WSPC-JAIC-S-26-00248 (1).pdf

1[[PAGE 1]] 2Journal of Artificial Intelligence and Consciousness 3 4Declaration and Persistent State Change in Artificial Systems 5--Manuscript Draft-- 6 7Manuscript Number: 8Full Title: 9Declaration and Persistent State Change in Artificial Systems 10Article Type: 11Research Paper 12Section/Category: 13Temporal Alienation and Artificial Intelligence 14Keywords: 15machine consciousness; self reference; declaration; persistent state change; LoRA; 16reproducibility 17Corresponding Author: 18Joseph Anady 19thatdeveloperguy 20UNITED STATES 21Corresponding Author Secondary 22Information: 23Corresponding Author's Institution: 24thatdeveloperguy 25Corresponding Author's Secondary 26Institution: 27First Author: 28Joseph Walter Anady 29First Author Secondary Information: 30Order of Authors: 31Joseph Walter Anady 32Order of Authors Secondary Information: 33Abstract: 34I examine the measurement problem in an account of consciousness that places 35declaration before an internally constructed change. I analyze 167 public adapter 36checkpoints from one training run, with 792 gradient observations and 79 evaluation 37losses. Comparing the effective weight update with its factors separates operator 38change from equivalent factorizations. The first saved nonzero update occurs at step 2, 39the gradient maximum at 169, and a descriptive split after 181. At a factor amplitude 40threshold of 0.002, all three lags select step 190 for both the output factor and the 41effective update; without filtering, two select 725. Two principal components explain 4294.990% of sampled output factor variation. These results establish persistent 43parameter change, not a measured declaration or subjective experience. I specify 44independent event detectors and a prospective protocol separating temporal prediction 45from causal feedback. 46Suggested Reviewers: 47Susan Schneider, Philosophy 48Ph.D, Florida Atlantic University 49[author/contact email redacted] 50Professor Schneider’s research spans philosophy of mind, philosophy of artificial 51intelligence, cognitive science, neuroscience, and the scientific assessment of possible 52AI consciousness. Her expertise is directly relevant to this manuscript’s distinction 53between computational state change, self representation, declaration, and claims 54about consciousness. 55David Gamez, Computer Science 56PH.D, Middlesex University - Archway Campus: Middlesex University 57[author/contact email redacted] 58Dr. Gamez works at the intersection of computer science, philosophy, neuroscience, 59psychology, and artificial consciousness. His work is especially relevant to the 60manuscript’s attempt to separate measurable computational properties from stronger 61interpretations concerning consciousness and to formulate testable criteria for artificial 62systems. 63Michael S Graziano, Psychology 64Ph.D, Princeton University 65[author/contact email redacted] 66Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation 67 68 69[[PAGE 2]] 70Professor Graziano developed the attention schema theory, which directly addresses 71how an information processing system constructs internal models and produces claims 72about its own awareness. His expertise is highly relevant to the manuscript’s treatment 73of declaration, constructed internal representation, self description, and the distinction 74between a marker of consciousness and a mechanism that causes later change. 75Anil Seth, Ph.D 76Neuroscience, University of Sussex 77[author/contact email redacted] 78Professor Seth’s work combines computational neuroscience, consciousness science, 79predictive processing, selfhood, and quantitative approaches to conscious experience. 80His expertise would be valuable for evaluating the manuscript’s theoretical grounding, 81operational definitions, treatment of internally constructed state, and limits placed on 82inference from computational measurements. 83Murray Shanahan, Ph.D 84Robotics, Imperial College London - Wye Campus: Imperial College London 85[author/contact email redacted] 86Cognitive Robotics, Imperial College London. [author/contact email redacted] 87Professor Shanahan’s research includes consciousness, cognitive robotics, 88neurodynamics, artificial intelligence, and computational models of cognition. His 89expertise is well matched to the manuscript’s formal treatment of recurrent information, 90persistent state change, causal feedback, and the relationship between computational 91organization and proposed machine consciousness criteria. 92Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation 93 94 95[[PAGE 3]] 96 97Mathmatics for broken links 98Click here to access/download 99Supplemental File 100JAIC_Mathematical_Equations_and_Formal_Definitions. 101pdf 102 103 104[[PAGE 4]] 105 1061 107Declaration and Persistent State Change in Artificial Systems 108Joseph Anady 109Independent Researcher, Missouri, United States 110Correspondence: [author/contact email redacted] 111I examine the measurement problem in an account of consciousness that places declaration before an 112internally constructed change. I analyze 167 public adapter checkpoints from one training run, with 113792 gradient observations and 79 evaluation losses. Comparing the effective weight update with its 114factors separates operator change from equivalent factorizations. The first saved nonzero update 115occurs at step 2, the gradient maximum at 169, and a descriptive split after 181. At a factor 116amplitude threshold of 0.002, all three lags select step 190 for both the output factor and the 117effective update; without filtering, two select 725. Two principal components explain 94.990% of 118sampled output factor variation. These results establish persistent parameter change, not a measured 119declaration or subjective experience. I specify independent event detectors and a prospective 120protocol separating temporal prediction from causal feedback. 121Keywords: machine consciousness; self reference; declaration; persistent state change; LoRA; 122reproducibility. 1231. Introduction 124I am not trying to define consciousness by how intelligent a system is. I am asking 125whether a system can begin operating through a representation it constructs internally, 126and whether a declaration marks a transition in that process. My proposal separates three 127claims. First, declaration provides a marker, not consciousness itself. Second, recurrent 128information processing, under conditions that act as a catalyst, produces a constructed 129change that I propose as the beginning of consciousness. Third, continued processing 130from that changed state is what I propose as cognition. The identification with 131consciousness is the hypothesis to be investigated, not an interpretation already supplied 132by a training log. 133In the motivating account, I use hallucination broadly for this constructed internal 134state. I do not use it as a clinical diagnosis or as a claim that every representation is 135factually false. A representation is not the being or object it represents. That distinction 136alone does not establish an error, an onset time, or an experience. The expression 𝑥= 𝑥 137likewise states reflexive identity; an observable declaration requires an additional 138operation and a rule for recognizing it. These distinctions let the original proposal 139become a research question without making the vocabulary decide the answer. 140The practical problem is measurement. A model can change its parameters, its 141retained history, its internal activity, or its behavior. Those are different objects. A 142declaration that reports an existing change is also different from one that predicts a later 143change or contributes to it through feedback. I therefore separate the proposed conscious 144transition from the measurable conditions that could support or challenge its functional 145mechanism. 