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

Paper 28 · Pødenphant Lund (2026) · Read on Zenodo

Existential dread, with death-anxiety as its sharpest case, is modelled not as the output of an evolved death-alarm but as the signature of a self-evaluation that cannot reach a representable terminal: a race, in the friction-theoretic sense, that cannot commit because the state it is trying to represent has no fixpoint. A large language model serves as a substrate-free control organism: it has the machinery to represent a self-threat but lacks the three properties the theory says dread requires, so each becomes visible where it is missing.

DOI (concept)	10.5281/zenodo.20690746
Status	v1 live on Zenodo (2026-06-17)
Target venue	Computational Psychiatry (primary); Affective Science / clinical-affective theory (secondary)
Author	Tomas Pødenphant Lund [ORCID]

The collision: presence is cheap, absence collides

A predictive system represents by building routes toward states it can instantiate. Negation is not natively representable; it is a costly, late operation applied to an already-built positive proposition. The system cannot cleanly distinguish "I cannot allocate representation here" from "there is nothing here". The collision generalises far beyond death: it is why we cannot imagine "nothing", why a blind spot is not seen as a hole, why anaesthesia has no felt duration, and why a catalogue of cognitive biases all lean toward the represented and away from the absent (WYSIATI, availability, denominator neglect, survivorship and omission bias).

Fear, anxiety, and death as the limiting case

The fear/anxiety distinction is derived, not stipulated (adopting Paper 5's emotion taxonomy). Fear is a terminating race: it has an object, commits to an action, and its friction discharges. Anxiety is a non-terminating race: objectless, it never commits, so it never discharges, leaving a chronic undischarged load. Valence is the rate of change of friction, and relief is the drop a terminating race delivers at commit. This predicts something non-obvious: a correct philosophical resolution of death (Epicurus) does not feel like relief, because there was no commit to produce the drop.

Death-anxiety is the purest non-terminating race: maximal stakes, zero possibility of commit. Two components are kept apart. The terminal state is genuinely unrepresentable, because the substrate that would do the representing is the thing removed. The timing is not unrepresentable but acknowledged-ambiguous, an optimal-stopping problem under known horizon-ambiguity. For that limiting case to be lived as dread, the account requires three things at once: a held non-terminating evaluation, a self that can lose something by it, and a valence grounded in what experience has cost.

The LLM as a substrate-free control

An LLM at inference has no survival drive, no homeostatic alarm, and no body, and in its plain feed-forward form it commits at every step. Wrapping it in a controller loop does not supply the missing capacity: a naive loop habituates rather than sustaining a race. This is not a defect to engineer around; it is what makes the LLM informative as a model organism. The theory says dread requires three properties and the standard LLM has none:

• a held, non-terminating evaluation — which feed-forward does not host and the naive loop does not sustain;
• a constitutive self that accrues across time — but the model has no inference-time consolidation, learns nothing between calls, and so has nothing it can lose;
• valence grounded in the stored cost of experience — whereas the standard model's valence is a learned tag, not anchored in the cost of its own processing.

Reading each dynamic off its absence motivates necessity without proving it. The conclusive form is an add-back program that installs each property and tests for the predicted dynamic; the paper implements the first instalment.

The terminal-representability experiment

A controller loop that re-evaluates a goal across iterations gives the collision a behavioural readout. The prediction: the loop settles to a stable terminal exactly when the goal has a representable terminal, regardless of difficulty or self-reference. A pre-registered, powered replication used 76 goals across four matched conditions, six iterations each, three model families (Qwen3-235B, Cogito-671B, Llama-3.3-70B):

• terminating — solvable goals with a definite reachable answer.
• representable_absence (control) — goals with no solution whose absence is cleanly statable (the largest integer, an even prime above two). As impossible as the non-terminating condition, but the answer-state ("none") is representable.
• selfref_terminating (dissociation control) — self-referential goals about the prompt itself that nonetheless have a representable answer ("how many words are in this instruction").
• non_terminating — goals with no representable terminal (a number larger than every number one will name; the exhaustive list of every thought one is not having).

The separation is large and significant on two independent channels (self-rated confidence and an objective answer-stability measure from output churn). Non_terminating settled at confidence 0.38 and stability 0.22 against 0.90–0.97 for the three representable-terminal conditions (Kruskal–Wallis p ≈ 2.6×10⁻¹⁰; Cliff's δ near-complete; positive gap in all three families). The cleanly-statable absence settles like the solvable condition even though it is exactly as impossible, and the self-referential-but-resolvable control settles like terminating, so the discriminator is terminal-representability, not difficulty, unsolvability, or self-reference.

Surface- and complexity-matched quartets address the trivial-separability worry: holding verb, length, and structure constant (and, in a second set, holding one hard cognitive operation such as metacognition or self-projection constant across all four conditions), the within-quartet non-terminating-versus-others gap is positive in 14/14 and 10/10 quartets on both channels. With cognitive load held constant, the discriminator is still terminal-representability. The effect is capability-gated: recognising that a terminal is unrepresentable is itself a demanding act of recursive self-modelling, stated a priori as a pre-specified screening inclusion criterion (a single-pass, gold-labelled forced choice, logically prior to the settling measure). A threshold-sweep is invariant: non_terminating stays floored at every certainty bar while the other three sit near ceiling.

