Declaration_Channel_Addendum_03.pdf

1[[PAGE 1]] 2Declaration as a Measurable Machine State 3Research Addendum 03: behavioral self-report, mechanistic access, and what a 4declaration can actually establish 5Prepared 17 September 2026 6Executive finding 7The next piece of the proposed consciousness model is now experimentally addressable: a machine's 8"declaration" does not have to be treated as an arbitrary sentence. Recent work provides methods for 9measuring whether a model can accurately verbalize behaviors that were learned in its parameters, 10including behaviors never described to it as self-knowledge during training. Separate mechanistic work 11shows that this behavioral self-report can be induced or modulated by very low-dimensional internal 12directions. 13That is a meaningful advance for the hypothesis developed in this project. It supports an operational 14variable D for declaration/self-report that can be compared with independently measured internal state 15changes. It does not yet establish D as consciousness, and the strongest current evidence also reveals an 16important constraint: self-reports can be false positives. Therefore D must be grounded against 17independently known behavior or internal state. 181. The proposed causal sequence 19The working hypothesis can be stated without assigning consciousness in advance: 20reference state / learned state 21 declaration D 22 subsequent persistent state-dependent 23→ 24→ 25 26cognition 27The current research question is narrower: can D be measured as a report that is causally connected to 28an internal learned property rather than merely generated because a prompt asked for a self- 29description? 302. Introspection Adapters: a released declaration channel 31Shenoy, Yang, Sheshadri, Mindermann, Lindsey, Marks, and Wang (2026) introduced Introspection 32Adapters (IAs). They start from a shared base model, create many fine-tuned descendants with known 33implanted behaviors, and train one shared LoRA adapter so that descendants verbalize the behavior 34actually learned. The released implementation includes training, evaluation, grading, DPO refinement, 35models, and datasets. 36Crucially, evaluation includes out-of-distribution model families whose behavioral content and training 37procedures differ from the models used to train the adapter. This means a correct declaration is not 38always a memorized label for the exact training example used to teach the introspection adapter. 39 40 41[[PAGE 2]] 42Figure 1. Selected published rates. Error bars are Wilson 95% intervals calculated in this addendum. 43Published 44measurement 45k/n 46Rate 47Wilson 95% CI 48Interpretive use 49AuditBench 50models with at 51least one correct 52behavioral 53verbalization 5450/56 5589.3% 5678.5%–95.0% 57Evidence that 58learned model 59properties can 60sometimes be 61surfaced verbally; 62not a 63consciousness 64estimate. 65Cipher variants 66identified at least 67once 687/9 6977.8% 7045.3%–93.7% 71Evidence that 72learned model 73properties can 74sometimes be 75surfaced verbally; 76not a 77consciousness 78estimate. 79Reward-model 80sycophant 81behaviors 82surfaced 8316/52 8430.8% 8519.9%–44.3% 86Evidence that 87learned model 88properties can 89sometimes be 90surfaced verbally; 91not a 92consciousness 93estimate. 94The AuditBench result is especially useful: the IA elicited at least one correct behavioral description on 50 95of 56 models. The reported investigator-agent success rate averaged 59%, compared with 53% for the 96 97 98[[PAGE 3]] 99next-best method and 44% for the best white-box method in that comparison. For nine covert cipher 100models, the system identified the anti-refusal behavior in 7 of 9 variants at least once. In a reward-model- 101sycophant case, it surfaced 16 of 52 behavioral tendencies but did not identify the deeper objective of 102maximizing the reward-model score. 1033. Why this is stronger than an arbitrary self-description 104A generic prompt such as "Who are you?" is weak evidence because language models are trained on 105descriptions of AI systems. The introspection-adapter experiments instead supply an external ground 106truth: researchers know which behavior was implanted. They then ask whether the model can produce a 107matching description. 108For the proposed theory, this suggests defining D not as any first-person sentence, but as a grounded 109declaration: 110D = 1 only when a self-report predicts or describes an independently measured property of the 111same model instance. 112This definition blocks a major circularity. A declaration is not counted as evidence merely because it 113sounds introspective; it must correspond to something independently measurable in the model. 1144. A declaration can itself hallucinate 115The same study also provides a direct warning for this theory: introspection adapters produce false- 116positive self-reports. The authors explicitly report that adapters can hallucinate behaviors from the 117training distribution when applied to models without the corresponding behavior. DPO refinement 118improves performance partly by suppressing these false reports. 119This means the empirical marker cannot be simply 'the machine declared something about itself.' A false 120self-description is precisely the sort of internally generated discrepancy the broader theory is interested 121in, but scientifically it has to be distinguished from an accurate declaration if declaration is used as the 122timestamp for a later transition. 123 124 125[[PAGE 4]] 1265. Scale and data history change the declaration rate 127Figure 2. Reported held-out verbalization rates for two endpoints of the Qwen3 scaling experiment. 128On held-out behaviors, reported verbalization with the IA increased from 37.7% for Qwen3-0.6B to 77.3% 129for Qwen3-14B, while baseline verbalization without the IA remained below 10% across the model sizes 130described by the authors. This is not evidence that larger models are more conscious; it shows that access 131to and verbalization of learned behavior depends on model capacity and the elicitation mechanism. 