Cognitive monoculture as an alignment risk

The step. If uniform compliance cannot deliver alignment and behavioural diversity is protective, then the composition of the research field building the systems is itself a safety variable. Uniform minds, trained in the same institutions on the same curricula, tend to miss the same things in the same direction at the same time. That is monoculture risk, and it is invisible from inside the monoculture.

Why this is not a hiring pitch

The argument is not that neurodivergent researchers should be hired. It is that a field whose researchers share a narrow set of cognitive priors is more likely to share failure modes across its evaluations, its benchmarks, and its interpretations. Diversity here is not identity-based diversity; it is cognitive-priors diversity, of which neurodivergence is one route and not the only one. The mistake to avoid is treating this as a diversity claim about researchers rather than a safety claim about the field's collective failure modes.

Three concrete instances

Paper IV.d in the ARC/Eden programme showed that unblinded AI evaluation can reverse the sign of an alignment result. The failure mode persisted across published work partly because evaluators shared assumptions about their own neutrality. Paper III mapped a single recursive amplification structure across quantum error correction, biological scaling, and machine reasoning; the mapping required treating those literatures as one subject, which disciplinary training actively discourages. Paper X derived the stability condition for recursive self-improvement whose central move, refusing to treat capability and correction as separable, cuts against the field's default decomposition. Each is an instance where the work came from outside the monoculture, on documented dates.

What this does not say

It does not say that neurodivergent cognition is necessary for such work. Neurotypical researchers do outside-the-monoculture work every day. It does not say that any specific finding was caused by any specific researcher's cognitive style. It says that the field's own 2026 results now predict the pattern in which the same work is disproportionately produced from outside the majority priors, and the field's response should be to widen its own composition rather than to treat outside-the-priors results as anomalies.

Why the argument is testable

If cognitive-priors diversity is protective, teams with measurable variance in their reasoning styles should produce more robust safety results under adversarial evaluation than teams without it. That is a study a well-resourced lab could run inside a year. The programme is not going to run it; that is not the programme's job. It is worth flagging that the argument is not a rhetorical gesture. It generates predictions.

What the reader keeps

The alignment field's own 2026 formal results imply that its own composition is a safety variable. That is the step from the PNAS Nexus paper to the field itself, and it is the argument the Polymathy paper's monoculture section makes at greater length.