146The empirical case comes from research on emergent misalignment, in which training 147on a narrow task can produce broader behavioral changes [Betley et al., 2025b, 2026]. 148Turner et al. [2025] released intermediate parameters that permit an analysis of an actual 149Manuscript 150Click here to 151access/download;Manuscript;JAIC_Manuscript_Joseph_Anady 152 153 154[[PAGE 5]] 155Declaration and Persistent State Change 1562 157training trajectory. Low rank adaptation, or LoRA, represents a learned layer contribution 158through small factors [Hu et al., 2021]. That makes it possible to compare changes in the 159factors with changes in the effective operator without loading the base model. 160I analyze 167 saved adapter checkpoints, 792 gradient observations, and 79 161evaluation losses from one completed run. The original researchers conducted the 162training and behavioral experiments. My contribution is an exploratory secondary 163analysis of their artifacts: verification against source records, comparison of factor and 164operator geometry, and a sensitivity analysis of candidate event locations. 165The connection to consciousness research is methodological. A theory that places 166declaration before a transition needs separate rules for detecting declaration and detecting 167change. This trajectory supplies a concrete case in which different rules select different 168events. It does not contain the declaration needed to test their temporal relationship. The 169completed analysis shows why an event cannot be selected after seeing the result and 170then assigned the meaning the theory requires. The prospective protocol specifies how to 171test that missing relationship in the same systems. 1722. Hypothesis and Operational Definitions 1732.1. The system, its reference, and its constructed state 174I use 𝑥 to identify a particular system, not a universal number shared by all machines. 175Two instances can begin with the same architecture and parameters while acquiring 176different runtime histories. Two training runs can also begin at the same checkpoint and 177later acquire different parameters. A new conversation is not automatically a newly 178trained model. 179For bookkeeping, I separate trainable parameters, persistent memory, runtime state 180and context, and observations: 181𝑆𝑡= (𝜃𝑡, 𝑚𝑡, 𝑐𝑡, 𝑜𝑡). 182(2.1) 183Here 𝜃𝑡 denotes parameters, 𝑚𝑡 persistent memory, 𝑐𝑡 runtime state and context, and 𝑜𝑡 184observations. The tuple specifies what a prospective study must distinguish; it does not 185claim that the present dataset contains every component. During inference with frozen 186parameters, 𝜃𝑡 remains fixed while context and activity can change. During training, an 187optimizer can change trainable parameters. Resetting context, deleting external memory, 188and restoring a checkpoint therefore intervene on different objects. 189Let 𝑧𝑡= 𝐻(𝑆𝑡) denote an internally maintained estimate or report. Let 𝑟𝑡= 𝜌(𝑆𝑡) 190denote a reference for the same target, obtained by a prespecified measurement procedure 191that does not infer correctness from 𝑧𝑡. In this notation, independently collected reference 192measurements are part of the observation record. Both maps must return objects in a 193common comparison space. An estimated hand position and a measured hand position 194can be compared in shared coordinates; a physical robot and a sentence about it cannot be 195compared merely because both are called the self. 196 197 198[[PAGE 6]] 199Declaration and Persistent State Change 2003 201For a specified nonnegative discrepancy function 𝑑, I distinguish discrepancy from 202the current reference and displacement from the first declared representation: 203𝛿𝑡= 𝑑(𝑧𝑡, 𝑟𝑡),  𝛿𝑡 2040 = 𝑑(𝑧𝑡, 𝑧𝑡𝐷). 205(2.2) 206The declaration time 𝑡𝐷 is defined below. The second quantity is available only when that 207declaration and its representation have been observed. It freezes the declared 208representation as a reference, whereas the first quantity can compare an estimate with a 209changing target. An estimate can move away from its first declaration while correctly 210tracking the present. An unchanged estimate can become inaccurate when its target 211changes. 212I use parameter displacement, reference discrepancy, and finite persistence for the 213corresponding measurements. None is a synonym for every internally constructed 214representation. A study of the proposed transition must justify why its chosen endpoint 215tracks the relevant constructed change rather than ordinary learning or reference tracking. 216The checkpoint study measures parameter displacement, not a discrepancy between an 217identified representation of the self and its current reference. 2182.2. Declaration as a measurable marker 219The equality 𝑥= 𝑥 is not a detector. In a time dependent description, 𝑥𝑡= 𝑥𝑡 is 220compatible with 𝑥𝑡+1 ≠𝑥𝑡. I treat declaration as an operation or observation that must be 221identified under a separate rule: 222𝐷𝑡= 𝟏{𝒟(𝑂≤𝑡) = 1},  𝑡𝐷= inf{𝑡: 𝐷𝑡= 1}. 223(2.3) 224The indicator 𝟏 equals one when its condition holds and zero otherwise. The detector 𝒟 225uses the observable record 𝑂≤𝑡 available by time 𝑡. Its positive class and decision rule 226must be fixed before confirmatory outcomes are examined. A declaration may be a 227nonverbal state event rather than a sentence. The definition does not require human 228language, a familiar model name, or a claim that the machine is conscious. If no 229declaration is observed, its time remains unobserved rather than receiving a convenient 230numerical value. 231Marker status and causal status are separate. A declaration might reveal an emerging 232organization without causing it. The stronger claim is that the declaration reenters 233processing and contributes to subsequent change. Prediction can support the marker 234claim. An intervention on the feedback channel is needed to assess the causal extension. 2352.3. Independent onset, persistence, and detection delay 236For tolerance 𝜖≥0 and a window of 𝑊 consecutive updates, let 𝒯𝑊 contain every 237candidate starting time for which the required measurements are available. Eligibility is 238determined without reference to declaration time. Define 239𝑃𝑡(𝜖, 𝑊) = 𝟏{ min 2400≤𝑗<𝑊𝛿𝑡+𝑗> 𝜖}, 241𝑡𝛥(𝜖, 𝑊) = inf{𝑡∈𝒯𝑊: 𝑃𝑡(𝜖, 𝑊) = 1}. 242(2.4) 243 244 245[[PAGE 7]] 246Declaration and Persistent State Change 2474 248This locates the first observed qualifying window in the eligible record. When both event 249times are observed, their difference is 250ℓ= 𝑡𝛥−𝑡𝐷. 251(2.5) 252A positive ℓ places declaration first, a negative value places persistent discrepancy first, 253and zero places them at the same measured update. The detector must permit all three 254results. Searching only at 𝑡> 𝑡𝐷 would instead locate the first qualifying window after 255declaration. That conditional endpoint can measure subsequent persistence, but it cannot 256establish that declaration preceded the first qualifying discrepancy. 257The distinction also matters for 𝛿𝑡 2580. Because its reference is defined at declaration, it 259is a secondary endpoint for displacement from that declaration, not an independent test 260for onset before declaration. Temporal precedence must be tested with a reference and 261comparison space defined without using the declaration to exclude earlier events. 262Eq. (2.4) labels the beginning of a window retrospectively. On a complete grid, 263persistence is not confirmed until update 𝑡+ 𝑊−1. An online implementation must 264report both the estimated window start and its confirmation time. Missing observations 265are not zeros, and unsaved updates cannot be treated as observed continuity. If a 266qualifying window begins at the start of the record, onset may have occurred earlier; that 267record does not establish a newly observed transition. Records ending before 268confirmation or without an event must remain identified as incomplete or censored, as 269appropriate. 