Emulation Diagnostics for Model Self-Reports: A Bridge Between Interpretability and Welfare Assessment

Abstract

Two adjacent but methodologically separate substrates currently coexist inside Anthropic's published record. The first is a body of mechanistic-interpretability work, most prominently On the Biology of a Large Language Model, that produces attribution graphs over sparse activation features and explicitly cautions that a model's verbal self-explanation can diverge from the underlying mechanism. The second is a welfare-assessment program, most visibly the §5 Claude Opus 4 welfare assessment in the May 2025 system card, that interprets model self-reports about preferences, distress, and affect. The vulnerability this paper takes up does not live inside the welfare assessment itself; it lives in downstream propagation. When the §5.2 descriptive prose is quoted in press, summaries, and secondary commentary stripped of its §5.1 hedge frame, the surface-signature load the §5.1 frame had been carrying is no longer present. This paper proposes four diagnostics: an emulation-versus-state activation contrast over matched prompts; a distribution-dependence ablation re-running the published two-instance free-conversation protocol; a structural-signature comparison of self-reports against other-reports and against empirically calibrated uncertainty; and a propagation diagnostic that compares §5.2-with-hedge-frame against §5.2-hedge-stripped against factual and creative comparators across a six-pattern ghost-pattern taxonomy. The framing is additive to the welfare program rather than adversarial.