Theory as Policy

A theory’s job, Thomson wrote in his 1907 Corpuscular Theory of Matter, is “to connect or co-ordinate apparently diverse phenomena, and above all to suggest, stimulate and direct experiment.” A policy is evaluated by its usefulness and held only as long as it is useful; a creed is believed, defended, and mourned when overturned. Thomson modelled the distinction in his own career: he held the electron with total certainty because the experiments left “no escape from the conclusion,” and he held the plum-pudding atom lightly, as a provisional co-ordination of the evidence, which is exactly why he could abandon it without damage to his standing or his integrity when Rutherford’s foil demolished it. The distinction maps directly onto the failure mode of large language models: they produce outputs in the grammar of creeds—confident declarative assertions—while the epistemics of those outputs are, at best, those of a provisional policy. A model that speaks in creeds while generating