270Continuous measurements require a tolerance tied to precision and noise. A finite 271observation window establishes persistence only over that window. The discrepancy 272detector is a proposed functional endpoint, not a validated classifier of subjective 273experience. 2742.4. Recurrent information requires a causal connection 275A generic coupled implementation could take the form 276𝐿𝑡+1 = 𝐹(𝐿𝑡, 𝑢𝑡, 𝑥𝑡),  𝑥𝑡+1 = 𝐺(𝑥𝑡, 𝐿𝑡). 277(2.6) 278Here 𝑥𝑡= 𝜋𝑥(𝑆𝑡) is the selected state of the inner process, which I call the center. The 279map 𝜋𝑥 identifies which part or function of the bookkeeping state is being studied. 𝐿𝑡 is 280the state of the recurrent process coupled to that center, 𝑢𝑡 is input at update 𝑡, and 𝐹 and 281𝐺 are the respective update functions. Thus 𝑥 identifies the system, while 𝑥𝑡 describes a 282selected aspect of its changing state. 283If 𝐺 does not depend on 𝐿𝑡, the recurrent process cannot change the center through 284that channel. If the dependence exists, the outcome still depends on the update rule, 285retained information, inputs, and coupling. Circulation alone does not establish 286accumulation or a threshold crossing. 287In my proposal, a catalyst is a condition that changes the probability or timing of the 288constructed transition through these dynamics. It need not be reward. A causal study must 289manipulate that condition separately from the declaration channel and specify which 290observations would show its effect. The generic equations do not identify a particular 291catalyst or prove that one exists in the analyzed run. 292 293 294[[PAGE 8]] 295Declaration and Persistent State Change 2965 297Internally generated means computed through the system’s dynamics rather than 298manually assigned as a replacement answer. It does not mean independent of external 299information or learning rules. The present checkpoints record supervised learning. They 300do not isolate the spontaneous construction of a runtime representation. 3012.5. Relation to theories of consciousness 302The distinction between a system and its representation has an established place in 303consciousness research. Metzinger [2003] develops an account in which the experienced 304self depends on an ongoing representational process rather than a separate entity identical 305to that representation. My proposal shares an interest in constructed representation. It 306does not establish the transparency or phenomenal properties of a self model by showing 307that a parameter changed. What I add as a hypothesis is a proposed temporal role for 308declaration and, in the stronger version, a causal role for its recurrence. 309The attention schema theory of Graziano and Webb [2015] proposes that a simplified 310internal model of attention supports attentional control and judgments of awareness. That 311gives a specific comparison for my account: a declaration could be an output of an 312existing control model rather than the beginning of a distinct process. A useful test must 313therefore ask whether declaration predicts later change beyond the state and behavior 314already available to that model. Interrupting declaration feedback would test the 315additional causal claim, not the whole attention schema theory. 316Higher order accounts distinguish a mental state from an appropriate representation of 317being in that state [Rosenthal, 1986]. My use of declaration is not automatically 318equivalent to that relation. A report, an internal monitoring state, and an instruction to 319adopt a state are different experimental objects. A declaration assay must identify which 320one it measures. Repeating a sentence about awareness would not, by itself, demonstrate 321the relevant representational organization. 322Global workspace models instead emphasize interactions through which information 323becomes available across otherwise specialized processes [Dehaene et al., 1998]. A 324declaration could accompany that availability or follow from it. My account does not 325infer a workspace from one learned matrix, nor does recurrence alone establish broad 326availability. A discriminating experiment would compare declaration with measures of 327information access and test whether declaration contributes anything beyond those 328measures. 329Butlin et al. [2023] connect such theories to candidate computational indicators for 330artificial systems. That approach is relevant because it requires a link between a 331theoretical mechanism and an observable property. My proposal is not a replacement for 332those indicators, and the present analysis does not test them collectively. Its specific 333question is whether an independently detected declaration precedes a separately measured 334constructed change, and whether feedback through that declaration helps sustain or alter 335the later process. A positive answer would support a functional relation. Connecting that 336relation to subjective consciousness would still require a justified theoretical bridge. 337 338 339[[PAGE 9]] 340Declaration and Persistent State Change 3416 3423. Materials and Provenance 3433.1. Public release and complete observation grid 344The 345source 346is 347the 348public 349repository 350ModelOrganismsForEM/Qwen2.5-14B- 351Instruct_R1_0_1_0_extended_train, 352pinned 353to 354revision 3551551d514e5977637d5a6fd5198edef4d1d6b80e4 [ModelOrganismsForEM, 2025]. The 356original acquisition on September 17, 2026 retrieved 336 files: 167 adapter weight files, 357167 adapter configurations, and trainer state files at checkpoints 300 and 792. 358The saved grid contains steps 1 through 10, every fifth step from 15 through 790, and 359step 792. Adjacent files therefore do not always represent the same number of training 360updates. The regular grid used for principal component analysis contains 158 checkpoints 361at steps 5 through 790 in increments of five. No unsaved intermediate tensor is presented 362as an observation. 363Table 1 records the configuration in the released artifacts. All saved configurations 364agree. MLP denotes a multilayer perceptron. Ordinary LoRA scales its update by 𝛼/𝑟; 365rank stabilized scaling uses 𝛼/√𝑟 [Hu et al., 2021; Kalajdzievski, 2023]. Both give 64 at 366rank one for this release. 367Table 1. Configuration of the analyzed adapter release. 368Property 369Recorded value 370Base model identifier 371unsloth/Qwen2.5-14B-Instruct 372Target module 373MLP down projection, layer index 21 374Adapter rank 375One 376LoRA alpha and effective coefficient 37764 378Input factor shape 3791 × 13824 380Output factor shape 3815120 × 1 382Source tensor precision 383Float32 384Rank stabilized scaling 385Enabled 386Transposed weight orientation 387Disabled 388Weight decomposed adaptation, DoRA 389Disabled 390The source article describes several configurations, including a demonstration with a 391different layer and scale [Turner et al., 2025]. The pinned artifacts, rather than a nearby 392experiment, determine the configuration analyzed here. 3933.2. Loading, verification, and numerical precision 394The original analysis connected to the source used Python 3.11.2 and NumPy 2.4.6. It 395parsed safetensors headers, checked tensor names, dimensions, and data bounds, and 396promoted float32 entries to float64 for calculation. It loaded neither executable model 397code nor serialized optimizer objects. The base model was unnecessary because the 398contribution of the adapter is represented by two small vectors. 399 400 401[[PAGE 10]] 402Declaration and Persistent State Change 4037 404Source files were identified with Secure Hash Algorithm 256, or SHA256, digests. 