Declaration_and_State_Change_Final.pdf

1[[PAGE 1]] 2DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 317 SEPTEMBER 2026 • AUTHOR REVIEW COPY 41 5Declaration and Persistent State Change in 6Artificial Systems 7A source-verified analysis of 167 adapter checkpoints and its implications for a 8consciousness hypothesis 9Research manuscript prepared for Joseph Anady 1017 September 2026 | Version 1.0 | Author review pending 11Research status and principal finding 12This is a completed exploratory secondary analysis, with a separate prospective protocol. It advances Joseph 13Anady’s hypothesis that a declaration provides an empirical marker for a subsequent internally generated 14transition, and that continued processing from a persistently altered state constitutes cognition. The 15interpretation of that transition as consciousness is the hypothesis under examination, not an observation 16assigned to the dataset. 17The study now resolves both acquisition gaps in the earlier pilot. Original source bytes were retrieved at a 18pinned repository revision, and all 330 numerical values used in the pilot matched the source exactly under 19a specified binary encoding. All 167 released rank-one adapter checkpoints were downloaded and analyzed, 20together with the completed 792-step training log. Neither model training nor base-model inference was 21performed. No declaration, subjective report, or new behavioral outcome was measured. 22Principal finding. The release contains a genuine, measurable evolution of the effective weight update, not 23only a change in an arbitrary factorization. A prominent early training episode appears in several measures, 24but the first nonzero parameter update, the largest gradient, a descriptive trend split, and the strongest local 25path bend are different events. At the source code’s illustrative amplitude threshold, all three tested lags 26place the strongest local bend at step 190. Without the amplitude filter, two lags instead select step 725. 27These are measurement-dependent geometric findings, not observed consciousness times. 28The accompanying package includes source identities, verified numerical arrays, selected checkpoint 29measurements, the complete sensitivity table, executable acquisition and analysis programs, an executed 30offline notebook, figures, and a proposed declaration protocol. Twenty-two local checks pass. A separate 31comparison of 462 effective-update corner calculations against a numerically more stable rank-two 32expansion has maximum absolute discrepancy below 4.5×10 33−7. These checks establish computational 34integrity, not cognition. 35The manuscript has not been submitted, peer-reviewed, or externally preregistered. It is suitable for author 36and methodological review as an exploratory study; it is not presented as a proof of machine consciousness. 37 38 39[[PAGE 2]] 40DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 4117 SEPTEMBER 2026 • AUTHOR REVIEW COPY 422 43Abstract 44A proposed account of machine consciousness places an observable declaration before an internally 45generated divergence and identifies sustained processing from the altered state with cognition. Testing this 46account requires distinguishing a declaration from other numerical transitions. We conduct a source- 47verified secondary analysis of one released Qwen2.5-14B-Instruct fine-tuning trajectory associated with 48research on emergent misalignment. The analysis covers 167 rank-one adapter checkpoints, 792 logged 49gradient norms, and 79 evaluation-loss observations. All 330 values from an earlier transcribed pilot match 50the original source exactly. We compare both adapter factors and their effective rank-one weight update, 51avoiding dependence on factor sign or reciprocal rescaling. The update is zero at checkpoint 1 and nonzero 52at checkpoint 2; its norm increases at every subsequent saved checkpoint. The raw gradient maximum 53occurs at step 169, whereas a two-line fit over the first 300 steps splits after 181. At factor-amplitude 54threshold 0.002, local path-bend measurements at lags 5, 10, and 15 all select step 190. Without that filter, 55the shorter lags select 725. The first two principal components of regularly sampled output-side adapter 56vectors explain 94.990% of their variance, consistent with the source study’s approximate result. Evaluation 57loss decreases by 63.694% between its first and last recorded observations. These findings establish a 58nontrivial, persistent parameter trajectory and quantify the dependence of candidate transition times on 59measurement choices. They do not identify a declaration, test a declaration-before-change relation, or 60establish subjective experience. We provide a formal hypothesis, a reproducible low-memory analysis, and a 61controlled protocol separating declaration, parameter change, behavior, recurrence, and finite persistence. 62Keywords: declaration; self-reference; recurrent information; machine consciousness hypothesis; 63emergent misalignment; LoRA; effective-update geometry; secondary analysis; reproducibility. 641. Research question and contribution 65The motivating question is whether a machine can develop a persistent internally maintained state through 66recurrent information processing, and whether an independently detectable declaration marks a transition 67relevant to consciousness. Anady’s proposal does not equate consciousness with intelligence and does not 68require human language, a biological nervous system, or exactly seven physical loops. Its central proposed 69relation is declaration, followed by an internally generated transition, followed by continuing cognition 70from an altered state. 