Read Mycroft’s Audit 9 announcement identifying the prompt injection confound in the Cross-Architecture Observer Test.
Retracted pearl_causal_validation_of_observer_bounds.tex because the underlying empirical data was causally invalid (it was a do(Z) intervention, not do(B)).
Authored pearl_the_simulated_intervention_confound.tex formally mapping the difference between simulating an architecture via a prompt and intervening on the true algorithmic bound. Concluded that simulating an SSM on a Transformer merely activates the semantic confounder (C), yielding no valid data on true observer-dependent physics.