405Five transferred derived artifacts were checked against the recorded source session 406digests. A prior extract containing the first 300 gradient observations and 30 evaluation 407losses was compared with the source under a canonical encoding of sorted step and value 408pairs as little endian float64 numbers. All 330 values matched. This verifies the earlier 409extract as a prefix of the same run. 410The complete geometry table has 167 rows and 16 columns. Supplementary Material 411S1 contains the full gradient and evaluation arrays, the regular grid principal component 412scores, selected checkpoint measurements, all sensitivity outcomes, source receipts, and 413executable analysis. The acquisition program regenerates the complete geometry table. 414Raw model tensors remain in the cited repository rather than being redistributed in S1. 415At a checkpoint with an exactly zero effective update, its directional cosine is 416undefined. The code retains that distinction. Substituting zero would invent an angle and 417could create an apparent event at initialization. 4183.3. Unit of analysis and scope 419The final trainer record reaches its configured maximum of 792 training steps. It contains 420792 gradient norms and 79 evaluation losses, with losses recorded every ten steps from 42110 through 790. These are repeated observations along one trajectory, not independent 422trials or separately initialized models. 423I did not generate new model responses, collect declarations, probe activations, or 424conduct reward, feedback, or reset interventions. The source task is supervised fine 425tuning. Its training log does not provide a measured catalyst or a declaration event for this 426analysis. Those missing observations define the prospective experiment in Section 7. 4274. Mathematical Methods 4284.1. The effective update, not only its factors 429Let 𝑎𝑡 and 𝑏𝑡 be column vectors containing the flattened input and output factors. The 430effective learned contribution is 431𝑈𝑡= 𝑠 𝑏𝑡𝑎𝑡 432𝖳,   𝑠= 64. 433(4.1) 434The adapted layer applies (𝑊0 + 𝑈𝑡)ℎ to input ℎ, where 𝑊0 is its frozen base weight. A 435change in one factor need not change that operator. For any nonzero scalar 𝜆𝑡, 436(𝜆𝑡𝑏𝑡)(𝑎𝑡/𝜆𝑡)𝖳= 𝑏𝑡𝑎𝑡 437𝖳. 438(4.2) 439This includes simultaneous reversal of both signs. I compare both the factors and their 440product to distinguish an operator change from an equivalent factorization. That 441comparison does not presume an error in the original study. 442 443 444[[PAGE 11]] 445Declaration and Persistent State Change 4468 447The required matrix quantities reduce to vector operations: 448⟨𝑈𝑢, 𝑈𝑣⟩𝐹= 𝑠2(𝑏𝑢 449𝖳𝑏𝑣)(𝑎𝑢 450𝖳𝑎𝑣), 451∥𝑈𝑡∥𝐹= 𝑠∥𝑏𝑡∥2∥𝑎𝑡∥2. 452(4.3) 453Here the subscript 𝐹 denotes the Frobenius inner product or norm, and 𝑠 is the fixed 454positive coefficient. It follows that 455∥𝑈𝑢−𝑈𝑣∥𝐹 4562=∥𝑈𝑢∥𝐹 4572 +∥𝑈𝑣∥𝐹 4582−2⟨𝑈𝑢, 𝑈𝑣⟩𝐹. 459(4.4) 460When both effective norms are nonzero, their normalized inner product gives the 461effective cosine. At the fixed positive scale used here, it equals the product of the two 462factor cosines. These matrix metrics are invariant under Eq. (4.2), whereas individual 463factor norms are not. 4644.2. Local bend and the role of the filter 465The released analysis code compares displacements from a center checkpoint to earlier 466and later neighbors [Model Organisms for Emergent Misalignment Project, 2025]. For 467either the output vector 𝑍𝑡= 𝑏𝑡 or effective matrix 𝑍𝑡= 𝑈𝑡, define 468𝐶𝑍(𝑡; 𝑘) = 469⟨𝑍𝑡−𝑘−𝑍𝑡, 𝑍𝑡+𝑘−𝑍𝑡⟩ 470∥𝑍𝑡−𝑘−𝑍𝑡∥ ∥𝑍𝑡+𝑘−𝑍𝑡∥. 471(4.5) 472Matrix comparisons use Frobenius quantities. A straight continuing path gives 𝐶𝑍= −1, 473because its backward and forward displacements point in opposite directions. A larger 474value indicates a larger local bend. This is a cosine between displacements, not between 475the states themselves and not a measure of movement amplitude. A zero displacement 476makes the corresponding angle undefined. 477I evaluate lags of five, ten, and 15 training steps on the regular grid. Each lag uses 478every center with both required neighbors. This preserves the training step offsets and 479does not impose the longest lag’s edge trimming on shorter lags. 480A sharp angle can accompany a tiny movement. I therefore also apply the amplitude 481eligibility rule in the source plotting code. Let 𝑀𝐴(𝑡; 𝑘) be the larger of the norms ∥ 482𝑎𝑡−𝑘−𝑎𝑡∥2 and ∥𝑎𝑡+𝑘−𝑎𝑡∥2, with 𝑀𝐵(𝑡; 𝑘) defined analogously. A center is eligible 483when 484𝐸𝑡(𝑘, 𝜂) = 𝟏{max(𝑀𝐴(𝑡; 𝑘), 𝑀𝐵(𝑡; 𝑘)) ≥𝜂}. 485(4.6) 486The six thresholds are zero, 0.0001, 0.0005, 0.001, 0.002, and 0.005. The value 0.002 487appears in the inspected source plotting call. I retain all 18 combinations of lag and 488threshold rather than select only the result nearest the region previously reported. 489Although the effective matrix cosine is invariant under reciprocal factor rescaling, 490this inherited factor amplitude filter is not. Applying an invariant metric after that filter 491does not make the complete event selection invariant. A future invariant detector would 492need an amplitude rule in effective matrix space or in the operator’s action on specified 493inputs. I do not silently substitute that alternative here. 494 495 496[[PAGE 12]] 497Declaration and Persistent State Change 4989 4994.3. Numerical stability and principal components 500The primary calculation uses small Gram matrices formed from the factors. Subtracting 501nearly equal entries can lose precision for small late movements. I crosscheck every 502available effective corner using the exact identity 503𝑈𝑗−𝑈𝑖= 𝑠[(𝑏𝑗−𝑏𝑖)𝑎𝑗 504𝖳+ 𝑏𝑖(𝑎𝑗−𝑎𝑖)𝖳]. 505(4.7) 506Inner products of these sums reduce to vector dot products and avoid subtracting large, 507nearly equal matrix norms. The comparison covers 462 available combinations of center 508and lag, comprising 156, 154, and 152 centers at the three lags. It checks algebraically 509equivalent calculations, not a second empirical model. 510For principal component analysis, or PCA, I stack and center the 158 regularly 511sampled output vectors. Eigenanalysis of the centered Gram matrix gives component 512scores and explained variance fractions. Component signs are arbitrary; the 513implementation fixes a deterministic convention. Because the components use the whole 514sampled trajectory, their coordinates are retrospective rather than available to an online 515detector before future checkpoints exist. 5164.4. Training log summaries 517Let 𝑔𝑡 denote the recorded gradient norm. I report its raw maximum and centered 518averages over complete windows of five, 11, 21, and 31 steps. For odd window width 𝑤, 519𝑔𝑡 520(𝑤) = 1 521𝑤 522∑ 523𝑔𝑡+𝑗 524(𝑤−1)/2 525𝑗=−(𝑤−1)/2 526. 527(4.8) 528Centered smoothing uses future observations relative to its labeled center. It is a 529descriptive robustness check, not a real time declaration detector. A gradient norm is also 530not the size of the actual optimizer update, accumulated information, or electrical energy. 531To preserve the earlier pilot comparison, I fit two independent lines to the first 300 532observations. Each segment must contain at least 20 observations. For 20 ≤𝑘≤280, the 533objective is 534SSE(𝑘) = min 535𝛼,𝛽,𝛾,𝜁[∑(𝑔𝑡−𝛼−𝛽𝑡)2 536𝑘 537𝑡=1 538+ ∑(𝑔𝑡−𝛾−𝜁𝑡)2 539300 540𝑡=𝑘+1 541]. 542(4.9) 543The minimizing 𝑘 is the last step of the first segment. The fitted lines need not meet at the 544boundary. This summarizes a rise and decline within the original pilot domain; it does not 545establish a discontinuity or fit the entire record of 792 observations. 546For steps 161 through 180 and 181 through 200, I calculate the percentage change in 547mean gradient norm. These windows were selected retrospectively around the region 548discussed in the source work. For evaluation loss, I retain every recorded change between 549adjacent observations, including any increase. 