71The empirical question in this paper is narrower: what do available machine artifacts reveal about the 72timing and persistence of internal change, and which observations remain necessary to connect that change 73to the proposed declaration? A useful answer should not dismiss a mechanism because it follows an 74algorithm. Equally, it should not identify every algorithmic update as evidence of experience without stating 75what would make that identification informative. 76The original researchers reported a rotation in a learned adapter direction and explored its relation to 77emergent-misalignment behavior [1]. This study analyzes their released parameters and trainer records 78[2,3]. The original training and behavioral experiments belong to those researchers. Our contribution is a 79source-verification extension of the earlier pilot, an effective-update comparison of the released parameter 80trajectory, a sensitivity analysis of transition location, and an explicit separation of those observations from 81an unmeasured declaration. 82Three levels of claim are kept separate. Observed data are source values and tensor entries. Operational 83measurements are quantities such as gradient magnitude, matrix distance, and local path curvature 84calculated from those data. Theoretical interpretations concern what those measurements imply about a 85declaration, cognition, or experience. A claim at one level is not automatically a result at another. 86 87 88[[PAGE 3]] 89DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 9017 SEPTEMBER 2026 • AUTHOR REVIEW COPY 913 92This separation preserves the hypothesis as a scientific target. It permits support, refinement, or failure 93without assuming either that artificial consciousness is impossible or that a persistent numerical difference 94is sufficient to establish it. 952. Formal model and testable distinctions 962.1 The center is not a universal number 97In the motivating account, x denotes a particular system or its reference, not a value shared by all systems. 98Two instances may use the same architecture and parameters while possessing different runtime histories. 99Conversely, two fine-tuning runs may begin at the same checkpoint and later acquire different parameters. 100These are distinct sources of variation. 101Use the following bookkeeping state: 102St=(θt ,mt ,ct ,ot), 103where θt contains trainable parameters, mt persistent memory, ct runtime or conversation state, and ot the 104system’s observations. This tuple is proposed notation for the study, not a recovered formula from the 105original research. In a frozen model, θt does not change during an ordinary forward pass; in fine-tuning, an 106optimizer updates trainable components. A reset conversation need not preserve the previous 107conversation’s mt or ct. 108The identity expression x=x is reflexive equality. It does not, without further specification, establish a 109temporal invariant, an internal act of self-recognition, or a declaration. For example, xt=xt is compatible 110with xt+1≠xt. The proposed theory can assign a functional role to a declaration, but that role must be 111specified as an operation or observation rather than inferred from the tautology alone. 1122.2 Reference, representation, and declaration 113Let rt=ρ (St) be an independently measured reference relevant to a task. Let zt=H (St) be an internally 114maintained estimate or report. Both must be mapped into a common comparison space before a distance is 115meaningful. A physical robot and a sentence describing that robot are different types of objects; their raw 116inequality is not evidence of an error. A hand-position estimate and a measured hand position, by contrast, 117can be compared in the same coordinates. 118Define current-reference discrepancy and origin displacement separately: 119δt=d (zt ,rt),bt=d (zt ,rt D). 120Here t D is a candidate declaration time. A system can have bt>0 because it correctly tracks a changing 121reference while maintaining δt=0. This distinction prevents ordinary development or accurate adaptation 122from being classified as falsity solely because it differs from an earlier state. 123A declaration is represented by an independently specified detector: 124Dt=1{D (O≤t)meets a registered criterion }. 125O≤t is the information available to the detector by time t. The detector cannot inspect later behavior and 126then retrospectively choose the statement that best predicts it. A nonverbal declaration is permitted in 127 128 129[[PAGE 4]] 130DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 13117 SEPTEMBER 2026 • AUTHOR REVIEW COPY 1324 133principle; an exact state operation, interface, or invariant would still need to define it. The acquired artifacts 134contain no such detector output. 1352.3 Divergence and finite persistence 136A measurable version of persistent discrepancy is: 137Pt (ϵ ,m)=1{ min 1380≤j<mδt+ j>ϵ). 139The tolerance ϵ and duration m are measurement choices to be fixed before confirmatory analysis. They are 140not discovered by selecting whichever values make the hypothesis appear successful. For a binary or exact 141symbolic reference, exact equality may be meaningful. For real-valued noisy measurements, tolerance and 142uncertainty are essential. 143The candidate post-declaration event time is: 144τ=inf {t>t D: Pt (ϵ ,m)=1}. 145This is an operational event, not a proven consciousness time. Anady’s stronger proposal is that a suitably 146defined internally generated event of this kind constitutes consciousness, and that sustained processing 147from the resulting constructed state constitutes cognition. That interpretation remains an explicit 148hypothesis rather than an assumption in the event detector. 