Mount Stupid in the machine: how evidence competition explains the Dunning-Kruger curve in a language model

The Dunning-Kruger “valley of despair” is, in a language model, a near-tie of about 0.4 nats in the substrate’s balance-of-evidence margin that greedy decoding renders as full stated confidence — and a validated log-probability proxy reads the latent decision variable directly, beating a fitted scalar-surprisal model (ΔAIC +54, pooled n = 749), including a matched-surprisal dissociation (Δ = +1.44) to which the field’s scalar measure is blind. The Dunning-Kruger pattern is among the most cited and most contested findings in metacognition, and the dispute is largely about measurement: the load-bearing quantity — how strongly a person’s competing answers compete before commitment — is never observed directly, but reconstructed from confidence and accuracy scores entangled with regression-to-the-mean and scale-use. This paper changes substrate. In a large language model that same quantity is read directly from the next-token probability distribution. That language models can display Dunning-Kruger-like overconfidence has been reported independently, as a calibration contrast between models (Ghosh & Panday, 2026, arXiv:2603.09985); no priority is claimed for the phenomenon itself. That report is treated as warrant rather than as a finding of this paper: an independently documented pattern is what makes the explanatory question worth asking, and what this paper supplies is the mechanism — within a single model, along a competence trajectory, in a variable read directly off the substrate. The reading is first validated as a proxy for the sequential-sampling (race / leaky-competing-accumulator) decision variable; the proxy is then used as a model organism for the Dunning-Kruger curve, yielding a partial reproduction with principled divergences. The confident-before- competent rise, the recognition near-tie, and the recalibration slope are architectural (and the rise plus recognition replicate on a second, structurally different hidden-rule task). Novice humility is absent under forced commitment but restored by an abstention channel; the affective valley is absent in behaviour yet present in the substrate margin. The paper thereby localises which phases of the human curve are properties of bounded competing-route computation and which require the biological substrate, and closes with the preregistered human study that would test the analogy directly. Companion papers (Friction Theory series): Behavioural Friction Theory (Paper 0); Friction as the Cost of Probabilistic Computation (Paper 1); Logic as Reactance (Paper 14); The Physics of Learning (Paper 16); Compete, Don’t Erase (knowledge-editing erase-vs-mask); An LLM as a Controllable, Fully-Inspectable Model of Measurement; The Delta (LLM as a subtraction-control for human neuroscience); and Nature and Nurture in a Language Model (installable value fields and intrinsic capacity). Data, prompts, scripts, and per-item outputs are released with the paper (see the Data and code availability section). Part of the Friction Theory research programme; companion papers above. New in version 2. A worked application of the measurement model: self-efficacy — a capability judgement studied in people — is, in the substrate, a capacity-gated readout of the same pre-commitment competing-routes race. The friction that degrades an answer and the signal that reports confidence in it are demonstrably one quantity, and they decouple only in smaller models, which keep the degrader but cannot read the reporter; the section is positioned head-on against the human confidence-readout and self-efficacy literatures. A per-field generalisation is added as an explicit, falsifiable prediction. v3 (August 2026) — related-work revision. The paper now engages the literature that reframed overconfidence. Moore and Healy (2008) separate overestimation of one's own performance, overplacement relative to others, and overprecision, and show the first two reverse by task difficulty — a rival account of the canonical curve that requires no metacognitive deficit, and the objection a reader of that literature has ready. It is stated alongside the older calibration work (Lichtenstein & Fischhoff, 1977). A provenance correction: the four popularised landmarks — novice humility, the peak, the valley, the slope — are not from Kruger and Dunning, who report a monotone relation between quartile and calibration error. They belong to the folk four-stages-of-competence model. Section 1 says so and uses the arc as a description of what the public curve claims rather than as a finding to reproduce, which is what the paper was already doing. Editorial pass. Earlier versions remain in the version history.