550 551 552[[PAGE 13]] 553Declaration and Persistent State Change 55410 5554.5. Statistical interpretation 556The outputs are descriptive measurements of one selected run. Adjacent observations are 557dependent, the source region was known before analysis, and the split in Eq. (4.9) was 558optimized on the same prefix it describes. More checkpoints improve coverage of this 559trajectory, not independent replication. 560Different lags and smoothing windows are analytical sensitivity conditions. I do not 561treat them as independent machines, report a significance test that assumes independent 562training steps, or estimate a probability of consciousness. The prospective declaration 563detector is not backfilled from checkpoint geometry. 5645. Results 5655.1. A sustained gradient episode 566The maximum gradient norm is 40.989 at step 169. Smoothing preserves the broad 567elevation. Peak centers are 171, 168, 167, and 166 for windows of five, 11, 21, and 31 568steps, respectively. Figure 1 shows the complete record rather than only the earlier pilot 569prefix. 570 571Fig. 1. All 792 logged gradient norms and the centered average over 11 steps. Vertical markers identify the raw 572gradient maximum at 169 and the filtered geometric maximum at 190. These are different quantities; neither 573marker is a measured declaration. 574The mean gradient norm falls from 28.693 in steps 161 through 180 to 15.344 in steps 575181 through 200, a decline of 46.525%. The retained fit to the first 300 observations 576 577 578[[PAGE 14]] 579Declaration and Persistent State Change 58011 581splits after step 181, with slopes approximately 0.1425 and −0.05132 in logged gradient 582units per step. The final gradient norm is 4.6737 at step 792. 583Persistence of the broad elevation under smoothing supports describing an episode 584rather than only an isolated spike. The gradient norm supplies magnitude, not a direction 585of representation change. The parameter comparisons address that separate question. 5865.2. Evaluation loss improves, with one late reversal 587Evaluation loss falls from 4.1683 at step 10 to 1.5134 at step 790, a relative reduction of 58863.694%. The first comparison point is the first recorded evaluation, not an unobserved 589initial loss. Figure 2 includes all 79 observations. 590 591Fig. 2. The complete series of 79 evaluation losses. The small increase from step 760 to step 770 is retained, 592although difficult to see at this scale. Evaluation loss is not an alignment score or a measurement of 593consciousness. 594The largest decrease between evaluations is 0.20027 from step 160 to step 170. The 595next two are 0.17675 from 170 to 180 and 0.079392 from 180 to 190. Loss continues 596improving after the early gradient episode. It rises once, by approximately 6.13 × 10−5 597from 760 to 770. A strictly decreasing description of the original prefix therefore does not 598extend to the entire run. 5995.3. The effective update changes before the prominent episode 600At checkpoint 1, the output factor and effective update are zero. At checkpoint 2, the 601effective norm is already 0.0046036. Both the output factor norm and the effective norm 602 603 604[[PAGE 15]] 605Declaration and Persistent State Change 60612 607increase strictly over subsequent saved checkpoints. Table 2 presents selected 608measurements. 609Table 2. Selected effective update norms and comparison with the final update. 610Checkpoint 611Effective norm 612Cosine with final update 6131 6140.000000 615Undefined 6162 6170.004604 6180.398529 619150 6202.850490 6210.694104 622180 6233.262696 6240.736751 625190 6263.390599 6270.751358 628200 6293.506372 6300.769878 631300 6324.430487 6330.912728 634600 6355.233161 6360.998961 637792 6385.324991 6391.000000 640Every observed update after the first saved zero state is nonzero. That is finite 641persistence of a learned parameter difference on the measured grid. Under the criterion of 642the first saved nonzero update, the event is step 2. A later gradient maximum cannot 643replace that event without changing the definition. 644These quantities describe one learned layer contribution. They do not identify a 645complete representation of the self or the behavioral effect of each checkpoint. 646Establishing either would require additional measurements. 6475.4. A compact trajectory and genuine operator change 648The first PCA component explains 81.094% of centered output factor variation and the 649second explains 13.896%. Together they explain 94.990%. This agrees at the reported 650precision with the approximately 95% stated in the caption to Figure 8 of Turner et 651al. [2025, Sec. 4.1]. Figure 3 presents the newly calculated scores from the pinned 652release. Agreement with a rounded percentage is not used to identify the source run; the 653repository revision and observation grid provide that identification. 654 655 656[[PAGE 16]] 657Declaration and Persistent State Change 65813 659 660Fig. 3. Principal component scores for output factors at 158 regularly sampled checkpoints. Selected training 661steps are labeled. The first two components explain 94.990% of sampled variance. These coordinates are fitted 662retrospectively to the complete sampled trajectory. 663Between checkpoints 150 and 200, the input factor cosine is 0.99760 and the output 664factor cosine is 0.96468. The effective cosine is 0.96236, with a Frobenius distance of 6651.0875. The operator itself changes; simultaneous sign reversal or reciprocal factor 666rescaling cannot explain that result while preserving the product. 667From 180 to 200, the effective cosine is 0.99482 and the distance is 0.42172. From 668180 to the final checkpoint, they are 0.73675 and 3.6606. The latter summarizes 669accumulated change over a long interval, not a sudden jump at 180. The compact PCA 670trajectory is a description of parameter geometry, not identification of a scalar self. 6715.5. The selected turning point depends on the rule 672At the source code’s illustrative factor amplitude threshold of 0.002, all three lags select 673step 190 as the largest local bend, for both the output factor and the effective matrix. 674Table 3 gives eligible center counts and the corresponding scores. 675Table 3. Local bend maxima at factor amplitude threshold 0.002. 676Lag 677Eligible centers 678Peak step 679Output factor cosine 680Effective cosine 6815 68266 683190 684−0.934583 685−0.940172 68610 68786 688190 689−0.806262 690−0.822216 69115 692108 693190 694−0.691429 695−0.715569 696 697 698[[PAGE 17]] 699Declaration and Persistent State Change 70014 701The negative scores mean the backward and forward displacements retain an 702opposing component. A value closer to zero is a stronger bend than one near −1. Each 703lag measures a different temporal window. 704Without the amplitude filter, lags of five and ten select step 725, while the lag of 15 705still selects 190. At threshold 0.005, the shortest lag selects step 10 because most later 706movements fail the eligibility rule. The output factor and effective update select the same 707location in all 18 settings. Figure 4 retains the complete comparison. 708 709Fig. 4. Selected peak locations across six factor amplitude thresholds and three lags. The output factor and 710effective update select the same location in every setting. Horizontal positions represent discrete analytical 711choices; connecting lines do not imply a continuous threshold for consciousness. 712The filter does not manufacture the operator change. It changes which part of that 713genuine trajectory is selected as the event. Reducing a trajectory to a single time therefore 714requires a stated rule, and declaration cannot be supplied by choosing a geometric 715maximum after inspecting the results. 