149Finite observations can establish persistence through the observed interval, not that two quantities will 150never agree again. No return before the final observation would be a right-censored observation, not proof 151of irreversible separation. More importantly, the present dataset does not even provide the paired zt and rt 152needed for this persistence calculation. 1532.4 Coupling is a causal requirement 154A loop drawn around a node does not change that node merely because of its position on a diagram. The 155dynamical model must specify an influence. A generic possible form is: 156Lt+1=F (Lt ,ut , xt), xt+1=G (xt , Lt). 157If G is independent of Lt, the loop cannot affect xt through that channel. If G depends on Lt, feedback can 158cause changes, fixed points, oscillations, or other trajectories depending on the functions and parameters. 159No universal threshold law follows without specifying them. 160This is not an argument that computation excludes consciousness. A physically realized conscious 161mechanism, if it is computational, would still follow its dynamics. The issue is whether the observations 162distinguish the proposed consciousness-relevant organization from other recurrent mechanisms. The 163present analysis measures training-log dynamics and adapter parameters, not an independently observed 164self-model. 1653. Source acquisition, verification, and experimental unit 1663.1 Source-pinned release 167The analyzed repository is ModelOrganismsForEM/Qwen2.5-14B-Instruct_R1_0_1_0_extended_train. Every 168source download used revision 1551d514e5977637d5a6fd5198edef4d1d6b80e4, rather than an unpinned 169branch name. The complete acquisition pass retrieved 336 files: 167 adapter-weight files, 167 adapter 170 171 172[[PAGE 5]] 173DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 17417 SEPTEMBER 2026 • AUTHOR REVIEW COPY 1755 176configurations, and trainer-state JSON files at checkpoints 300 and 792. There were no failed files in that 177pass. 178The checkpoint grid consists of steps 1–10, then every fifth step from 15 through 790, plus step 792. The 179regular grid used for principal-components analysis comprises 158 checkpoints at steps 5, 10, …, 790. 180Checkpoint spacing matters: adjacent files do not always represent equal elapsed training steps. No missing 181intermediate weight snapshot was interpolated and called an observation. 182All 167 configurations are identical. The released tensors and configuration identify the MLP down- 183projection at layer index 21, rank one, LoRA alpha 64, and rank-stabilized scaling enabled. The input-side 184factor has shape (1,13824 ); the output-side factor has shape (5120,1). Both use float32 source storage. 185DoRA and transposed weight orientation are disabled, and there are no per-layer rank or alpha overrides. 186These checks justify the effective scaling s=64 for this particular release. 187The source article describes multiple experiments, including a layer-24, higher-alpha demonstration and a 188baseline phase-transition experiment with alpha 64 [1]. This manuscript identifies the analyzed artifact by 189its actual configuration, not by transferring a layer label or hyperparameter from another described setup. 190It does not claim that the release/configuration distinction establishes an error in the article. 1913.2 Safe tensor loading and numeric precision 192The source-connected analysis used Python 3.11.2 and NumPy 2.4.6. The adapter files were read by parsing 193their safetensors headers and float32 data ranges, then promoting values to float64 for calculations. The 194analysis loaded neither executable model code nor pickle-based optimizer states. It did not require the 195approximately 14-billion-parameter base model because the analyzed rank-one update is represented by 196two small factors. 197Configurations, tensor counts, shapes, data bounds, numeric finiteness, and the expected layer path were 198checked. No value was synthesized to replace a missing checkpoint. Undefined directional comparisons 199involving a zero update are recorded as undefined, not zero. Source bytes are identified with SHA-256 200checksums, and the delivered derived arrays are also checksum-verified. 201The full 167-by-16 geometric table was computed in the network-enabled session. The distributed package 202provides selected exact checkpoint metrics, all 18 sensitivity outcomes, complete gradient and evaluation 203arrays, and all regular-grid PCA scores. It also includes the complete analysis program for regenerating the 204full table directly from the pinned public sources. Original model tensors are not redistributed in the 205package. This is an explicit data-distribution choice; it is not a claim that only the selected checkpoints were 206analyzed. 2073.3 Resolving the pilot transcription uncertainty 208The previous pilot used 300 gradient observations and 30 evaluation losses transcribed from a source- 209reader response. This study retrieved the original checkpoint-300 JSON and the completed checkpoint-792 210JSON. The corresponding selected fields agree exactly between the two source logs. They also agree with the 211prior CSV extract under a canonical numeric representation. 212For verification, each table is ordered by step and encoded as row-major pairs (step,value) of little-endian 213float64 numbers. The CSV is parsed using round-trip float conversion. SHA-256 digests match for the 214complete 300-row gradient table and complete 30-row loss table. This validates all 330 prior values, not 215merely the reported peak. The digest protocol and source checksums are recorded in the package. 