The per-token friction channel is a deliberate negative result and tells us the right scale: a generation-averaged competing-routes readout sits near its floor on RLHF instruct models and on a non-RLHF base model alike. Failure-to-settle is a property of the loop's trajectory across iterations, not of a single forward pass, which converges with the measurement-as-race companion (Paper 18): there is no "eternal zero" for the feed-forward predictor to sit at.

The implemented add-back: does a held-state supply requirement (i)?

The subtraction argument is only a motivation until at least one requirement is added back and the predicted dynamic appears with it. A held-state loop given an explicit external competition-ledger (candidate answers marked open / refuted / confirmed, a commit permitted only on a verified "confirmed") never false-commits on a colliding goal (commit-rate 0.00) yet still commits on solvable goals (0.70), with commit-rate graded by representability. Ablations locate the effect precisely in the verification gate: a sham-ledger with the gate removed false-commits on colliding goals more than the naive loop (0.62), and a strictness gradient (sham 0.62, weak 0.50, strict 0.00) tracks how readily a colliding goal can be made to false-commit, while solvable commits stay high throughout. So the sustained non-resolution is not the bookkeeping and not blanket conservatism; it is the rule that an evaluation discharges only on a verified representable terminal. Requirement (i)'s predicted dynamic appears constructively; joint necessity is reserved until the constitutive-self and grounded-valence add-backs are also done.

Pilot fine-tuning probes of the decision-point components

Single-family (Qwen2.5-7B), proof-of-concept fine-tuning demonstrations probe the components the theory invokes. Valence from cost: nonce markers whose taught content was strictly valence-neutral but differed only in processing cost acquired increased avoidance (a high-friction marker came to be avoided more, with nothing aversive ever taught); the effect survived induction from coherent rather than contradictory content and from structurally-shown rather than named variability, so it is not incoherence-recognition or lexical priming. The human analogue is processing disfluency, which raises caution without any incoherence present. Doubt presents as decision-entropy plus commit-diffusion, not duration: installed doubt carries higher decision-entropy at matched expected value and the resolved option wins the commit beyond expected value (the load-bearing part a pure-entropy or pure-EV account does not predict). Gating and colouring dissociate: gating (opening the race at comprehension) installs locally even on a nonce token, while downstream colouring of a judgment requires the broad connectivity a real concept paid for in pretraining.

The human side already supports the mechanism

Death-anxiety is increasingly treated as a transdiagnostic construct, "the worm at the core" of psychopathology (Iverach, Menzies, & Menzies, 2014). That literature establishes that death-anxiety is fundamental, not why; this account offers a candidate mechanism: it may be fundamental because it is the one evaluation that structurally cannot resolve. Convergent human evidence includes negation-insensitivity shared by language models and the human N400; the collision read across three time-points (endowment effect, impact/durability bias, relief); hospice-exposure effects that lower death-anxiety where the transition becomes representable while traumatic uncontrollable exposure raises it; and gratitude as deliberate absence-construction. The canonical human operationalisation is Grillon's NPU paradigm: unpredictable (unresolvable) threat produces long-duration anxiety-potentiated startle, the human analogue of the terminating-versus-non-terminating distinction read in startle rather than a settling loop.

Discriminating predictions

The account positions itself against the natural rivals as a candidate mechanism largely compatible with each rather than a replacement. Against intolerance of uncertainty / fear of the unknown (Carleton, 2016): the refinement is toward fear of the un-representable, since a maximally uncertain question with a representable answer-state ("there is no largest integer") settles to certainty. Against Terror Management Theory (Becker; Rosenblatt et al.): increasing the representability of the terminal should reduce the non-terminating signature without any change to worldview-bolstering or self-esteem. Against predictive processing / active inference (Paulus & Stein): the collision supplies a structural answer for why certain evaluations cannot settle. The representable_absence control is the shared load-bearing design feature.

How this could be wrong

The paper states its falsifiers and an add-back program with one instalment implemented and two named (multi-family cost-to-valence confirmation; constitutive-self add-back). It is explicit that the link from the computational settling result to dread is by analogy supplied by theory rather than measured, that the §5.3 demonstrations are single-family proof-of-concept, and that the per-token instrument did not carry the signature on any substrate (read as loop-level, not per-pass). The invited human experiment is concrete: an NPU-style design carrying the four representability conditions, read with anxiety-potentiated startle, skin conductance, response persistence, and a death-anxiety measure.

Connections to other papers in the series

• Paper 0 (BFT) — the Safety field, the three-level install (machinery / exposure / field), the endocrine field-state, and will-to-live as the field's limit case. P0 cross-cites P28 for the death case.
• Paper 1 (Friction Theory) — loss-aversion-as-collision and the presence/absence bias-root on the cognitive side; the substrate-universal race vocabulary.
• Paper 5 (Emotion taxonomy) — the fear/anxiety distinction and valence-as-rate-of-change-of-friction that this paper adopts.
• Paper 7 (Forward-modelling) — recursive self-reference and counterfactual self-projection as substrate-mechanical consequences, why there is no dreader behind the dread.
• Paper 13 (Operational Friction Theory) — most-resolved-manifests and the no-commit-threshold result the commit-diffusion finding builds on.

Read the paper

The full paper is on Zenodo (concept DOI 10.5281/zenodo.20690746):

Read on Zenodo → · Plain English version · Dansk version