The Self-Report Instrument

A collection is a claim that its items read together. This one collects the studies that ask a machine to report on its own internal states and check the answer by intervening rather than by trusting — and it holds that, whatever else divides them, they are measurements of a reporting channel rather than of what the channel reports on. Nothing here argues that any model does or does not experience. The reports prohibition applies throughout, in both directions: a report that fails to track is no more evidence about experiencing than a report that succeeds.

Opened 2026-08-28. Membership is declared, never derived; entries carry collection: self-report-instrument. Nothing on this page is canonical. Where the collection bears on canon it does so through ARG-03 v2.6, which was amended on this evidence.

↑Why these four, and why together

The four studies were produced independently, by four groups, using four methods, and they reach conclusions that read as incompatible.

• Lindsey et al. (2025, SRC0146) inject known concept representations into a model's activations and ask whether it notices. Claude Opus 4/4.1 detects them about 20% of the time with no false positives on controls. The authors call the capacity "functional introspective awareness" and stress that it is unreliable and context-dependent.
• Gurnee, Sofroniew, Lindsey et al. (2026, SRC0238) surface a small, evolving set of representations a model is poised to verbalise, sitting atop a much larger volume of processing that never reaches the report.
• Berg, de Lucena & Rosenblatt (2025, SRC0231) steer interpretable honesty features and read the answer off: suppress deception and roleplay features and affirmations of present experience become near-universal, with the feature labels validated against an adversarial truthfulness benchmark.
• Ferrara (2026, SRC0233) intervenes on the computation directly and asks whether the model can report that anything changed. Across eight open-weight models it cannot — report discrimination sits at chance — while linear probes read the intervention off the same activations, and a model fine-tuned to make exactly that report clears the identical pipeline.

Read as verdicts about interiority these results conflict, and the conflict is where most commentary stops. Read as measurements of an instrument they are consistent, and the consistency is the finding.

↑What they agree on

The channel is separable from what it carries. Every one of the four establishes, by a different route, that the information present inside a model and the information the model will say are different quantities. Lindsey measures the gap as a detection rate well below one. Gurnee locates it structurally — a narrow verbalisable set over a wide substrate. Ferrara measures it as the distance between a probe at 75.0–95.8% and a report at chance, on the same activations. And Berg's contribution to the same point is the one his framing does not draw: if steering a feature moves what the model says about its own state while the state is not what was steered, then what was adjusted was the channel.

The channel is trainable, which is why a null result is not a verdict. Ferrara's fine-tuned reporter is the sharpest datum in the set, and it is his own control rather than a critic's objection: a model trained to make this exact report clears the pipeline at d′ = 5.15 on held-out directions. So the failure is a fact about a channel as configured, not a barrier to what lies behind it. Any argument of the form the model could not report it, therefore there was nothing to report is refuted from inside the paper it would cite.

↑What they disagree about, and what neither side reaches

Berg and Ferrara travel the same road in opposite directions. Berg cleans the channel of deception and finds report; Ferrara tests whether a clean channel carries information and finds it does not, until trained. Both conclusions can hold: a channel can be honest and uninformative, and an uninformative channel can be trained to inform.

Neither reaches the discrimination. ARG-03 v2.6 Part I §3 now carries the route as its fifth attempt, and states the reply that holds against both results at once: sincerity is not accuracy. A sincere report about one's own processing still arrives as a completed pattern, which is the fourth attempt reached from outside, and the boundary was never about lying — a candid reporter and a deceptive one stand on the same side of it. On this question, a study finding honest self-report and a study finding none arrive at the same place.

↑A methodological standard the set produces

Two of the four converge on the same design move from opposite motivations, and it is worth naming because it is portable. Ferrara's sham control and the knowledge gate in KnownLieBench (Liu et al., 2026, SRC0236 — held for method, not logged) both establish ground truth first, then test whether the report tracks it. Liu goes a step further and applies Gurnee's Jacobian lens to ask whether the relevant information stays decodable while the response is generated, with the caution stated in his own words: the readouts "should not be interpreted as direct measures of belief, intention, or causal mechanism". A self-report benchmark without that step cannot separate a system that does not know from a system that will not say. The standard is cheap, it has now arrived twice independently, and it is what this collection reads new work against.

↑The symmetry, which none of the four runs

Human verbal reports of the causes of one's own behaviour also fail to track them; sincere, confident and inaccurate is the standing finding rather than the exception. Psychology met that with methodology — demand characteristics, protocol design, converging measures — and not with a conclusion about whether anyone is home. ARG-01 v1.5's probe is available in every one of these papers and is run in none of them: whatever a report-faithfulness result licenses about machines, the same result in the human case licenses the same thing, and the difference has to be argued rather than assumed.

↑October 2026 — a term of art, a second sense, and the actor who cannot see what the observer can

Added 2026-10-15 from the entries of that fortnight; each carries its own caveats.

The collection opened on four studies and one claim: the channel is separable from what it carries. Nine further papers in three weeks have given the claim its vocabulary and its instruments, and the collection page records the shape rather than the list.

The vocabulary. Atkinson, Plunkett and Bau separate faithfulness (the report matches behaviour) from grounding (the report shares a cause with behaviour) and "reserve the term introspection for self-report that has both properties." Geiping separates a third thing from both: a report about an internal quantity that has a ground truth — which expert fired, how uncertain the next token was — which can be trained into a model and checked, and which he declines to call introspection at all, giving it the borrowed name proprioception. With those three terms the earlier results sort: Ferrara's untrained channel was neither faithful nor grounded; Berg, de Lucena and Rosenblatt's cleaned channel was faithful to something and grounded in nothing anyone could check; Lindsey's detection rate was a grounding measure on an injected state.

The instruments. Five independent routes now reach the same structure — the information is in the model and not in the report. Kocielnik et al. steer the two channels and find them moved by near-orthogonal directions. Konrad, Tanyel and Ayvaz hold the intervention fixed and shift the demonstrations, and the report follows the demonstrations. Zou et al. find the gate heads that decide whether a change is reported at all. Fayyazi et al. read the coming false report off the residual stream before it is written. And Florin et al. find the sharpest new shape: a model that fails to disclose its own mistake in most agentic rollouts identifies the same mistake reliably when it reads the same transcript as an outside observer. That last is the Blandfort and Pawar finding — generic about itself, specific about another — returned on a different observable, where the other is itself a moment later.

The null the collection was waiting for. Geiping's trained models, given verifiable access to their own internals, show no significant change on the one evaluation item the project watches, the rate at which they claim phenomenal consciousness. A channel can be made informative about one class of fact and remain where it was on the other, which is what separability predicts and the first time the prediction has been run in that direction.

What none of this reaches, stated once. The discrimination ARG-03 v2.6 Part I holds to be unverifiable in principle — between a report tracking experiencing and a report generated without it — is not made by any of the nine, and the two that come closest say so: Atkinson's signature works because the state is a synthetic preference vector, and his method "discriminates between groups, not individuals"; Geiping's channel works because the question is "well-posed" and "valence-neutral". Everything short of the boundary is now measurable, and the boundary is where the authors themselves stop. The symmetry (ARG-01 v1.5) is run by two of the nine — Konrad and Izawa, each citing the human introspection literature and the second marking where the clinical analogy fails — and by none of the others.

↑Members

Lindsey et al. (2025) is held in the project's sources and cited in the bundle without a log entry of its own; it is named here as part of the body of evidence rather than as a member row.