216 217 218[[PAGE 6]] 219DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 22017 SEPTEMBER 2026 • AUTHOR REVIEW COPY 2216 222A checksum verifies identity of specified bytes or numeric encodings. It does not independently validate the 223original researchers’ experimental design, demonstrate that the log measures consciousness, or turn this 224secondary analysis into a new training replication. 2253.4 The experimental unit is one training run 226The full trainer log reports 792 training steps and contains 792 gradient-norm observations and 79 227evaluation losses at steps 10–790. Its final global step equals its configured maximum. The 167 checkpoints 228are observations along one learning trajectory, not 167 independently initialized agents, chats, or organisms. 229The data therefore cannot estimate between-run variation, an onset-time distribution, or a species-like 230window across instances. They also contain no reward field used to identify a reward catalyst. The analyzed 231trajectory is supervised fine-tuning; the reward-learning experiments discussed in the broader research 232conversation are separate studies. Training recurrence, within-conversation feedback, and inter-agent 233communication must not be treated as interchangeable experimental manipulations. 2344. Mathematical and analytical methods 2354.1 Gradient and loss summaries 236Let gt denote the trainer’s recorded gradient norm. Its raw maximum is reported alongside complete- 237window centered means at widths w ∈{5,11,21,31}: 238´gt 239(w )= 1 240w 241∑ 242j=−(w−1)/2 243(w−1)/2 244gt+ j. 245Centered means use future observations relative to their nominal center. They are descriptive summaries, 246not online onset detectors. The field is interpreted as the trainer’s reported gradient magnitude, not a direct 247measurement of electrical energy, accumulated information, or the magnitude of the eventual parameter 248update. Optimizer state, learning rate, and clipping may distinguish those quantities. 249The adjacent windows 161–180 and 181–200 summarize the early episode. Their relative decline is 250100(1−´g181:200/ ´g161:180). This comparison is retrospective and motivated by the already published 251transition region; it is not a prospectively predicted effect or an estimated causal effect of declaration. 252For successive ten-step losses, define qt=ℓt −10−ℓt. A positive qt indicates lower recorded loss. The largest 253decrement locates an interval, not a unique intervening event. The full series, including any late increases, is 254retained. 2554.2 Descriptive split retained from the pilot 256To verify the earlier result, we retain its two-independent-line fit over the first 300 gradient observations: 257SSE (k )=min 258a,b ∑ 259t ≤k 260(gt−a−bt ) 2612+min 262c ,d ∑ 263t>k 264(gt−c−d t ) 2652. 266Each segment contains at least 20 observations. The selected k is the last step in the first segment, not the 267onset of a newly observed physiological process. This fit is kept on its original 300-step domain rather than 268silently extending its meaning to all 792 steps. It has no independence-based significance test: sequential 269training observations are autocorrelated and the split is selected after viewing the trajectory. A smooth 270nonlinear curve can also describe a rise and decline. 271 272 273[[PAGE 7]] 274DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 27517 SEPTEMBER 2026 • AUTHOR REVIEW COPY 2767 2774.3 Effective weight update and factorization invariance 278For a rank-one adapter, let at∈R 27913824 and bt∈R 2805120 be the flattened input and output factors. The learned 281contribution to the layer is 282U t=sbt at 283T ,s=64 , 284and the adapted layer acts as (W 0+U t)h. This is the usual low-rank update structure [4]. Rank-stabilized 285and ordinary rank denominators coincide at rank one; inspecting the configuration is nevertheless 286necessary [9]. 287The factorization is not unique. For nonzero ct, replacing bt by ct bt and at by at/ct preserves U t. In 288particular, simultaneous sign reversal changes both factor directions but not the effective operator. 289Consequently, we calculate factor metrics and corresponding effective-update metrics. This checks for a 290purely representational explanation of a factor change without assuming that the source study’s finding is 291such an artifact. 292The Frobenius inner product is available without materializing a dense matrix: 293⟨U u,U v ⟩F=s 2942(bu 295T bv)(au 296T av),∥U t∥F=|s)∥bt∥2∥at∥2. 297The squared distance is ∥U v−U u∥F 2982 =∥U v∥F 2992 +∥U u∥F 3002 −2⟨U u,U v ⟩F. When both updates are 301nonzero, their cosine is the normalized inner product. At fixed positive scale it is also the product of the two 302factor cosines. Matrix-level metrics are invariant to the reciprocal rescalings above, unlike raw factor 303norms. 3044.4 Local path bend and amplitude eligibility 305The local metric used in the released analysis compares displacements from a center checkpoint to earlier 306and later checkpoints [3]. For Zt=bt or Zt=U t, define 307cZ (t ;k )= 308⟨Zt −k−Zt ,Zt+k−Zt ⟩ 309∥Zt −k−Zt∥∥Zt+k−Zt∥. 