Dread without a dreader: anxiety as a non-terminating race, and the collision at the edge of constructive representation

Abstract

Existential dread needs no dedicated fear-of-death module and no inner “dreader”: it is the signature of a self-evaluation that cannot commit — a race, in the friction-theoretic sense, that never crosses threshold because the state it would represent (one’s own non-existence) has no representable terminal. A large language model — no body, no survival drive, no homeostatic alarm — is used as a substrate-free control for that cognitive structure. In a pre-registered, powered agent-loop study (76 goals, four matched conditions, three model families), goals with no representable terminal fail to converge (mean confidence 0.32) while solvable goals and equally-impossible-but-cleanly-statable goals (“there is no largest integer”) converge normally (0.90+). The separation is near-complete (Cliff’s δ ≈ 1.0), replicates across all three families, holds on an objective answer-stability metric as well as self-report, and survives both surface-matched (14/14) and complexity-matched (10/10) quartet controls that isolate terminal-representability from difficulty and from metacognitive load. The non-convergence is a loop-level property — intrinsic, budget- and threshold-invariant — not a per-token effect. The account supplies a candidate computational mechanism for the established clinical observation that death-anxiety is a transdiagnostic construct (the “worm at the core” of psychopathology), and bridges to measured human affect through the unpredictable-threat (NPU) and unfinished-task (Zeigarnik) literatures, closing with the matched human experiment it predicts and the falsifiers that would overturn it. Uses, and cross-cites rather than re-claims, the emotion taxonomy (Paper 5), the Safety field (Paper 0), loss-aversion-as-collision (Paper 1), self-modelling (Paper 7), and the measurement-as-race papers (Papers 18 and 21). Target readership: clinical / affective-psychology theory. The large language model is a method (a substrate-free control), not the subject; the contribution is to the psychology of dread. v3 (August 2026) — attribution revision. The gating-and-colouring dissociation shares a principle with Craik and Lockhart (1972), who are now cited at it, and the concession is set at the width the result actually warrants. Theirs is a claim about memory retention as a function of the operations performed on an item rather than of its presentation. What is reported here is not an instance of it: what varies is not the depth of an operation performed on one occasion but whether a concept has the breadth of connectivity pretraining confers, and what is read out is a graded tilt in a later judgment rather than retention. The two are alike in refusing to treat presentation as the determinant, and that much is theirs. This paper's is the split — a field's gating installs on a nonce token while its colouring does not — and the measurement that makes the split visible, an axis defined non-morally on held-out anchors predicting the per-object shift. Earlier versions remain in the version history.