Four Properties in Search of an Architecture

This paper is released as a Perturbation Draft (Temte 2026b): a provisionally complete work published early to maximise the perturbation window, the span of time during which the argument is exposed to external challenge. The draft designation is not an apology for incompleteness but a methodological commitment to iterative knowledge production. This is the second perturbation cycle for the argument; §6 names the exposed joints where external critique is most likely to be productive. An earlier Perturbation Draft (Temte 2026, PD v1.0) proposed that a subset of autistic cognition implements LLM-style architecture: threshold-based statistical pattern learning with regenerative inference. This revision retains the empirical and theoretical observations that motivated the proposal but rejects the bundling. Four architectural properties travel together often enough to constitute an identifiable cluster: threshold-dependent acquisition rather than power-law gradualism; regenerative re-computation rather than cached heuristic response; statistical-pattern rule-use without explicit symbolic representation; and density-matched specialisation through sustained domain engagement. Each is defensible on biological and behavioural evidence independent of any machine-learning parallel. The LLM case becomes existence proof for the computational viability of such a cluster, not a claim of architectural identity. Reframing the proposal this way addresses three weaknesses in PD v1.0: it decouples the argument from the Schaeffer, Miranda, and Koyejo (2023) critique of LLM emergence; it refuses the autism-versus-neurotypical categorical framing that is itself an instance of Binary Bias (Temte 2026a); and it isolates the social-cognition sub-claim, the weakest joint, for honest exposure rather than rhetorical cover.