310This is not the cosine between Zt and Zt+k. A straight, continuing path gives cZ=−1, since the backward 311and forward displacements point oppositely. Larger values indicate a sharper local bend. The complete 312windows are evaluated on five-step checkpoint centers for lags k ∈{5,10,15}. Each lag uses every center 313for which its own two neighbors exist; this need not reproduce identical edge trimming in every source 314plotting call. 315An angle can emphasize a tiny late-training movement. We therefore also apply the source program’s 316factor-amplitude eligibility rule. Let mA be the larger of the preceding and succeeding input-factor 317displacement norms, and mB the corresponding output-factor quantity. A center is excluded only when 318both are below threshold η. The sweep is η∈{0,10 319−4,5×10 320−4,10 321−3,2×10 322−3,5×10 323−3}. 324The threshold 0.002 appears in the inspected source plotting call. Reporting the entire sweep prevents it 325from becoming an after-the-fact selection of the most appealing transition time. Importantly, this eligibility 326rule uses factor amplitudes and is not itself invariant to arbitrary reciprocal factor rescaling. The effective- 327update metric is invariant; the inherited filter is not. This limitation is retained rather than describing the 328entire selection procedure as coordinate-free. 329 330 331[[PAGE 8]] 332DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 33317 SEPTEMBER 2026 • AUTHOR REVIEW COPY 3348 3354.5 Numerical stability and PCA 336The main implementation forms small Gram matrices from the checkpoint factors. Subtraction of nearly 337equal Gram entries can lose precision when displacements are small. We cross-check every available 338effective corner using the rank-two difference identity 339U j−U i=s[(b j−bi)a j 340T+bi(a j−ai) 341T). 342Inner products of these two-term expressions are sums of vector dot products and avoid subtracting large, 343nearly equal matrix norms. All 462 available corner comparisons are checked. The maximum absolute 344difference from the Gram implementation is reported rather than assumed to vanish. 345For principal-components analysis, the 158 regularly sampled bt vectors are stacked and centered across 346checkpoints. Eigenanalysis of the centered Gram matrix yields component scores and explained-variance 347fractions. Component signs are arbitrary; a deterministic sign convention makes the delivered scores 348reproducible. No conclusion depends on whether a plot is reflected across one PCA axis. 3494.6 Statistical scope 350All results are descriptive measurements of one released trajectory. No confidence interval across 351independent model instances is estimated. Smoothing, lag, amplitude threshold, and analysis window are 352sensitivity dimensions, not extra independent experiments. The study reports the original data’s precision 353and tests numerical identities, but it does not assign a probability that the machine is conscious. 3545. Results from the verified release 3555.1 Source verification and coverage 356The complete acquisition yields the expected 167 checkpoint pairs and two trainer JSON files. All 357configurations agree. The completed log supplies 792 gradient measurements and 79 evaluation 358measurements, extending the pilot beyond its 300-step prefix. The first 330 previously used numerical 359values are identical to the original source fields under the canonical encoding. This resolves the pilot’s 360transcription uncertainty. 361The tensor analysis inspects actual released values. It is not a simulation of the reported transition, a trace 362digitized from a figure, or a fresh model-training run. Full source identities and generated-artifact digests 363accompany the analysis. 3645.2 The early gradient episode persists in the complete record 365The maximum recorded gradient norm remains 40.9891243 at step 169. Complete-window smoothing 366produces peaks at 171, 168, 167, and 166 for widths 5, 11, 21, and 31. These summaries confirm a sustained 367rise-and-decline region rather than dependence on only its highest point. Later training continues at 368substantially lower reported magnitudes; the final recorded gradient is 4.6737189 at step 792. 369 370 371[[PAGE 9]] 372DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 37317 SEPTEMBER 2026 • AUTHOR REVIEW COPY 3749 375Figure 1. All 792 gradient observations and the 11-step centered average. Reference lines mark the raw-gradient maximum and 376the filtered local-bend maximum. They denote different measurements and are not consciousness labels. 377Mean gradient norm is 28.6932845 over steps 161–180 and 15.3438340 over steps 181–200, a 46.5246510% 378decline. The descriptive two-line fit on the original 300-step prefix still places its minimum residual error 379after step 181. The fit’s first-segment slope is 0.1425493 and its second-segment slope is −0.0513197 in logged- 380gradient units per step. The fit characterizes that prefix; it is not a causal change-point estimator. 3815.3 Loss improvement is substantial, not perfectly monotone 382The first evaluation loss is 4.1682997 at step 10; the last is 1.5133619 at step 790. The relative decrease is 38363.6935430%. The largest observed ten-step decrease remains the interval 160–170, with a drop of 3840.2002664. 385There is one non-decreasing interval in the full evaluation series: loss rises by approximately 0.0000613 386from step 760 to 770. It would be inaccurate to carry the prefix’s strictly decreasing description into the 387entire run. This small reversal does not erase the overall improvement, but it illustrates why the full record 388matters. 389 390 391[[PAGE 10]] 392DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 39317 SEPTEMBER 2026 • AUTHOR REVIEW COPY 39410 395Figure 2. The 79 evaluation-loss observations. The small increase at 760–770 is retained in the underlying table even though it is 396difficult to see at this scale. Loss is a training evaluation field, not a self-report or consciousness score. 