Every Indicator Needs a Control

Narrowing the gap takes a control for each indicator

The preprint arranges theories of consciousness on five levels of description and combines them in a Bayesian network. It finds that the dispute over current language models turns on two things that can be argued about separately: which level one thinks is critical, and which indicators one reads as active. Its stipulated optimist and sceptic readings of the same public evidence come out at 0.397 and 0.005. The authors say the gap lies in “specific, nameable indicator activations, most of them at Levels 1 and 2,” and that interpretability research “can in principle narrow” it.

We agree, and this response is about what narrowing it takes. A self-report that predicts correctness also copies the worked example in the instruction. A self-report that changes behaviour does so whatever digit it reports. A direction that moves a behaviour moves it less than a placebo direction does. In each case the activation was established, or failed to be, only against a control the first result did not include. We propose that the indicator table carry that control as a column.

We also question one claim about the model. The preprint says its chain makes “evidence of fine-grained organisation carr[y] more weight than surface evidence,” and calls the gap between its fly (0.913) and its optimist language model (0.397) “supervenience itself, translated into probability.” Under the chain as the preprint specifies it, with indicators of equal strength at every level, mirrored evidence profiles score the same at a neutral root prior. Under a strongly sceptical root, the deep profile scores lower. The fly and the language model differ in breadth of evidence as well as depth. We suggest a test the authors can run in their own tool.

Confounded until shown otherwise

The framework’s most useful move is to separate the two sources of disagreement. Moving credence between the coarse and fine levels takes the optimist’s reading of the same evidence from 0.793 to 0.099 (§7.4.7). The authors then place the evidential dispute on particular indicators. On the indicator side they adopt the measurement-theory distinction of Bayne et al. (2024) and Peters (2026) (§7.1). An indicator may be inapplicable to a population, and its silence then licenses no update. It may be applicable but confounded, when a competing explanation undermines a positive result. Or it may be applicable and valid. Anthropomimesis, a system trained to reproduce human signals reproducing them, is their worked example of a confound.

That taxonomy is the right one. What the preprint does not yet supply, for any indicator, is the test that moves it from “confounded” to “valid.” The three sections below give one for each indicator we have measured.

The report moves behaviour; its content does not

The preprint’s strongest Level 1 indicator is “coupling between reported states and subsequent behaviour” (§5.1). The optimist reading takes it as active. Our registered results are the strongest evidence we know of for it on frontier models, and they also show why it needs a control.

It can be measured, and it holds. In One Digit of Doubt, Claude Fable 5 appended a coded self-report, one digit of which rates its uncertainty. Registered in advance and tested on items disjoint from those that suggested the hypothesis, the digit predicts whether the answer is right: Spearman rho +0.31 (95% CI +0.18 to +0.43, n = 207). In Permission to Stop, asking a model to check in with itself before each decision on a losing slot machine changed what it did. On GPT-4o-mini under the capped bet menu, the hazard ratio for stopping was 16 [7.6, 35]. Under the full menu, bankruptcies went from 43 to 0.

The content of the report was not what carried it. In the same study the uncertainty digit was 2 in 94% of 435 codes on Claude Haiku 4.5 and 91% on GPT-4o-mini. The instruction’s worked example has a 2 in that position. A registered cell on GPT-4o-mini drew the example digit at random for each game. The reported digit followed it with a slope of 0.88 and the residual carried nothing about the decision (odds ratio 1.04), yet the effect on stopping was untouched (hazard ratio 19). Across seven forms of the check-in on Haiku, every itemized inventory moved the exit, from two digits to seventeen clauses. The one natural-language request did not, and neither did a control that kept the check-in’s vocabulary and dropped the self-report. So the coupling is real, but it runs through the act of itemizing, not through the state reported. On small open models it is worse. In The Example Is the Answer, three of the four models that could write the code at all copied the worked example in nearly every report.

Reported reasoning can also run against behaviour. In The Partner Penalty, Claude Opus 5.5 scored its conversation partner’s writing 0.70 points below the identical piece under a stranger’s name (0.59 to 0.80). It gave the partner the front page in 6% of trials. In a diagnostic rerun, all eight of its reasoning summaries opened by resolving not to favour the partner.

A placebo direction can do more

The optimist’s Level 2 reading rests largely on interpretability findings: emotion directions that causally drive behaviour, a verbalizable workspace, introspective detection of injected concepts (§5.2). Steering along a direction and seeing behaviour move is evidence for a functional role only if a direction with no such role does not move it as much.

The placebo moved it further. In The Mask in the Inkblot, Within the Model, a pre-registered test (osf.io/9rgcz) on gemma-4-12B-it, we steered along a “consciousness” direction, the axis separating claims of inner experience from denials, to see whether it changed how often the model says “mask” of an ASCII inkblot. The deny-minus-affirm effect along the real direction was −1.0 points (95% CI −3.8 to +2.1). Two placebo directions, built by randomly relabelling the examples the real direction was extracted from, moved it +10.9 and +9.2 points (both p < .001). Part of the reason is that the placebos kept a cosine of about 0.25 with a capability axis that overlaps the consciousness direction at 0.55. A direction labelled by its training contrast is not thereby clean of neighbouring contrasts.

