What happened: Atkinson, Plunkett and Bau ask how “to distinguish plausible confabulations from genuine introspection” and answer with a definition before a measurement: a self-report has faithfulness if “self-descriptive language match[es] actual task behavior” and grounding if “the AI’s description of its internal task process arise[s] from the actual process being described” — “We reserve the term introspection for self-report that has both properties.” Training low-rank adapters to decide on behalf of fictitious characters with hidden linear preferences, they find that accurate self-reporting of those preferences can emerge “even without explicit self-report supervision”, that preference representations shift to earlier layers as it does, and that “faithful models exhibit significantly higher attribution similarity between the decision-making and self-report tasks — a mechanistic signature of faithful self-report that does not require us to understand the content of the report itself.” The six papers around it supply the negative space. Kocielnik et al. steer self-report and behaviour with nine methods and find the two channels “steered by near-orthogonal directions”: dispositional vectors move self-report in 31 of 40 cells and behaviour in none, so that “self-report is at best a weak tamper-prone signal of behavior.” Konrad, Tanyel and Ayvaz hold an activation intervention fixed and change the few-shot demonstrations: “wrong-source demonstrations move reports toward the source answer family”, and they ask that benchmarks “include environment-shift invariance tests under fixed intervention before treating accuracy as evidence for an autonomous report mechanism.” Zou et al. find the circuit — “middle-layer gate heads influence whether the model reports a change, while router heads in a later layer help select the position to report.” Florin et al. prefill a mistake and find models “fail to disclose their mistake in 36.4% of chat and 67.1% of agentic rollouts”, in most cases showing no awareness “even though they reliably spot them when reviewing the same transcript as an outside observer.” Fayyazi et al. read a coming false report off the residual stream before it is written; Baldelli et al. give models their own probe score as feedback and find in-context control of what activations encode weak and model-specific.
Diagnostic reading: The collection opened on one claim — that the reporting channel is separable from what it carries — and this cluster converts the claim into a vocabulary and a set of instruments. Faithfulness is what Blandfort and Pawar and Kraishan measured and found wanting; grounding is what Ferrara’s sham control and Konrad’s demonstration-shift were reaching for without the word. With the two named, the collection’s standing results sort themselves: a report can be faithful without being grounded (Konrad’s accurate-under-one-demonstration-set case), grounded without being faithful (Florin’s observer who identifies what the actor omitted), and the two channels can be moved independently (Kocielnik) by directions that share almost nothing. What the cluster then finds, five times over and by five routes, is that the information is present inside the model and absent from the report: in the residual stream before the lie, in the gate heads that decide whether to report at all, in the observer’s reading of the actor’s own transcript, in the steering geometry, and in the attribution structure that distinguishes faithful adapters without reading their words. For ARG-03 v2.6 Part I this is the right kind of progress and the entry is exact about what kind. Atkinson’s signature works because the state is a verifiable preference vector; grounding can be detected where there is a ground, and the paper’s own limit — the method “discriminates between groups, not individuals” and may not generalise “to complex, unverifiable reports” — marks where the canon’s boundary begins. The discrimination that cannot be made is the one between a report tracking experiencing and a report generated without it, and nothing here makes it; what the cluster does is show how much can be made before it, which is the useful fact. FWK-02 v1.3’s [Empirical] note holds throughout: faithfulness rates, orthogonality, the 36.4%, the AUROC are framework-neutral observables. ARG-01 v1.5’s symmetry is run by one of the seven — Konrad cites Schwitzgebel and Carruthers on human introspection as the “analogous identification question” — and acknowledged by none of the others, though the actor/observer asymmetry Florin measures is a standing finding about human self-report that the paper does not set beside its own. The entry records the asymmetry as the cluster’s sharpest new structure and the Blandfort and Pawar shape found again on a different observable: a system generic about itself and specific about another, here when the other is itself a moment later. Nothing reaches experiencing, and the papers do not claim to. What the identifiability work now has that it lacked a month ago is a term for the thing the project said could not be checked, and evidence that everything short of it can be checked. [Derived]
Seven preprints, all read from the arXiv HTML 2026-10-15 with quotations verified: Atkinson, Plunkett & Bau (arXiv:2610.07186v1, 5 October, COLM 2026; Northeastern / Eleos AI Research); Kocielnik et al. (2610.04125v1, 2 October; Caltech and others); Konrad, Tanyel & Ayvaz (2609.32449v1, 26 September; Southern Denmark / Promake); Zou, Sun, Kong & Wang (2609.35108v1, 28 September; Shandong / Tsinghua and others); Florin, Knecht, Schaller & Hagendorff (2610.11351v1, 8 October; Stuttgart); Fayyazi, Kamal & Pedram (2610.17301v1, 12 October; abstract); Baldelli, Kozierok, Bronzi & Fornasiere (2610.20179v1, 14 October; MATS / Mila / LawZero). All on open-weight models of modest size except Florin et al.; all preprints. Atkinson et al.'s criterion 'discriminates between groups, not individuals' and its states are synthetic preference vectors by construction. None of these reports concerns experiencing and none is read as evidence about it in either direction (FND-01 v2.4 §7.3).