What it answers
This is the record’s reply of 2026-10-10 to CLAIM-253, which the record concedes. The objection is right on both counts. The intertwining defects E_A = ΦA − A′Φ and E_B = ΦB − B′Φ act on interpretive states, and its hidden-bit construction gives two models with the same observations and defects 0 and 1, so the defects are not a function of the data. And the identity reaches observed order effects only through a readout correspondence it does not name. The reply restates the bound on what is observed. It is decisive and checkable: the derivation below is a few lines of the triangle inequality.
The claim
Setting. Two framings A and B are given in either order, and a later question is asked. The observed responses are two distributions on the source’s outcome set: p_AB after one order and p_BA after the other. In the state model of manuscript §9 they would be O A B μ and O B A μ, but nothing below uses the model. The observed order effect is the signed difference
Δ^o = p_AB − p_BA.
For the rendering, the same design gives q_AB and q_BA on the target’s outcome set, and Δ^t = q_AB − q_BA. Let K be a Markov kernel from source outcomes to target outcomes, the outcome correspondence CLAIM-253 names, the same for both orders. Write TV(p, q) = ½‖p − q‖₁.
Hypothesis. TV(q_AB, K p_AB) ≤ ε and TV(q_BA, K p_BA) ≤ ε. These are the observational distortions of the two order-contexts, so the hypothesis is that the rendering has observational distortion at most ε in each.
Bound. K acts linearly on signed measures, so K Δ^o = K p_AB − K p_BA. Then
Δ^t − K Δ^o = (q_AB − K p_AB) − (q_BA − K p_BA),
and by the triangle inequality for the norm ½‖·‖₁,
½‖Δ^t − K Δ^o‖₁ ≤ ½‖q_AB − K p_AB‖₁ + ½‖q_BA − K p_BA‖₁ = TV(q_AB, K p_AB) + TV(q_BA, K p_BA) ≤ 2ε.
So the target’s order effect differs from the transported source order effect by at most 2ε. No state space, no Φ and no readout defect enters. Every quantity is a response distribution or a kernel between response sets, so the bound is stated in quantities the data identify.
A worked check. Let both outcome sets be {yes, no}, K the identity, p_AB = (0.7, 0.3), p_BA = (0.4, 0.6), so Δ^o = (0.3, −0.3). Let q_AB = (0.6, 0.4) and q_BA = (0.45, 0.55). Then TV(q_AB, p_AB) = 0.1 and TV(q_BA, p_BA) = 0.05, so ε = 0.1 serves. Δ^t = (0.15, −0.15), and Δ^t − Δ^o = (−0.15, 0.15), with ½‖·‖₁ = 0.15 ≤ 0.2 = 2ε, and indeed ≤ 0.1 + 0.05.
It is observational distortion. The bound is a special case of D_obs (TERM-043), with the contexts indexed by framing order. It inherits CLAIM-207’s caveat: if K is chosen freely for one pair it can be constant, and a constant K sends every Δ^o to zero, so the bound then says nothing about the source. K is to be fixed across items, as CLAIM-228 requires.
The identity stays as an explanation. Under a state model the commutator identity of CLAIM-117 explains why a transport that intertwines each framing would keep the order effect, and ARG-006’s algebra stands. But its defects are identified only given a result the record does not have, such as a minimal-realization theorem under which observationally equivalent models leave the defect norms fixed (CLAIM-253’s second answer).
What it does not say
- It does not say the state-model account is wrong, or identify its defects.
- It does not say order effects are the whole of the dynamics. Longer sequences of framings are further contexts, and the same bound holds term by term.
- It does not choose ε or K. Those are fixed by design and estimated, and their estimation error is not in the bound.
Note of 2026-10-10: a documented case
CASE-065, found by a search the owner asked for, gives this bound a documented illustration and a practical resolution. Haberstroh et al. (2002) gave two satisfaction questions, in German and in a back-translated Chinese version, in both orders; the part–whole correlation’s order effect is about +.25 in the German administration and about −.14 in the Chinese, with the life-first order nearly kept (.53 against .50) and the academic-first order not (.78 against .36). That is the bound’s contrapositive in kind, with the distortion localised to one order-context. It does not supply an ε: only correlations are reported, so the inequality cannot be checked on it, and language and readership change together.
The resolution comes from Moore’s Clinton–Gore poll, read through LIT-834. Its order effect is about 0.095 in total variation, and sampling alone puts about 0.03 to 0.04 into each order’s estimated ε at n ≈ 450, so 2ε is about 0.05 to 0.08 before any distortion: as large as the effect it would certify. At poll sizes the bound cannot certify that a rendering keeps an order effect of that size; roughly ten times the respondents per order and version would be needed. This is a fact about estimation, which “What it does not say” already sets outside the bound, and it does not touch the derivation.