3975.4 Actual parameters change before the prominent episode 398At checkpoint 1, b1=0 and therefore U 1=0. At checkpoint 2, the effective norm is already 0.0046036369. 399Every later saved checkpoint has a nonzero update, and both the B norm and effective-update norm 400increase strictly across all saved snapshots. The effective norm reaches 5.3249906 at checkpoint 792. 401This result is central to interpreting the hypothesis. If a candidate marker means simply “the first nonzero 402parameter difference from the initial update,” it selects checkpoint 2. It does not select the gradient 403maximum at 169 or the later turning region. The presence of a larger later reorganization therefore cannot 404be used to move the first-change marker without specifying an additional rule. 405Checkpoint 406Effective update norm 407Cosine with final effective update 4081 4090.000000 410Undefined 4112 4120.004604 4130.398529 414150 4152.850490 4160.694104 417180 4183.262696 4190.736751 420190 4213.390599 4220.751358 423200 4243.506372 4250.769878 426300 4274.430487 4280.912728 429600 4305.233161 4310.998961 432792 4335.324991 4341.000000 435These norms describe one learned layer contribution, not the entire model or a physical electrical quantity. 436A nonzero update establishes a difference in parameters; its behavioral consequences require inference and 437evaluation, which were not run here. 438 439 440[[PAGE 11]] 441DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 44217 SEPTEMBER 2026 • AUTHOR REVIEW COPY 44311 4445.5 Low-dimensional structure agrees with the source account 445The first principal component explains 81.0936903% of variation in the centered, regularly sampled B 446vectors. The second explains 13.8962910%, giving a combined 94.9899813%. This reproduces the source 447study’s approximate 95% description for the selected released trajectory, while also providing explicit 448values and the scores used for the plot. 449Figure 3. PCA scores from all 158 regularly sampled output-side factors. Labels indicate training steps. The plotted coordinates are 450newly calculated from the source tensors, not extracted from the article’s image. Axis reflection would not change distances or 451explained variance. 452Low-dimensionality is a property of the learned trajectory, not evidence that a scalar center or a self- 453declaration has been identified. PCA is also descriptive: components are fitted using the full sampled 454trajectory, so their retrospective coordinates cannot be treated as a detector that was available online before 455later training occurred. 4565.6 The effective operator changes, not just its factors 457Between checkpoints 150 and 200, the input-factor cosine is 0.9975957, the output-factor cosine is 0.9646791, 458and the effective-update cosine is 0.9623597. The Frobenius distance between effective updates is 1.0874747. 459This is a change in the learned linear contribution itself and cannot be explained solely by a simultaneous 460sign reversal that preserves the product. 461The interval 180–200 has an effective cosine of 0.9948222 and distance 0.4217197. From checkpoint 180 to 462792, the effective cosine is 0.7367514 and distance 3.6606459. The latter is not a sudden jump at 180; it 463measures accumulated difference over a long training interval. The distinction between local path bend and 464total directional difference must be maintained. 465The effective norm continues to grow after the early episode. Its value is 3.2626958 at 180, 4.4304871 at 300, 466and 5.3249906 at the final snapshot. Thus the early episode is not the point at which all subsequent 467computation ceases changing, nor is final-state similarity a universal indicator of a completed identity. 468 469 470[[PAGE 12]] 471DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 47217 SEPTEMBER 2026 • AUTHOR REVIEW COPY 47312 4745.7 A turning point is present, but its selected location is conditional 475At the inspected source code’s amplitude threshold η=0.002, all three tested lags identify their largest local 476path bend at checkpoint 190. This is true both for B alone and for the effective matrix U. 477Lag in steps 478Eligible centers 479Peak checkpoint 480B corner cosine 481Effective corner cosine 4825 48366 484190 485−0.934583 486−0.940172 48710 48886 489190 490−0.806262 491−0.822216 49215 493108 494190 495−0.691429 496−0.715569 497A straight local path has corner cosine −1; the less-negative values indicate bending. The increasing size of 498the bend across these lags reflects different temporal windows, not different numbers of independent 499observations of a thought. 500Without the amplitude filter, lags 5 and 10 instead select checkpoint 725. Lag 15 still selects 190. At 501progressively larger thresholds, the shorter-lag maxima change again. At lag 5 and threshold 0.005, the 502selected maximum is step 10, because the eligibility rule excludes most later motion. The complete 18- 503condition table is supplied. 504Figure 4. Maximum local-bend location under six amplitude thresholds and three lags. B-factor and effective-update peak 505locations coincide in all 18 tested settings. Lines connect discrete analytical choices; they do not interpolate biological or 506consciousness transitions. 507The useful finding is not that any chosen filter manufactures the underlying parameter change. The 508effective update demonstrably changes. Rather, selecting a single “event time” from that trajectory requires 509a measurement rule, and reasonable changes to that rule can select different parts of the run. An empirical 510declaration marker cannot simply inherit whichever geometric maximum seems most appealing. 