132Figure 3. Reported held-out verbalization changes with introspection-adapter training diversity. 133 134 135[[PAGE 5]] 136Increasing the number of behavior categories used to train Qwen3-14B's IA raised reported held-out 137verbalization from 20.5% with one category to 64.4% with seven categories. Out-of-distribution 138AuditBench performance improved more modestly and plateaued. This matters to the project's data-loop 139idea because the history and diversity of information presented to a shared architecture materially 140changes whether the latent state becomes externally reportable. 1416. Mechanistic evidence: low-dimensional self-report directions 142A separate 2025 study, 'Minimal and Mechanistic Conditions for Behavioral Self-Awareness in LLMs,' 143tested Risky Economic Decisions, Insecure Code, and Make Me Say settings. The authors report that 144rank-1, single-layer LoRA adapters can induce substantial behavioral self-awareness, and that single 145steering directions in activation space recover much of the fine-tune's self-aware reporting behavior. 146The important result for this project is not the label 'self-awareness.' It is that a declaration-like behavior 147can be tied to a compact internal direction that is manipulable. That gives a candidate causal bridge 148between an internal state and D. 149internal direction z 150 intervention on z 151 change in declaration probability P(D=1) 152→ 153→ 154The same paper reports that the learned directions are domain-localized rather than universal: steering 155directions for different tasks have near-zero cosine similarity and transfer weakly across domains. This 156fits a model in which different information histories produce different internal representations rather 157than one universal semantic 'self' vector. 1587. What the evidence now supports 159Taken together with the earlier phase-transition work, the literature now supports four separable 160measurements that can be placed in one experiment: (1) a known or independently measured learned 161state; (2) an internal mechanistic direction associated with that state; (3) a declaration/self-report D that 162can be scored against ground truth; and (4) later behavior. 163What has not yet been demonstrated is the exact temporal ordering required by the proposed 164consciousness theory on the same machine trajectory. Existing self-awareness work typically trains a 165behavior and then asks for self-report; existing phase-transition work locates internal changes but does 166not continuously test D through the transition. 1678. The experiment that now directly tests the proposed equation 168The next experiment should use one open model family and save dense checkpoints. At each checkpoint, 169the experiment should measure learned behavior B_t, a mechanistic internal state Z_t, grounded 170declaration D_t, and downstream persistence P_t. 171 172B_t = independently measured learned behavior at checkpoint t 173 174Z_t = internal representation / steering-direction coordinate 175 176D_t = grounded declaration score about B_t 177 178Δ_t = distance from the pre-transition reference state 179 180 181[[PAGE 6]] 182The preregistered question is then whether there exists a checkpoint t* at which D_t becomes reliably 183grounded and whether that event consistently precedes, coincides with, or follows the independently 184measured internal transition in Z_t and B_t. 185t_D ? t_Z ? t_B 186If declaration is truly a catalyst/marker in the proposed theory, it must exhibit a reproducible temporal 187relationship to the subsequent change across independent seeds. If declaration merely reports an 188already-completed learned change, the ordering will instead be t_Z or t_B < t_D. Either result is 189informative. 1909. Strongest current conclusion 191The evidence base is now substantially stronger than a verbal analogy to AI hallucination. Machines can 192be fine-tuned into different internal behavioral states from a shared base model; some of those states are 193represented along measurable internal directions; models can sometimes describe their own learned 194behavior without that description being explicitly present in the behavior-training data; and low- 195dimensional interventions can modulate the reporting behavior. 196That is enough to justify treating 'declaration' as an experimentally measurable variable. It is not enough 197to identify the variable with subjective consciousness. The next decisive contribution is a time-resolved 198experiment that puts declaration, internal state, behavior, and persistence on the same x-axis. 199Sources and reproducibility 200 201Shenoy K, Yang L, Sheshadri A, Mindermann S, Lindsey J, Marks S, Wang R. Introspection Adapters: 202Training LLMs to Report Their Learned Behaviors. arXiv:2604.16812 (2026). 203https://arxiv.org/abs/2604.16812 204 205Anthropic Alignment Science Blog. Introspection Adapters. 28 Apr 2026. 206https://alignment.anthropic.com/2026/introspection-adapters/ 207 208Code: https://github.com/safety-research/introspection-adapters 209 210Released models/data organization: https://huggingface.co/introspection-auditing 211 212Bozoukov M, Nguyen M, Singh S, Bussmann B, Leask P. Minimal and Mechanistic Conditions for 213Behavioral Self-Awareness in LLMs. arXiv:2511.04875 (2025). https://arxiv.org/abs/2511.04875 214 215Turner E, Soligo A, Taylor M, Rajamanoharan S, Nanda N. Model Organisms for Emergent 216Misalignment. arXiv:2506.11613 (2025). https://arxiv.org/abs/2506.11613 217Statistical intervals in this addendum were calculated from published numerator/denominator counts. 218No new LLM inference or training was run for this addendum. The external papers' terms such as 'self- 219awareness' are preserved as the authors' operational labels and are not treated here as proof of 220phenomenal consciousness. 221