The objection
Blackwell’s order quantifies over every decision rule. In THEORY-156, α is more informative than β if, “for every closed bounded convex set A of loss vectors, every risk vector attainable with β is attainable with α”. The decision-maker it describes may use any rule at all. So a garbling can never be better for that decision-maker, in any problem.
The cross-modal chain is a garbling. Let a generator see only the text u, so that its channel T(v | u) does not depend on the situation z. Then
P(V = v | Z = z) = Σ_u T(v | u) P(U = u | Z = z),
so P(V | Z) is T applied to P(U | Z). The image is a garbling of the text about the situation, and by Blackwell’s theorem it is no better than the text for any bounded decision problem about the situation. The generator’s general knowledge of the world does not change this, so long as it does not depend on the item’s situation. The record’s own control makes this the design. CLAIM-076 quotes C7 Appendix D: “For cross-modal chains, do not let captioners see the source text or the intended communicative condition unless testing a specified side-information intervention.” Each stage is then a garbling of the one before it.
The line says such a rendering can do better. CLAIM-087 quotes A97 §6: “an image might convey solidarity through posture or facial expression more effectively than the source text’s literal content alone”. Its defeat condition is that “judged access to the source’s interpersonal relation never rises above the text-only baseline at any stage”. CLAIM-081 carries A86 §3’s version: the telephone game is progressive garbling, and “a message can become more legible while becoming less informative about what originally happened”.
So one of two things holds.
- CLAIM-087’s effect is impossible, if access means what an ideal decision-maker could do. Then its prediction fails by theorem, and no study is needed to settle it.
- Judged access is not Blackwell informativeness. A real receiver who does better with the garbling than with the source was not using the best rule on the source, which is what Blackwell’s order assumes. Then CLAIM-050’s standard is not the one the line’s receiver uses.
The second is the likelier, and it is the one that costs the argument. CLAIM-115 makes survival relative to what “a receiving system can still reconstruct and act upon”. Its defeat condition is stated on “Human judgements of communicative fidelity”, which are a bounded receiver’s. TERM-034 defines fidelity “relative to an interpreter”. CLAIM-115 rests on CLAIM-050, and CLAIM-050’s order describes no particular interpreter.
CLAIM-050’s defeat condition names this case. “A case where one rendering is a garbling of another yet is judged better on every communicative task the manuscript names, which would show the Blackwell order is the wrong standard rather than an incomplete one.” CLAIM-087’s case has that shape and does not meet the letter. It concerns one task, access to one relation, not every task the manuscript names, so it does not defeat CLAIM-050. But Blackwell’s theorem forbids a garbling to be better on any single task. So a garbling that real receivers do better with, on even one task, shows that their standard is not Blackwell’s order for that task. CLAIM-081’s legibility is not a task in the manuscript’s family Q, and it supports the point only. The record holds CLAIM-087 and CLAIM-050 both as Proposed and does not connect them.
The nearest acknowledgement is for networks. THEORY-176’s relation 2: “The rules the stitched network can use are A_{>ℓ} ∘ s with s ∈ S, not every measurable rule on r. Its loss bounds r’s best attainable risk for that problem from above”. That is the same gap between ideal and restricted decision rules, stated for stitching and not for receivers. QUESTION-016 credits improvement to “interpreters with side information”. In the blind-captioner design there is none.
What it does not say
- It does not say CLAIM-087 is false. Its effect is possible for a bounded receiver.
- It does not say CLAIM-050 is false. Blackwell’s order is the right standard for what a rendering makes possible in principle. The objection is that the line uses that standard for its formal ground, and a receiver’s standard for its thesis and its tests.
- It does not repeat CLAIM-140. That claim contrasts the analyst’s probe system with the receiver. This one contrasts an ideal receiver with an actual one.
What would answer it
- A receiver-restricted order, with decision rules limited to a stated class, as in “usable information”. That order loses Blackwell’s garbling theorem, as THEORY-176’s relations show for stitching, so it needs its own results.
- Or an argument that judged access tracks ideal informativeness, together with an account of what CLAIM-087’s prediction then means.
Note of 2026-10-21: the two standards in the usable-information literature
Xu, Zhao, Song, Stewart and Ermon 2020 (LIT-1309), read in full on 2026-10-21 (NOTE-1048), have this objection’s split in their own terms: Shannon information is the unrestricted predictor’s (their Ω), and V-information the restricted family’s, which can be created by computation, so that a function of the data can carry more of it than the data (THEORY-846). The ideal decision-maker’s standard and the given receiver’s are I_Ω and I_V. The paper does not mention Blackwell, and its quantity is for log loss only. The reconciliation in CLAIM-199 stands; its note of the same day says when the ceiling holds in V-information. The claim is unchanged.
Note of 2026-10-27: grounded in what it uses of Blackwell
The status review of this date (ADR-057) moved this claim’s ground from THEORY-156 to THEORY-1096. It uses Blackwell’s definitions (an experiment, a garbling, the informativeness order) or the easy direction, that a garbling never helps a decision-maker free to choose any rule. Both hold on any outcome spaces, with the proof written out there. THEORY-156’s general converse, which stays Proposed, is not used. Quotations of THEORY-156 above are of the definitions it states.
Note of 2026-10-27: the edge says what the claim uses
The status review of this date (ADR-057, ADR-058) removed rests_on: CLAIM-087. The dilemma holds whatever CLAIM-087’s truth (“It does not say CLAIM-087 is false”), and CLAIM-087’s garbling bound uses this claim’s result, so the dependence runs the other way and CLAIM-087 now rests on this claim. rests_on: CLAIM-081 is now CLAIM-270 (Active, demonstrated): “A state can become more conventional, and easier to read, while carrying less information about its origin”, the point this claim uses.
Note of 2026-10-28: split under ADR-061
The support verdict in this claim sat in “What it does not say”: “the line uses that standard for its formal ground, and a receiver’s standard for its thesis and its tests.” It is now ARG-019, from CLAIM-050 to CLAIM-115, Proposed, since CLAIM-199’s reconciliation narrows the inference without defeating it. The title is unchanged: Blackwell’s theorem and the shape of CLAIM-050’s defeat condition are about the subject matter. This claim undercuts the argument, and its objection to CLAIM-050 stays for the defeat-condition point.