The objection
It needs the state. Blackwell’s order compares experiments P(V | Z) over a state Z. To test it between a source and a rendering one needs the joint law of Z and each signal: for each source item, the situation it is about, and the distribution of the signal given that situation. CLAIM-050’s tasks are of this kind: “tell teasing from reproach, recognize a position, choose a reply”.
For natural texts Z is not observed. It is what interpretation is meant to recover. So it has to come from somewhere.
- The design. CASE-040 crosses topic with “pragmatic act P ∈ {teasing, reprimand}”, and “makes the act a manipulated factor rather than an outcome”. There the state is whatever the item writers made it.
- The analyst’s labels.
- The judges, whose judgements are then both the state and the outcome.
In each case “the information needed to make the relevant distinctions about the communicative situation” (TERM-028, from A86 §3) is measured against whoever fixed Z. CLAIM-140 found that identifiability in the October thesis is relative to the analyst’s probe system. The same holds here for decision-relevance: it is relative to the labeller of the state.
It orders procedures. An experiment is “an n-tuple of probability measures on a common space, one per state” (THEORY-156). A single rendering of a single source is one draw, under one state. Henley’s rendering of Villon (CASE-010) is one text, about one situation. Its Blackwell fidelity is undefined. What has one is the procedure that produced it, applied across a distribution of situations. So CLAIM-050’s defeat condition, “A case where one rendering is a garbling of another”, is not well typed as written: garbling is a relation between channels, not between texts. CLAIM-207 makes the same point for the outcome-level kernels of the manuscript’s transport, which constrain a translation only when fixed across utterances.
The record has met the measurement gap once. CASE-042: A207 §6 calls a held-out comparison of decision risk between two encodings “a direct Blackwell experiment”, and “By the record’s reading it is not one. It compares a chosen family of losses, which THEORY-176 shows is not Blackwell’s order”. CLAIM-126 limits its own certificate to “decisions about the label used to define the experiment”. Neither says that for natural texts the label is someone’s choice. For stitching, the label comes with the data set. For translation it is the situation, and someone assigns it.
What it does not say
- It does not say a design-based state is illegitimate. Experiments routinely fix the state by design, and CASE-040 is a careful one. It says that the decision term then measures agreement with the design’s labels, and that the theory’s notion of fidelity must not be built into those labels (QUESTION-005).
- It does not say a single translation cannot be judged. It says Blackwell’s order does not judge it.
What would answer it
- A design-based Z with an argument that the design’s labels do not build in the theory’s preferred fidelity (QUESTION-005).
- CLAIM-050’s defeat condition restated over translation procedures, applied across a stated distribution of situations.
- For single renderings, a different measure, stated as such.
Note of 2026-10-29: not an objection to CLAIM-126
The eleventh status review (ADR-057) removed objects_to: CLAIM-126. The
paragraph “The record has met the measurement gap once” cites it for the limit it already states, and says that for
stitching the label comes with the data set, so nothing here contests what
it says.