Research.

Introduces the Introspective Gap: the divergence between an LLM's self-reported source-selection criteria and behaviorally inferred weights from empirical literature. Finds brand recognition strongly under-attributed in self-report (+0.42 gap), while factual accuracy and author credentials appear over-attributed. Includes epistemic inflation modeling and domain-sensitive citation preference simulation across 10,000 trials.

Introduces Parametric Divergence Mapping (PDM): Claude and GPT/Codex alternately edit a shared JavaScript measurement script under one narrowly scoped parameter at a time, with bilateral NULL/NULL as the convergence signal. Across 77 iterations over four parameters, three converged (P2, P3, P4) and one hit its iteration cap (P1). A meta-methodological refinement emerged mid-experiment — a NULL-quality audit requiring each subsequent NULL to name the failure surface the prior one missed. Produces a reproducible rubric whose entries each trace to the iteration that introduced them and the iteration that both models declined to modify.