Reading the substrate, not the score: a self-calibrating friction signature that warns of fine-tuning over-fitting and maps when interventions help

Early warning of a training-time collapse has so far been read from a model’s internals. This paper shows the same precursor is legible in the model’s own output distribution: per-token entropy variance and lag-1 autocorrelation, computed from the log-probabilities any inference interface already returns, rise before held-out accuracy drops when narrow fine-tuning begins to erode general reasoning. The two precursors replicate across three model families (Qwen2.5-1.5B, Phi-3.5-mini, SmolLM2-1.7B) with multi-seed error bars, and the lead over the accuracy drop runs from several optimizer steps to coincident depending on the model. The signature is packaged as a self-calibrating monitor whose trigger is relative to each run’s own opening baseline, so there is no global threshold and no separate calibration pass. It flags the collapse in three of three families that collapse, while a difficulty-masking gate correctly skips a fourth. The load-bearing contrast is that the failure is capability-erosion rather than over-fit to the training task: the in-task validation loss falls straight through the collapse and never warns, and a ceiling-masked accuracy probe never moves, while the friction signature rises through the transition. An estimator ablation shows the advantage of the distributional moments over a scalar entropy level is transfer rather than lead — the level matches them on two families and fails on the third. Read across model capacity instead of training time, the same signature yields a regime-map that predicts when an instruction frame helps: it bites on pressed substrates and does nothing on frozen ones. A published account of entropy collapse proves it a first-order transition for feedback-amplified systems, across which variance and autocorrelation stay finite, so that no such precursor can exist there. That result and this one concern different system classes: its defining feature, the model’s own output re-entering its input, is absent from supervised fine-tuning on a fixed external corpus. The paper states as an explicitly untested conjecture that the signature should fail on recursive self-training, and names the experiment that would decide it. The monitor’s statistical construction is inherited rather than proposed: it is a control chart with a baseline-relative limit, the Phase-I / Phase-II practice of statistical process control and change detection. What is specific here is the observable — a friction quantity read from the model’s own output distribution rather than a task metric or an internal state — and the honest limits that follow: the threshold is set to a ceiling on benign drift rather than to a false-alarm rate, and a panel of never-collapsing runs is named as the validation that would supply the guarantee sequential methods provide. Target venue: Transactions on Machine Learning Research. Data, notebooks, and analysis scripts are released with the paper. Part of the Friction Theory research programme; companion papers are linked in the related identifiers. New in version 2. The empirical results are unchanged; the revision is one of positioning. Version 1 shipped without a related-work section. This version adds one covering six literatures, narrows the novelty claim from the early-warning phenomenon (which prior work established) to the observable (that the precursor is legible in the output distribution, hence deployable without weight access), and reconciles the regime-map against its incumbents. It also positions the monitor against the statistical field that owns its construction — sequential change detection and control charts — conceding that the construction is inherited and locating the specific contribution in the observable. An estimator ablation on the existing trajectories, added here, reports the distributional moments as buying transfer across families rather than a longer lead. Reviewed by external dual peer-style review of the revision. Builds on v1. v3 (August 2026) — scope and provenance revision. The claim that the signature leads the accuracy score is now tied to the channel and the model family it was measured in, rather than stated of the signature as a whole. The zero lead of the deployed variance trigger is a property of Qwen2.5-1.5B; Phi-3.5-mini and SmolLM2-1.7B are named with their own leads, and the appendix and discussion are scoped the same way. Every figure now carries the run it was computed from, and a new paragraph states what that labelling made visible: the same model's collapse onset moves by several steps across run-sets differing in seed, budget, probe filter and scan cadence — a spread larger than most of the leads being claimed. Each step-lead is therefore restricted to its own run-set, and a lead of about one scan interval is to be read as such rather than as a horizon to plan against. A checkpoint-versus-step unit error and two ambiguities about onset are corrected. The estimators are placed in the tradition that owns them: reading the spread of a choice distribution as an unresolved decision is the accumulator literature — Luce (1959) for the choice axiom the softmax implements, Ratcliff (1978) for evidence accumulated to a bound, and Usher and McClelland (2001) for the leaky competing accumulator — and the friction signature is a family of estimators of that quantity computed on a language model's output distribution. Companion references now carry their published identifiers. Earlier versions remain in the version history.