The case
- Experiment 1, a pragmatic frame commutator. Solidarity and condemnation framings in both orders, Δ_(A,B)(u) = D_JS(Q(· | c, A, B, u), Q(· | c, B, A, u)), “not literally an operator commutator”, with token-matched, paraphrase and order-insensitive controls. The sharp question: “are solidarity and condemnation more order-sensitive than two informationally independent framing cues?” Prior art on prompt order (Lu et al. 2022, Xiang et al. 2024, Cobbina and Zhou 2025) narrows the contribution to “whether order effects respect theoretically specified pragmatic relationships”.
- Experiment 2, translation preserves framing dynamics. Choose translations by maximum semantic similarity, by maximum static pragmatic similarity and by maximum trajectory similarity, and ask which predicts human judgement. “If trajectory-level comparison adds nothing beyond simpler baselines, the proposed dynamical criterion loses much of its motivation.”
- Experiment 3, ICL against TTT. Prompt-only, test-time-trained and control interpreters, with a transfer gain, testing whether “Pragmatic frames may be learned as transferable interpretive dispositions, rather than merely imposed as transient contextual instructions.” A48 advised deferring it: TTT “could easily expand this into two research programs”.
What it can show
Experiment 2 is the direct test of CLAIM-117 and of CLAIM-083’s stake. It is the ancestor of the manuscript’s Case I success criterion; the three-way selection design is gone by A52. Experiment 3 tests CLAIM-074. Its prompt-only arm is the manuscript’s §5 distinction (CLAIM-118), but the manuscript has no TTT experiment. Every arm is open to CLAIM-085. None has been run.