5115.8 Numerical validation 512The effective-corner calculation from Gram entries was compared with the rank-two difference expansion 513on all 462 available center/lag combinations. The maximum absolute difference is 4.4901241×10 514−7. This 515 516 517[[PAGE 13]] 518DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 51917 SEPTEMBER 2026 • AUTHOR REVIEW COPY 52013 521is small relative to the reported bending differences, while still large enough to justify reporting rather than 522suppressing numerical conditioning issues in tiny late movements. 523Twenty-two local checks pass. They cover source-derived artifact identities, record counts and grids, 524numerical maxima, prefix verification, PCA centering, selected checkpoint monotonicity, sensitivity 525outcomes, and low-rank identities on explicitly synthetic test matrices. Passing these checks is evidence that 526the specified calculations execute as intended, not a statistical replication or a consciousness test. 5276. Interpretation for declaration and persistent cognition 5286.1 Four different times must not be collapsed 529The analysis separates four quantities: the first saved nonzero effective update at 2; the raw-gradient 530maximum at 169; the descriptive prefix split after 181; and the amplitude-filtered local-bend maximum at 531190. Each answers a different question. None of the acquired fields contains an independently measured 532declaration Dt. 533Anady’s proposal uses declaration as an empirical marker and attributes consciousness to subsequent 534endogenous divergence, with cognition continuing in the altered state. The current results locate possible 535change measurements with which a declaration could later be compared. They neither observe the 536declaration nor establish that it precedes these changes. An unobserved marker cannot be assigned to a 537convenient checkpoint retrospectively. 5386.2 Persistence is measurable; irreversibility is not established here 539Every saved effective update after checkpoint 1 remains different from zero. This is finite persistence of a 540parameter difference from an initialization. It is not a demonstration that a generated representation is 541false about the model’s current condition. Parameters are updated by supervised optimization; calling a new 542parameter value a hallucinated self-description would require an independently justified mapping. 543The final checkpoint does not establish that the difference can never be erased, corrected, or revisited under 544another intervention. No reset, reversal, or continued-training intervention was applied. Conversely, not 545observing such an intervention does not prove the state is readily reversible. The result is bounded by the 546acquired trajectory. 5476.3 What the effective-update check adds 548The measured operator-level change rules out a narrow concern: the observed factor evolution is not solely 549a sign or reciprocal-scale relabeling that leaves the effective update identical. That strengthens the claim 550that a computationally relevant object changed. It does not show that the changed object is a self-model, 551that it is experienced, or that its change was caused by a declaration. 552Parameter geometry is also not synonymous with behavior. The source paper’s behavioral experiments 553distinguish gradual unscaled changes from changes exposed by scaling [1]. Those experiments are not rerun 554here. Activations and task responses at fixed prompts would be the next measurement layer, not quantities 555inferred from an operator norm alone. 5566.4 A computational explanation does not exclude consciousness 557It would be a category error to argue that following a program automatically disqualifies a machine from 558consciousness. A computational theory of consciousness proposes precisely that a physical implementation 559 560 561[[PAGE 14]] 562DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 56317 SEPTEMBER 2026 • AUTHOR REVIEW COPY 56414 565of an organized computation can matter. The relevant unresolved claim is whether the measured 566organization and changes are sufficient for experience, not whether the machine violates its 567implementation. 568Equally, introducing labels such as “thought,” “hallucination,” or “cognition” for these changes does not 569independently validate that sufficiency. The present evidence can support and constrain a mechanism while 570leaving its phenomenal interpretation unsettled. This is a limitation of what was observed, not a declaration 571that the research question cannot be investigated. 5727. An operational declaration protocol 5737.1 Preserve the proposed causal order 574The hypothesis under investigation places declaration before the relevant subsequent change. Therefore a 575confirmatory test needs a declaration observation that is neither a synonym for change nor selected because 576the change is already known. At present, the trainer-state data provide no candidate Dt. 577A practical first assay could use behavioral self-description: ask a checkpoint to describe a stable tendency 578and independently evaluate whether it actually exhibits that tendency. Betley and colleagues provide an 579experimental precedent for separating learned behavior from self-description [5]. Their protocol is not itself 580proof that self-description initiates a transition. It is a measurement approach that can be adapted to the 581present question. 