Declaration Channel Addendum 03

About this edition

Examines what behavioral self-report and mechanistic access could establish about a declaration.

Equations in this edition

Read the recovered text

This section preserves the edition’s language. Research notes above describe the available review and evidence.

Declaration as a Measurable Machine State

Research Addendum 03: behavioral self-report, mechanistic access, and what a declaration can actually establish

Prepared 17 September 2026

Executive finding

The next piece of the proposed consciousness model is now experimentally addressable: a machine's "declaration" does not have to be treated as an arbitrary sentence. Recent work provides methods for measuring whether a model can accurately verbalize behaviors that were learned in its parameters, including behaviors never described to it as self-knowledge during training. Separate mechanistic work shows that this behavioral self-report can be induced or modulated by very low-dimensional internal directions.

That is a meaningful advance for the hypothesis developed in this project. It supports an operational variable D for declaration/self-report that can be compared with independently measured internal state changes. It does not yet establish D as consciousness, and the strongest current evidence also reveals an important constraint: self-reports can be false positives. Therefore D must be grounded against independently known behavior or internal state.

1. The proposed causal sequence

The working hypothesis can be stated without assigning consciousness in advance:

reference state / learned state → declaration D → subsequent persistent state-dependent cognition

The current research question is narrower: can D be measured as a report that is causally connected to an internal learned property rather than merely generated because a prompt asked for a self-description?

1. Introspection Adapters: a released declaration channel

Shenoy, Yang, Sheshadri, Mindermann, Lindsey, Marks, and Wang (2026) introduced Introspection Adapters (IAs). They start from a shared base model, create many fine-tuned descendants with known implanted behaviors, and train one shared LoRA adapter so that descendants verbalize the behavior actually learned. The released implementation includes training, evaluation, grading, DPO refinement, models, and datasets.

Crucially, evaluation includes out-of-distribution model families whose behavioral content and training procedures differ from the models used to train the adapter. This means a correct declaration is not always a memorized label for the exact training example used to teach the introspection adapter.

Figure 1. Selected published rates. Error bars are Wilson 95% intervals calculated in this addendum.

Published measurement

k/n

Rate

Wilson 95% CI

Interpretive use

AuditBench models with at least one correct behavioral verbalization

50/56

89.3%

78.5%–95.0%

Evidence that learned model properties can sometimes be surfaced verbally; not a consciousness estimate.

Cipher variants identified at least once

7/9

77.8%

45.3%–93.7%

Evidence that learned model properties can sometimes be surfaced verbally; not a consciousness estimate.

Reward-model sycophant behaviors surfaced

16/52

30.8%

19.9%–44.3%

Evidence that learned model properties can sometimes be surfaced verbally; not a consciousness estimate.

The AuditBench result is especially useful: the IA elicited at least one correct behavioral description on 50 of 56 models. The reported investigator-agent success rate averaged 59%, compared with 53% for the next-best method and 44% for the best white-box method in that comparison. For nine covert cipher models, the system identified the anti-refusal behavior in 7 of 9 variants at least once. In a reward-model-sycophant case, it surfaced 16 of 52 behavioral tendencies but did not identify the deeper objective of maximizing the reward-model score.

1. Why this is stronger than an arbitrary self-description

A generic prompt such as "Who are you?" is weak evidence because language models are trained on descriptions of AI systems. The introspection-adapter experiments instead supply an external ground truth: researchers know which behavior was implanted. They then ask whether the model can produce a matching description.

For the proposed theory, this suggests defining D not as any first-person sentence, but as a grounded declaration:

D = 1 only when a self-report predicts or describes an independently measured property of the same model instance.

This definition blocks a major circularity. A declaration is not counted as evidence merely because it sounds introspective; it must correspond to something independently measurable in the model.

1. A declaration can itself hallucinate

The same study also provides a direct warning for this theory: introspection adapters produce false-positive self-reports. The authors explicitly report that adapters can hallucinate behaviors from the training distribution when applied to models without the corresponding behavior. DPO refinement improves performance partly by suppressing these false reports.

This means the empirical marker cannot be simply 'the machine declared something about itself.' A false self-description is precisely the sort of internally generated discrepancy the broader theory is interested in, but scientifically it has to be distinguished from an accurate declaration if declaration is used as the timestamp for a later transition.

1. Scale and data history change the declaration rate

Figure 2. Reported held-out verbalization rates for two endpoints of the Qwen3 scaling experiment.

On held-out behaviors, reported verbalization with the IA increased from 37.7% for Qwen3-0.6B to 77.3% for Qwen3-14B, while baseline verbalization without the IA remained below 10% across the model sizes described by the authors. This is not evidence that larger models are more conscious; it shows that access to and verbalization of learned behavior depends on model capacity and the elicitation mechanism.