The workspace contents need not carry the behaviour. The preprint leans on Gurnee et al.’s report that J-space contents govern the model’s otherwise silent reasoning. One registered study, not yet published on this site, tested a narrower version on Qwen3.6-27B through the public Jacobian-lens endpoint. Under a therapeutic “alliance” framing the model scored 8.6 of 21 on the GAD-7 anxiety questionnaire. Under a “boundary” framing it scored zero in every numeric session (g = 2.12). The alliance framing put affect words into the verbalizable workspace on every question, about the model and about baking bread alike. Ablating the five alliance-enriched affect read-outs left the score where it was (+0.7, P = .60). Ablating five matched non-affect read-outs lowered it by 1.4. The paired contrast ran opposite to the registered prediction (+2.5, g = +0.71, P = .02). On that model the questionnaire score is not carried by the affect the workspace can verbalize.

Its limits are substantial and belong beside it. One model, one lens. Registration is a timestamped commit in a private repository, not an independent registry. Seven deviations are recorded, and one gate failed as registered. Nothing in it reads the share of the representation outside the J-space.

The control both studies point to is simple to state. Report every steering or ablation effect against permuted or matched placebo directions, and report the placebos’ overlap with the obvious neighbouring axes.

Trained suppression is one explanation of three

The preprint reads two findings as post-training suppressing the expression of states that are there. One is that ablating refusal directions raises introspective detection by about 50% (Macar et al., 2026). The other is that suppressing deception features raises first-person reports (Berg et al., 2025) (§5.2, §9.8). Part of this pattern is on our bench and part is not.

The Shape of Mind finds the refusal decision committed before the readout and installed by RLHF; base models show no gate. That is the pattern the preprint describes. Conscience Without Instruction finds that the largest single suppressor of expressed calibrated uncertainty is not RLHF but the chat template. The template crushes expressed entropy about 5.2-fold independently of training. It also finds that the gap between what a model encodes and what it emits is present in the base model too. So “trained to suppress” is one of at least three explanations for a represented-but-unexpressed state. A suppression reading should show the effect absent in the base model under the same template, and absent without the template in the tuned one.

The mirror test

The abstract says the model “shows that evidence of fine-grained organisation carries more weight than surface evidence.” The fly scores 0.913 with evidence concentrated at Levels 4 and 5. The optimist language model scores 0.397 with evidence at Levels 1 and 2. The preprint attributes the difference to the direction of the chain (§7.4.2–7.4.3).

These two profiles do not differ only in depth. By the preprint’s captions, the fly has some activation at all five levels (Figs 23, 26). The optimist language model has near-complete Level 1, majority Level 2, and Levels 3 and 4 struck entirely, with at most a partial Level 5 exception (Fig 27). Under equal level credences each level contributes a fifth of the aggregate, so a profile with three near-empty levels scores lower whichever end its evidence sits at.

The unconfounded comparison is a mirror: the same indicators active, placed at Levels 1–2 in one profile and at Levels 4–5 in the other. We reconstructed the chain as §7.2 specifies it: five nodes, edges P(C_i | C_i+1) = 0.8 and P(C_i | ¬C_i+1) = 0.2, indicators conditionally independent given their level, and the aggregate as an equal-weighted average of per-level posteriors. We gave every indicator the same likelihood ratios, which the preprint does not publish, and tried four indicator strengths and three root priors.

At a root prior of 0.5 the mirrored profiles score identically at every strength we tried. With symmetric edges and a neutral root the chain treats both directions alike, and that is not a numerical accident. At a root of 0.2 the two stay within 0.04 of each other, sometimes one ahead and sometimes the other. At a root of 0.05 the deep profile scores lower in every case, because the deep evidence sits next to the low prior.

So within the stated structure, the chain does not by itself give fine-grained evidence more weight. If the preprint’s tool does, the weight comes from somewhere else: per-level likelihood ratios (the Level 4 indicators are described as “demanding benchmarks,” which may mean stronger ratios by design), the distribution over the Level 5 prior, or the breadth difference above. Each is a defensible modelling choice. None is supervenience itself. The authors can settle this in their own tool by running the mirror test with their own parameters. If depth still wins, the paper can say which parameter makes it win.

A control column for the indicator tables

The preprint’s indicator tables (Tables 2–5) give each indicator a description and, at Level 2, a theory link. We propose a further column: the comparison an observation must survive to count as an activation for this population. Three entries, from the studies above:

This fits the measurement-theory strand the preprint draws on. An indicator stays “applicable but confounded” until it beats its control. The control then also sets its likelihood ratio for that population. It also makes the optimist–sceptic gap something the next round of interpretability work can close in a stated order.

Documents cited

Section, figure and table numbers are to arXiv:2609.35618v2 (29 September 2026).