E3: Semantic Layer Ablation
experiment
ClaimProtocol for testing whether LLM semantic modeling yields value-consistent action selection beyond non-LLM alternatives.
Goal: test whether LLM semantic modeling yields value-consistent action selection beyond non-LLM alternatives.
Design
Keep state estimator constant, compare action selection using (a) LLM semantic layer + commitment store vs (b) rules + small models + same commitment store.
Outcomes
User audit agreement ("this matches what I care about"), explanation trace quality, goal substitution detection rate.
Falsifier triggers
No advantage in value-consistency; LLM increases confabulation and reduces trust.
Tests
Architectural necessity of the understanding layer (D4).
Depends onP4: Understanding Layer Adds Value