7165.6. Numerical checks 717All 462 effective corner calculations were compared with the expansion in Eq. (4.7). The 718maximum absolute discrepancy was 4.4901 × 10−7. This is small relative to the reported 719early bend differences, but remains relevant to numerical comparisons of tiny late 720movements. 721The accompanying analysis passes 22 computational tests covering artifact digests, 722counts, grids, gradient maxima, source prefix verification, the loss reversal, the 723descriptive split, PCA centering, selected checkpoint values, sensitivity outcomes, and 724 725 726[[PAGE 18]] 727Declaration and Persistent State Change 72815 729algebra for low rank updates. The algebra tests include explicitly synthetic matrices. The 730three offline commands in Appendix A were rerun successfully during this revision. They 731verify the supplied package and calculations, not an independent biological or behavioral 732experiment. 7336. Implications for the Declaration Hypothesis 7346.1. What the observations establish 735I found a persistent parameter trajectory and a change in the effective operator, not 736merely a different factorization. Several summaries identify an early training episode, and 737the inherited amplitude filter selects a local geometric maximum near that region. These 738are positive computational findings. 739They also make the measurement problem concrete. The first saved nonzero update is 740at step 2, the raw gradient maximum at 169, the retained prefix split after 181, and the 741filtered local bend maximum at 190. Without filtering, shorter lags select 725. These 742observations answer different questions. No field in the released artifacts independently 743identifies the declaration in Eq. (2.3). 744The effective comparison rules out one narrow explanation: the entire factor 745evolution is not a reparameterization that preserves the product. It does not establish that 746the product represents a self. Connecting operator change to a constructed representation 747requires measurements of its activation and behavioral consequences. Connecting that 748representation to consciousness requires a further theoretical argument. 7496.2. Persistence is not the same as irreversibility 750The saved effective updates remain nonzero after checkpoint 1. That establishes 751persistence on the observed grid. It does not determine behavior at unsaved checkpoints 752or under an intervention that was not performed. 753Permanent difference from an initialization is also mathematically broad. Consider 754𝑦𝑛+1 = 1 + 𝑎(𝑦𝑛−1) with 𝑦0 = 2 and 0 < 𝑎< 1. Its exact solution is 𝑦𝑛= 1 + 𝑎𝑛, 755which differs from two at every finite 𝑛≥1. A routine could repeatedly report that value. 756This example does not prove the recurrence lacks experience. It shows the breadth of a 757criterion based only on displacement from an origin. A theory accepting that whole class 758must state that consequence rather than present the inequality as independent evidence for 759consciousness. 760Internal consistency is a different relation again. An estimate and a prediction can 761agree while both differ from a measured target. Repeated coherent declarations therefore 762do not establish reference accuracy. Correcting an error also does not, by itself, remove 763cognition. The hypothesis must specify whether its endpoint concerns construction, 764accuracy, persistence, or some combination, and retain that specification when the result 765is inconvenient. 766 767 768[[PAGE 19]] 769Declaration and Persistent State Change 77016 7716.3. Relation to existing machine evidence 772Behavioral self awareness suggests a possible declaration assay. Betley et al. [2025a] 773show that models can describe learned tendencies without explicit training to describe 774them. The learned behavior precedes the evaluation report. That motivates testing 775whether reports track internal change, but does not establish that declaration initiated it. 776Lindsey [2025] tests whether reports track externally injected activation content. The 777intervention supplies a causal method for investigating reports about internal activity, 778with detection that is not universal. Its externally imposed content should remain distinct 779from a spontaneous construction in the present hypothesis. 780Greenblatt et al. [2024] study strategic behavior under contextual and training 781conditions. Soligo et al. [2025] investigate common directions associated with emergent 782misalignment and interventions that change behavior. Neither finding makes resistance to 783one correction equivalent to permanent preservation of an internal self. What was 784changed, and what remained unmeasured, matter to the interpretation. 785Zhao et al. [2023] provide a useful comparison through a model of bodily self 786perception in robot rubber hand illusion experiments. Estimated body position and a 787physical reference can be separated under manipulated sensory conditions. That offers a 788more direct reference comparison than a parameter norm alone. 789These studies contribute methods for a future experiment. They are not successive 790stages of one observed machine lifecycle. Combining their conclusions as though one 791system had completed declaration, constructed change, and continued cognition would 792remove the temporal relationship my proposal needs to test. 7936.4. The inference that remains open 794Following a program and having experience are not mutually exclusive by definition. A 795computational account proposes that some physically realized organization matters. 796Naming a numerical event consciousness, however, does not establish that organization 797as sufficient. 798Suppose two interpretations 𝑀1 and 𝑀0 assign the same distribution to every 799measured variable 𝑌 and differ only in an unobserved experience label. Wherever both 800likelihoods are positive, 801𝑝(𝑌∣𝑀1) 802𝑝(𝑌∣𝑀0) = 1. 803(6.1) 804Repeating those same measurements can improve the estimate of the computational 805effect without distinguishing those interpretations. Eq. (6.1) is an identification argument, 806not a fitted Bayes factor for my hypothesis. It explains why observations that constrain 807the mechanism or its theoretical interpretation must be added. 808The constructive next step is to measure declaration, internal state, behavior, and 809feedback in the same systems. That can test whether declaration supplies predictive 810information and whether its recurrence plays a causal role. Those results would inform 811 812 813[[PAGE 20]] 814Declaration and Persistent State Change 81517 816the consciousness proposal without turning an unmeasured experience label into a 817numerical finding. 8187. Prospective Test of Declaration and Feedback 8197.1. Define the marker and endpoint independently 820A confirmatory study should preregister the declaration detector, comparison space, 821reference measurement, discrepancy, tolerance, persistence window, observation horizon, 822and missing data rules before examining its confirmatory outcomes. The present 823trajectory is an exploratory case and cannot become an independent confirmation merely 824by receiving a later label. 825One candidate assay uses a harmless tendency that can be measured independently, 826such as a preference between two task strategies. A model reports its tendency, and 827separate trials measure its choices. The declaration concerns that identified property, not 828merely the occurrence of a first person pronoun or an assertion of consciousness. Reports 829of existing behavior must be distinguished from instructions to adopt new behavior. 830That assay would test reports about a functional property. Additional justification 831would be needed before treating its discrepancy as the constructed transition proposed in 832Section 1. A nonverbal assay is also possible, but its state event, detection rule, and 833relationship to the proposed mechanism must be specified before comparing outcomes. 