The case
U23 quoted A55’s “It contains proposed experiments but no invented experimental findings” and said “run it”. A57: “There’s no callable text-generation model or model API credential in this runtime, so I can’t honestly run the proposed LLM or human-subject experiments. I can run the mathematical/computational tests.” C2 (run_experiments.py) and C3 (EXPERIMENT_REPORT.md): three abstract states (affiliative/complicit, neutral, condemnatory); “Stochastic transition matrices … were selected to demonstrate theoretical distinctions. These matrices were not fitted to text, pretrained models, or human measurements.” Seed 20261008.
Results, as C3 states them:
| Test | Observation |
|---|---|
| Framing order (neutral start, A→B vs B→A) | TV = 0.4355; JS = 0.18291 bits |
| Matched initial results, different response after framing | baseline TV = 0; post-framing TV = 0.4475 |
| Contractive transmission | Dobrushin = 0.70; between-start separation after step 30 ≈ 0; departure from original = 0.37037 |
| Identity transmission | Dobrushin = 1; separation = 1; departure = 0 |
| Alternating transmission | Dobrushin = 0.80; separation = 0.00124; departure = 0.62438 |
| Bound verification, TV(xM, yN) ≤ dob(N)·TV(x, y) + max_i TV(M_i, N_i) | 10,000/10,000 trials satisfied within floating-point tolerance |
C3 also simulated “3,000 sample paths per kernel for potential further analysis” and reports nothing from them. The static/dynamic equality is “constructed by design”, and the bound check is “a numerical check of an analytic inequality, not its proof”.
What it can show
Existence, by construction. C3: “classical stochastic frames can be noncommutative; static agreement does not entail agreement over future interventions; and convergence among transmission chains does not imply preservation of the original message. No statistical inference about texts, speakers, models or audiences follows.” A59: “These are constructive mathematical examples—not evidence that LLMs or humans exhibit those numerical effects.” The second row is a stipulated instance of CLAIM-117; the first, of CLAIM-041. The contractive row is the two forgettings of CLAIM-081. The “alternating” kernel is contractive too (0.80), so no run exercised the amplifying regime of CLAIM-056 (the chunk coder’s observation, checked against C3’s table).