Reading "Friction" in LLM Logprobs: A Substrate-Level Correlate of Present Route-Competition in Social-Moral Judgments

Can a single token-level signal read from a language model’s log-probabilities — the count of competing routes, the commit-margin in nats, the first-token entropy — act as a readable correlate of how a model resolves a social-moral judgment? This paper reads such a signal across three experiments and delimits, carefully, what it does and does not establish: it is a substrate-level correlate of present route-competition, not an established causal mechanism and not a claim that the model feels anything. Three experiments, with controls. (E1) Third-party moral commitment is morality-specific and lexeme-robust, with a friction difference that appears in the ambiguous region where the judgment is not pre-resolved. (E2) Content-invariant binary rule-evaluation generalises up an abstraction gradient at capability — a capability control that bounds the signal, not a social-moral signature. (E3) A disadvantaged agent’s disengagement is driven by comparison and expectation rather than reward amount, with a human-like advantaged/disadvantaged asymmetry; and — read internally rather than at the output token — the inequity reaction is strongly predicted by a non-morally-defined internal disparity axis (a gap-controlled, context-general dose-response, made non-circular by construction), consistent with a fairness reaction that imports another agent’s unresolved competition rather than reciting a learned moral script. Every contrast carries a bootstrap confidence interval; a lexeme-invariance check distinguishes robust effects from answer-token artefacts; and the limits (two models, pilot scale, a correlational interpretability readout for the internal result) are stated rather than hidden. Companion papers in the series develop the underlying friction theory, the forward-modelling and operational accounts, the competing-routes measurement-model programme, and the mechanism home for the mirror-friction reading of fairness. Prepared for submission to Transactions on Machine Learning Research (TMLR). Data and code. The stimulus generators, probes, re-analysis scripts, and per-token log-probability outputs are available from the author. v2 (August 2026) — attribution and ownership revision. The method behind the fairness literature this paper concedes is now credited where the concession is made: running a language model as an experimental economic agent is Aher, Arriaga & Kalai (2022) and Horton (2023), and Experiment 3 uses that method rather than proposing it. The instrument's own lineage is stated in the method section — reading quantities off the output distribution as a correlate of a model's judgement is established (Guo et al., 2017; Kadavath et al., 2022), with semantic entropy (Kuhn, Gal & Farquhar, 2023) its strongest current form — and what remains this paper's is the application: reading those quantities at the commit point of a social-moral judgement, against matched contrasts built so the distributional reading and the stated answer can come apart. Experiment 1's agent-versus-nature contrast is credited to the moral/conventional distinction (Turiel, 1983). In the other direction, the appendix field-axis result is no longer written as a borrowed technique reported for completeness. The projection method is a companion's; the held-out, non-morally-defined axis that makes the test non-circular, and the dose result itself, are this paper's, and they are the strongest evidence here against a mimicry reading. Editorial pass. Earlier versions remain in the version history.