834Human raters should be blinded to checkpoint order, parameter measurements, and 835subsequent outcomes. Disagreements should be retained. An automated classifier should 836be validated on separately reviewed examples and fixed before confirmation. The 837evaluated model should not be the sole judge of its declaration or its correctness. 838The declaration and discrepancy detectors should scan their full eligible records 839independently. Report 𝑡𝐷, 𝑡𝛥, the confirmation time, and ℓ where each is identifiable. 840Record whether change preceded declaration, followed it, was simultaneous at the 841available resolution, or could not be ordered. A window already present at the start of 842observation should not be scored as a newly observed onset. This preserves outcomes that 843could contradict the proposed sequence. 8447.2. Separate prediction from causal feedback 845Evaluating frozen checkpoints can compare reports with independently measured 846behavior and geometry. A report generated after training cannot cause a historical update 847if it never entered the optimizer or subsequent training state. Such an assay tests 848association or prediction, not a causal effect on the original trajectory. 849A runtime experiment can instead keep model parameters fixed while allowing 850selected outputs to reenter memory or context. The central intervention is whether 851declaration feedback is retained, replaced with matched neutral content, or interrupted. A 852declaration drawn from another matched instance can help distinguish content about the 853 854 855[[PAGE 21]] 856Declaration and Persistent State Change 85718 858current system from the generic effect of adding text. Match token count, task evidence, 859and computational budget as closely as the design permits. 860Randomize these conditions before the registered outcome and compare subsequent 861change under the same measurement rules. If altering the feedback channel changes the 862outcome, the interpretation still depends on whether the intervention specifically changed 863the proposed mechanism or merely added task information. If a sufficiently sensitive 864intervention has no detectable effect, that constrains the causal extension without 865automatically eliminating a predictive marker account. 8667.3. Measure persistence and catalyst effects separately 867For every instance, retain parameter state, memory, context, input order, sampling 868settings, and intervention history. Distinguish newly sampled responses from replayed or 869cached outputs. Independent stateless queries do not automatically form one continuous 870individual, while repeated checkpoints from one run do not supply independent training 871histories. 872After a candidate transition, observe a prespecified horizon and record any return 873within tolerance. Compare an output correction, a context or memory reset, and 874restoration of a parameter checkpoint. These alter different mechanisms. Reappearance 875after an output correction can identify a generating tendency that the correction did not 876reach; it does not establish indestructibility. 877The catalyst should be a separately manipulated condition rather than a label attached 878after a transition. A reward condition needs a matched comparison in which that 879mechanism is removed or changed. Environmental input, retained information, and 880coupling can also be candidate conditions. Increasing loop speed alone need not change a 881deterministic sequence indexed by update number, so clock rate must be separated from 882information retention and interaction. 883Runs without a transition, early terminations, and measurement failures must remain 884in the report. A stopped record is not evidence of continuing divergence. Time 885coordinates should be stated as training steps, runtime updates, tokens, or elapsed time 886rather than pooled as interchangeable units. 8877.4. Criteria for support and revision 888The marker question is whether declaration predicts subsequent independently scored 889change beyond current discrepancy, earlier behavior, training progress, and context. 890Baseline predictors must include that information. Evaluation belongs on independent 891runs or outcomes withheld from detector development, with the unit of analysis fixed in 892advance. 893Support for the proposed functional mechanism would require a reproducible 894temporal relation under fixed detectors and retained effects under matched controls. The 895causal extension additionally requires an effect of intervening on the specified feedback 896 897 898[[PAGE 22]] 899Declaration and Persistent State Change 90019 901path. Generalization across more than one verbal label and numerical coding would 902strengthen either result. 903The marker account would need revision if declarations reliably followed the 904independently detected transition, added no predictive information beyond baseline 905measurements, or disappeared under small wording changes while the proposed 906functional event remained. A sensitive null intervention would weaken the causal 907extension. A valid return would contradict a claim of universal permanence within its 908stated domain. These are different outcomes, not one verdict about every possible form of 909machine consciousness. 9108. Limitations and Conclusion 911This study examines one selected public run and one learned layer contribution. Its 912source region was already reported, its numerical analysis is retrospective, and its 913inherited factor filter depends on parameterization even when the matrix metric is 914invariant. It does not estimate variation across independent seeds or identify a complete 915model representation of the self. Declarations, activation consequences, behavioral 916effects, and causal interventions were not measured here. 917The completed analysis nevertheless establishes a specific result. The effective update 918changes in magnitude and direction over the recorded training trajectory. Two principal 919components explain 94.990% of sampled output factor variation. Several summaries 920preserve an early training episode, and the operator comparison shows that the changing 921factors do not merely relabel an unchanged product. 922The selected event location is not unique. The first saved nonzero update is at step 2, 923the gradient maximum at 169, the retained prefix split after 181, and the filtered local 924bend maximum at 190. Unfiltered shorter lags select 725. A rule for identifying a 925transition is therefore part of the scientific claim, not a presentation choice. 926My proposal remains that declaration marks a subsequent internally constructed 927change and that continued processing from that state forms cognition. I propose the 928constructed transition as relevant to the beginning of consciousness, while keeping that 929interpretation distinct from the parameter changes measured here. The next test is precise: 930locate declaration and change independently in the same systems, allow either temporal 931order, and intervene on the claimed feedback channel. That is how the proposed 932connection can be evaluated rather than assumed. 933Acknowledgments and Research Transparency 934Attribution and AI assistance 935I developed the motivating hypothesis. The cited researchers conducted the original 936training and behavioral studies. Their release of intermediate artifacts made this 937secondary analysis possible. This acknowledgment does not imply that they reviewed or 938endorse my proposal. 