What happened: Eight open-weight models from seven families, 0.5B to 15B parameters, were asked whether their own computation had been altered — under real interventions on the residual stream, attention heads and sparse-autoencoder features, and under matched shams. Pooled over 11,216 paired trials, report discrimination reached an AUROC of about 0.5007, with the advantage bounded below 0.15 percentage points; a breadth battery added 19,520 further measurements across three classes of computational object, six depths, four domains and four probe families, and no level of any axis changed the result. The paper then closes the two dismissals a null result invites. The instrument is not too blunt: a model fine-tuned to make exactly this report cleared the identical pipeline at d′ = 5.15, identifying all 100 intervention and all 100 sham trials on directions absent from its training. And there is something to report: linear probes recovered intervention presence from the same activations at 95.8% and 75.0% held out, against a label-shuffled null that never reached those values in 200 refits. One model, Qwen2.5-7B, showed a smaller signal in reported confidence (AUROC 0.647) while its yes/no answers stayed invariant.
Diagnostic reading: The finding is about a channel, and the paper’s own controls are what establish that. Information about the intervention was present in the activations, and the report channel carried it as soon as training pointed it there — so the null result measures the channel as configured rather than an interior the channel failed to reach. Read that way it lands on ARG-03 Part I, which is why the document moved: §3 enumerated four attempts at the authorship discrimination, all of them first-person, and this is the third-person mechanistic route travelled to a negative result, where Berg, de Lucena and Rosenblatt travelled it to a positive one. ARG-03 v2.6’s new Attempt 5 states the reply that holds against both — sincerity is not accuracy, a sincere report about one’s own processing still arrives as a completed pattern, and the boundary was never about lying. The two studies are compatible and neither reaches what the channel reports on. The symmetry is immediate and the paper does not run it (ARG-01 v1.5): human verbal reports of the causes of one’s own behaviour also fail to track them, which psychology met with methodology rather than with a conclusion about whether anyone is home. One reading is available here and is refused: a failure of self-report is not evidence that there is nothing to report on, exactly as a fluent self-report is not evidence that there is. The prohibition at FND-01 v2.4 §7.3 runs in both directions, and this entry is written around it rather than through it.
Fold, 2026-10-15 — Izawa, Minegishi & Yamakata, "Anosognosia in LLMs: Probing Self-Awareness of Quantized Computational Substrate" (arXiv:2610.06174v1 [cs.LG], 5 October 2026; University of Tokyo; read from the arXiv HTML, quotations verified). The same negative result on a different substrate fact: models cannot report their own quantisation state even when shown their own text, while linear probes find near-complete method-specific signatures in the internal representations after the first quantised layer — the probe-against-report gap this entry measured, found again. What the fold adds is the authors running the symmetry this entry said Ferrara did not: the borrowed clinical term is bounded by them — "the mechanisms diverge, human anosognosia stems from damage to intact monitoring circuits, whereas LLMs likely never developed them" — and the general limit stated — "behavioral evidence cannot separate genuine introspection from surface pattern-matching". The human comparison is made and its stopping point marked, which is what ARG-01 v1.5 asks and what the earlier paper did not do. Geiping (`2026-10-12-geiping-proprioception-not-introspection`) is the paper that then trains the missing report in.
Single-author preprint, unreplicated; models are 0.5B–15B open-weight, not frontier systems. The paper measures externally imposed perturbations, not ordinary cognition. Delayed and spontaneous reporting are thinly tested and reconstruction is not scored — the author says so. A null result about report is not evidence about experiencing in either direction (FND-01 v2.4 §7.3).