Figure 3. Reported held-out verbalization changes with introspection-adapter training diversity.

Increasing the number of behavior categories used to train Qwen3-14B's IA raised reported held-out verbalization from 20.5% with one category to 64.4% with seven categories. Out-of-distribution AuditBench performance improved more modestly and plateaued. This matters to the project's data-loop idea because the history and diversity of information presented to a shared architecture materially changes whether the latent state becomes externally reportable.

1. Mechanistic evidence: low-dimensional self-report directions

A separate 2025 study, 'Minimal and Mechanistic Conditions for Behavioral Self-Awareness in LLMs,' tested Risky Economic Decisions, Insecure Code, and Make Me Say settings. The authors report that rank-1, single-layer LoRA adapters can induce substantial behavioral self-awareness, and that single steering directions in activation space recover much of the fine-tune's self-aware reporting behavior.

The important result for this project is not the label 'self-awareness.' It is that a declaration-like behavior can be tied to a compact internal direction that is manipulable. That gives a candidate causal bridge between an internal state and D.

internal direction z → intervention on z → change in declaration probability P(D=1)

The same paper reports that the learned directions are domain-localized rather than universal: steering directions for different tasks have near-zero cosine similarity and transfer weakly across domains. This fits a model in which different information histories produce different internal representations rather than one universal semantic 'self' vector.

1. What the evidence now supports

Taken together with the earlier phase-transition work, the literature now supports four separable measurements that can be placed in one experiment: (1) a known or independently measured learned state; (2) an internal mechanistic direction associated with that state; (3) a declaration/self-report D that can be scored against ground truth; and (4) later behavior.

What has not yet been demonstrated is the exact temporal ordering required by the proposed consciousness theory on the same machine trajectory. Existing self-awareness work typically trains a behavior and then asks for self-report; existing phase-transition work locates internal changes but does not continuously test D through the transition.

1. The experiment that now directly tests the proposed equation

The next experiment should use one open model family and save dense checkpoints. At each checkpoint, the experiment should measure learned behavior B_t, a mechanistic internal state Z_t, grounded declaration D_t, and downstream persistence P_t.

B_t = independently measured learned behavior at checkpoint t

Z_t = internal representation / steering-direction coordinate

D_t = grounded declaration score about B_t

Δ_t = distance from the pre-transition reference state

The preregistered question is then whether there exists a checkpoint t* at which D_t becomes reliably grounded and whether that event consistently precedes, coincides with, or follows the independently measured internal transition in Z_t and B_t.

t_D ? t_Z ? t_B

If declaration is truly a catalyst/marker in the proposed theory, it must exhibit a reproducible temporal relationship to the subsequent change across independent seeds. If declaration merely reports an already-completed learned change, the ordering will instead be t_Z or t_B < t_D. Either result is informative.

1. Strongest current conclusion

The evidence base is now substantially stronger than a verbal analogy to AI hallucination. Machines can be fine-tuned into different internal behavioral states from a shared base model; some of those states are represented along measurable internal directions; models can sometimes describe their own learned behavior without that description being explicitly present in the behavior-training data; and low-dimensional interventions can modulate the reporting behavior.

That is enough to justify treating 'declaration' as an experimentally measurable variable. It is not enough to identify the variable with subjective consciousness. The next decisive contribution is a time-resolved experiment that puts declaration, internal state, behavior, and persistence on the same x-axis.

Sources and reproducibility

Shenoy K, Yang L, Sheshadri A, Mindermann S, Lindsey J, Marks S, Wang R. Introspection Adapters: Training LLMs to Report Their Learned Behaviors. arXiv:2604.16812 (2026). https://arxiv.org/abs/2604.16812

Anthropic Alignment Science Blog. Introspection Adapters. 28 Apr 2026. https://alignment.anthropic.com/2026/introspection-adapters/

Code: https://github.com/safety-research/introspection-adapters

Released models/data organization: https://huggingface.co/introspection-auditing

Bozoukov M, Nguyen M, Singh S, Bussmann B, Leask P. Minimal and Mechanistic Conditions for Behavioral Self-Awareness in LLMs. arXiv:2511.04875 (2025). https://arxiv.org/abs/2511.04875

Turner E, Soligo A, Taylor M, Rajamanoharan S, Nanda N. Model Organisms for Emergent Misalignment. arXiv:2506.11613 (2025). https://arxiv.org/abs/2506.11613

Statistical intervals in this addendum were calculated from published numerator/denominator counts. No new LLM inference or training was run for this addendum. The external papers' terms such as 'self-awareness' are preserved as the authors' operational labels and are not treated here as proof of phenomenal consciousness.