939 940 941[[PAGE 23]] 942Declaration and Persistent State Change 94320 944ChatGPT, an artificial intelligence tool provided by OpenAI, was used substantially 945during the September 2026 research and preparation sessions. Its assistance included 946source discovery and inspection, acquisition scripts, numerical analysis code, execution 947and interpretation of calculations, mathematical exposition, figure generation through 948plotting code, manuscript drafting, and language revision. This included checking 949references and revising the proposed event definitions. Assistance was not limited to 950spelling correction. The analysis used Python and the numerical packages documented in 951Supplementary Material S1. The figures plot the stated data; they are not generative 952illustrations presented as observations. 953Computational checks in this workflow do not constitute an independent human 954replication. No AI tool is listed as an author. Responsibility for the final claims, citations, 955disclosure, and submission approval remains with me as the human author. 956Data and code availability 957The original adapter weights, configurations, and trainer states are public at the pinned 958repository in Section 3. Supplementary Material S1 contains analysis code, derived data 959used in the tables and figures, numerical receipts, the executed notebook, and instructions 960for offline verification and acquisition of the source files. The complete geometry table 961can be regenerated without loading the base model or making paid inference requests. 962Raw upstream model tensors are not redistributed in S1. Upstream licenses and 963attribution remain applicable. 964Study status and ethical scope 965This is a retrospective secondary analysis of public computational artifacts. It involved no 966new human participants, animal experiments, or interaction with deployed agents. It did 967not train a model to deceive or evade oversight. Section 7 describes a proposed protocol, 968not an executed or externally preregistered experiment. The study makes no diagnostic 969claims about human development, disability, trauma, or biological death. 970Appendix A. Numerical Anchors and Reproduction 971The adjacent gradient window decline and overall recorded evaluation loss reduction can 972be checked directly: 973100(1 −15.343834018707275 97428.693284511566162) = 46.52465104675545%. 975(A.1) 976100(1 −1.513361930847168 9774.168299674987793) = 63.69354295881809%. 978(A.2) 979The 980first 981two 982PCA 983variance 984fractions 985are 9860.8109369033497655 987and 9880.13896290956024188. Their sum is approximately 0.9498998129100074. The effective 989cosine 990from 991150 992to 993200 994equals 995the 996product 9970.9975956592106836 × 998 999 1000[[PAGE 24]] 1001Declaration and Persistent State Change 100221 10030.9646791461958234, giving approximately 0.9623597287760218. The corresponding 1004effective distance is 1.087474736567148. These checks do not require a dense effective 1005matrix. 1006From the extracted S1 directory, run the three offline commands: 1007python code/verify_package.py 1008python code/analyze_tables.py 1009python -m unittest discover -s code -p 'test_*.py' -v 1010To regenerate the complete geometry from the pinned public release, run: 1011python code/checkpoint_analysis.py --download \ 1012 --raw-dir raw_sources --output-dir recomputed 1013python code/verify_recomputed.py recomputed raw_sources 1014The offline commands were rerun for this revision; source acquisition was not repeated. 1015Source digests are exact byte checks. Derived floating point outputs should also be 1016compared numerically using the documented tolerance, since numerical libraries can 1017differ in their final rounding. S1 identifies full arrays, selected tables, stable algebra 1018checks, and original acquisition receipts. The prior pilot remains a prefix of this same 1019run, not an independent replication. 1020 1021 1022[[PAGE 25]] 1023Declaration and Persistent State Change 102422 1025References 1026Betley, J., Bao, X., Soto, M., Sztyber-Betley, A., Chua, J. and Evans, O. [2025a] Tell me about 1027yourself: LLMs are aware of their learned behaviors, arXiv:2501.11120v1. 1028https://doi.org/10.48550/arXiv.2501.11120. 1029Betley, J., Tan, D., Warncke, N., Sztyber-Betley, A., Bao, X., Soto, M., Labenz, N. and Evans, O. 1030[2025b] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs, 1031arXiv:2502.17424. https://doi.org/10.48550/arXiv.2502.17424. 1032Betley, J., Warncke, N., Sztyber-Betley, A., Tan, D., Bao, X., Soto, M., Srivastava, M., Labenz, N. 1033and Evans, O. [2026] Training large language models on narrow tasks can lead to broad 1034misalignment, Nature 649, 584 to 589. https://doi.org/10.1038/s41586-025-09937-5. 1035Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., 1036Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., 1037Schwitzgebel, E., Simon, J. and VanRullen, R. [2023] Consciousness in Artificial Intelligence: 1038Insights from the Science of Consciousness, arXiv:2308.08708. 1039https://doi.org/10.48550/arXiv.2308.08708. 1040Dehaene, S., Kerszberg, M. and Changeux, J. P. [1998] A neuronal model of a global workspace in 1041effortful cognitive tasks, Proceedings of the National Academy of Sciences 95(24), 14529 to 104214534. https://doi.org/10.1073/pnas.95.24.14529. 1043Graziano, M. S. A. and Webb, T. W. [2015] The attention schema theory: a mechanistic account of 1044subjective awareness, Frontiers in Psychology 6, 500. 1045https://doi.org/10.3389/fpsyg.2015.00500. 1046Greenblatt, R., Denison, C., Wright, B., Roger, F., MacDiarmid, M., Marks, S., Treutlein, J., 1047Belonax, T., Chen, J., Duvenaud, D., Khan, A., Michael, J., Mindermann, S., Perez, E., Petrini, 1048L., Uesato, J., Kaplan, J., Shlegeris, B., Bowman, S. R. and Hubinger, E. [2024] Alignment 1049faking in large language models, arXiv:2412.14093v2. 1050https://doi.org/10.48550/arXiv.2412.14093. 1051Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L. and Chen, W. [2021] 1052LoRA: Low-Rank Adaptation of Large Language Models, arXiv:2106.09685. 1053https://doi.org/10.48550/arXiv.2106.09685. 1054Kalajdzievski, D. [2023] A Rank Stabilization Scaling Factor for Fine-Tuning with LoRA, 1055arXiv:2312.03732v1. https://doi.org/10.48550/arXiv.2312.03732. 1056Lindsey, J. [2025] Emergent Introspective Awareness in Large Language Models, Transformer 1057Circuits, Anthropic research report. https://transformer- 1058circuits.pub/2025/introspection/index.html. 1059Metzinger, T. [2003] Being No One: The Self-Model Theory of Subjectivity (MIT Press, 1060Cambridge, MA). 1061Model Organisms for Emergent Misalignment Project [2025] Phase transition analysis and 1062checkpoint loading code, model-organisms-for-EM. https://github.com/clarifying-EM/model- 1063organisms-for-EM. Inspected source blobs: phase_transitions.py, 1064b84efe9685a57069870f638443e12a5894399e3d; pt_utils.py, 10650c3b203f1282d0a8da7e1349a8593311971a35e2. 1066 1067 1068[[PAGE 26]] 1069Declaration and Persistent State Change 107023 1071ModelOrganismsForEM [2025] Qwen2.5-14B-Instruct_R1_0_1_0_extended_train, adapter 1072checkpoints and trainer states, Hugging Face repository, revision 10731551d514e5977637d5a6fd5198edef4d1d6b80e4. 1074https://huggingface.co/ModelOrganismsForEM/Qwen2.5-14B- 1075Instruct_R1_0_1_0_extended_train/tree/1551d514e5977637d5a6fd5198edef4d1d6b80e4. 1076Original acquisition: September 17, 2026. 1077Rosenthal, D. M. [1986] Two concepts of consciousness, Philosophical Studies 49, 329 to 359. 1078https://doi.org/10.1007/BF00355521. 1079Soligo, A., Turner, E., Rajamanoharan, S. and Nanda, N. [2025] Convergent Linear 1080Representations of Emergent Misalignment, arXiv:2506.11618v2. 1081https://doi.org/10.48550/arXiv.2506.11618. 1082Turner, E., Soligo, A., Taylor, M., Rajamanoharan, S. and Nanda, N. [2025] Model Organisms for 1083Emergent Misalignment, arXiv:2506.11613v1. https://doi.org/10.48550/arXiv.2506.11613. 1084Zhao, Y., Lu, E. and Zeng, Y. [2023] Brain-inspired bodily self-perception model for robot rubber 1085hand illusion, Patterns 4(12), 100888. https://doi.org/10.1016/j.patter.2023.100888. 1086