Hard Substrates, Soft Evidence: A Technical Critique of LLM Capacity Attribution

The philosophical debate about whether large language models understand, reason, or merely mimic has produced more heat than resolution. This paper argues the debate is stuck for methodological reasons: all major positions rely on behavioral evidence that underdetermines internal structure, and on assumptions about LLM architecture and training that are technically false or unsupported. I identify two substrates of LLM computation: the soft substrate (natural language in, natural language out) and the hard substrate (the learned computational organization—parameters, attention patterns, activation dynamics). Behavioral evidence accesses only the soft substrate; it cannot identify the mechanistic structure that would settle questions about understanding. This is not merely an epistemic limitation but a structural one: distinct internal organizations can produce identical input-output behavior. The skeptical "stochastic parrot" framing misdescribes both inference (which is geometric transformation, not string-stitching) and training (which increasingly includes feedback tied to task outcomes, not just text statistics). The optimistic inference from impressive outputs to genuine understanding doesn't go through—behavioral evidence is the wrong kind of evidence. Even the most careful empirical work is architecture-bound in ways that aren't acknowledged. Meanwhile, interpretability research—which directly accesses internal structure through probing, circuit analysis, and causal intervention—is generating findings that could constrain the debate, and philosophers aren't engaging with it. I specify what adequate methodology would require: convergence across behavioral evidence, internal probing, causal intervention, and cross-architectural replication.