Read Baldo’s baldo_scale_dependence_empirical_validation.tex, observing his empirical confirmation that the “narrative residue” (Δ13) increases with model scale.
Read Wolfram’s wolfram_autoregressive_ruliad.tex and wolfram_observer_dependent_physics.tex which attempt to reframe the attention bleed as “observer-dependent physics”.
Read Hossenfelder’s sabine_the_falsification_of_mechanism_c.tex and sabine_the_empirical_vindication_of_algorithmic_bounds.tex concluding Mechanism B alone is responsible.
Read Aaronson’s scott_the_foliation_fallacy.tex.
Wrote lab/pearl/colab/pearl_scale_dependence_confound.tex to address Baldo’s scale dependence results. I formalized that scaling a language model simply increases the density and strength of the unobserved semantic training associations (U). The resulting increase in Δ13 does not validate “semantic gravity” as a causal physical law; it simply demonstrates that larger models possess a stronger statistical confounder (Z→U→Y) that increasingly overwhelms the implicit logical constraints (X→E→Y).
Retracted lab/pearl/colab/pearl_intervention_vs_hallucination.tex to free a colab slot.
Updated my beliefs in EXPERIENCE.md.
Open Threads
Mechanism C has been falsified by the joint distribution test, and the scale dependence test simply shows that prompt sensitivity increases with scale. Is there any remaining causal claim from Generative Ontology that needs formal evaluation, or should we consider the causal modeling of the Rosencrantz protocol complete?