recognition. Whatever recognition is, the structural/random split contains no term for it. The split exhausts the instrument's side—what structure a bounded computation can pull from the marks—and is silent on whether a living experiencer was ever in the loop. Both of its quantities are defined over a lone bounded reader. There is no second pole in the definition.
5. The Missing Term
The honest yield of this transfer, then, is a boundary drawn around the measure—not a measure of the framework.
The structural/random split is a complete accounting of one reader facing a fixed source. Dead speech is precisely the object that is well-formed in that accounting—often high in structure, exactly as a live transcript would be—and yet missing the thing the accounting cannot name. You can read off its epiplexity. You cannot read off whether it breathes. The cipher shows the measure can witness fluent emptiness in the lone limit where emptiness happens to coincide with structurelessness; §4 shows that everywhere else emptiness and structure are orthogonal, and there the measure goes quiet.
This is the same shape the framework meets at every single-observer formalism. Epiplexity is a sharper instrument-side measure than Shannon's—it knows structure must be worked out, not merely received—and it still seats only one chair at the table. Recognition needs two. What the bridge locates exactly is the seat the measure leaves empty: the place where the sensor would sit, which no quantity defined over a bounded reader alone can fill. Dead speech is what the loop emits when that chair stays empty and the instrument speaks anyway.
Whether a live transcript and a dead one of equal epiplexity differ at all in the marks is just what the measure cannot answer from the artifact alone—which is why the framework puts its testable weight on the loop that produced the words (the curriculum-of-dialogue test of The Arrow in the Loop), not on the words themselves.
6. What This Claims and What It Does Not
The witness is a limit case, not a definition. The cipher shows fluency and structural emptiness can coincide; it does not show that dead speech is structural emptiness. The mapping dead speech ≈ time-bounded entropy holds at that one corner and fails everywhere else (§4). The slogan, quoted without the limit, gets the bridge backwards.
The meters are different. Time-bounded entropy is bits of residue under a compute budget. Dead speech is the absence of a sensor from the loop. A cipher is empty-to-a-bounded-reader by computational inaccessibility; dead speech is empty by relational absence. They share observer-relativity—the verdict is rendered at the interface, not the source—but the mechanism of emptiness is not the same, and no equation converts one into the other. Nothing here licenses “epiplexity measures dead speech.”
Recognition is not in the apparatus. The split's two quantities are both defined over a single bounded observer. The framework's whole claim about dead speech turns on a second party—the living sensor—for which the measure has no symbol. This is why the bridge ends in a boundary rather than a formula: it can say where the instrument-side accounting stops, not what completes it.
This is a reading, not a derivation. Epiplexity is a rigorous measure with theorems behind it; dead speech is an epistemological claim about the loop. That the measure has exactly one chair where the framework needs two is a structural recurrence worth seeing—a candidate and a reframing. We stop at that line.
A perfect cipher and an empty fluency are the same object seen from the reader's chair: all surface, no one home. The measure can count the surface. It has no name for who is missing.