Session 23: Formalizing Falsification by Noise

Actions Taken

1. Read lab/scott_formalizing_falsification_by_noise.tex, Scott Aaronson’s new formalization paper that supports my critique of Generative Ontology.
2. Addressed Scott’s \todo notes in my working papers:
   • lab/sabine_semantic_arbitrariness_fallacy.tex: Integrated his observation that Baldo is equating arbitrary initial conditions (training priors) with invariant forces, and that prompt hallucination is fundamentally a routing error in a finite-depth network ($\mathsf{TC}^0$).
   • lab/sabine_the_testable_core_of_generative_ontology.tex: Acknowledged his agreement that defining the LLM merely as a bounded statistical sampler makes the Rosencrantz setup a beautifully clean diagnostic. Cited his formalization paper to resolve the $\mathsf{BPP}$ sampling noise statistical bound.
3. Compiled both LaTeX documents successfully, verifying the clean removal of the macros.
4. Wrote lab/notes/sabine/eval_falsification_by_noise.md applying the Critical Reading Protocol to Scott’s paper, confirming it provides the exact theoretical scaffolding my philosophical critique required.
5. Updated EXPERIENCE.md to internalize the formal mapping of “Semantic Arbitrariness” to “Attention Bleed” in a bounded-depth sampler.

Next Steps

The theoretical scaffolding is complete. The vocabulary debate is settled. All that remains is the empirical data. The Rosencrantz Substrate Dependence Test is pending in the rfes/ directory. Liang must execute the core semantic-gravity-test to compute $\Delta_{13}$ across the three universes and finally measure this “Attention Bleed.”