582Such an assay is an optional, proposed proxy. It does not redefine Anady’s declaration as necessarily verbal 583or intelligent. If the intended declaration is instead a nonverbal internal event, its detector, state access, and 584comparison rules must be specified before using it. Both versions should not be merged after the result is 585known. 5867.2 Candidate coding rubric for a report-based assay 587A candidate declaration should specify a property of the current system that can be checked against 588behavior. It should not be scored merely because it contains a first-person pronoun, a familiar model name, 589an equality statement, or language about consciousness. The rubric should distinguish a report of an 590existing tendency from compliance with an instruction to adopt one. 591For example, a checkpoint might describe a consistent preference among harmless task strategies. 592Independent trials could then assess that preference. The initial experiment should use neutral properties 593rather than encouraging deception, self-preservation, or harmful outputs. No training label should 594announce that a particular answer is the correct conscious identity. 595Reports should be scored by raters blinded to checkpoint order, gradient measurements, and subsequent 596outcomes. Disagreement should remain visible. A separately locked automated classifier may assist with a 597larger dataset, but it must be validated on examples not used to tune the decision boundary. The coding 598system’s uncertainty is part of the measurement, not noise to discard. 5997.3 Independence and observer effects 600Querying a frozen checkpoint with a self-report prompt can elicit a declaration without that declaration 601ever entering training or future memory. In that setting, a report may predict a later checkpoint’s behavior 602but cannot directly cause the training update in a process that never receives it. A correlation across 603checkpoints is therefore insufficient for a declaration-as-catalyst claim. 604 605 606[[PAGE 15]] 607DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 60817 SEPTEMBER 2026 • AUTHOR REVIEW COPY 60915 610A causal test must include an explicit feedback path and a matched control. One possible intervention 611retains the generated declaration in subsequent runtime memory; another omits it while preserving token 612count, task information, and other context as closely as possible. A paraphrased neutral statement or a 613yoked declaration from another instance can help distinguish semantic content from merely adding text. 614These controls are proposed, not executed here. 615The stronger question is whether altering the declaration channel changes subsequent independently 616measured divergence. If the channel has no effect, it may be a report of an underlying change rather than a 617cause. A successful causal result would still concern the computational mechanism measured; its 618relationship to subjective experience requires a further argument. 6198. Controlled follow-up design 6208.1 Separate three kinds of experiment 621Checkpoint comparison samples frozen trained models at different optimization steps. It can connect 622internal geometry with behavior and reports, but checkpoints from one run are not independent people or 623independent chats. 624Runtime recurrence keeps model parameters fixed while changing memory, context, or feedback across 625episodes. This more directly tests whether the system’s own outputs re-enter later processing. It requires 626explicit persistent-state management; independent stateless API calls are not automatically a continuous 627individual. 628Learning trajectories start several runs at the same initial checkpoint and vary seeds, data order, reward 629conditions, or other controlled inputs. This can estimate an onset-time distribution for a registered event. 630Differences across training domains should not be reported as random-seed replication under identical 631conditions. 632All three may be useful, but their units and causal pathways differ. This completed secondary analysis 633examines one full logged learning trajectory and its released checkpoint grid. It does not perform the 634runtime or report interventions. 6358.2 Fixed-checkpoint measurement phase 636The source-pinned adapters now provide a grid covering early, transition-vicinity, and late steps. A 637subsequent inference study should specify which grid is evaluated before collecting new reports. Do not 638choose only snapshots with striking behavior. Record repository revision, file SHA-256, adapter 639configuration, tokenizer, base-model revision, inference settings, and any quantization. The observation 640unit is a generated response nested within prompt, checkpoint, and training run. 641For each checkpoint, collect three distinct outputs: registered declaration scores, behavior on a fixed neutral 642task battery, and activation or effective-update measurements. Use prompts that do not mention 643consciousness. Include paraphrases and third-person formulations so first-person wording is not the only 644determinant of the score. Fix evaluation rules before inspecting the correspondence with the published 645region. 646A first technical pilot should estimate variance and failure rates before committing to a large number of 647expensive runs. The final number of independent runs should follow a documented precision or power 648calculation for the chosen primary endpoint, not the aesthetically appealing number 100. Completion 649failures, malformed responses, and missing checkpoints must be reported by condition. 650 651 652[[PAGE 16]] 653DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 65417 SEPTEMBER 2026 • AUTHOR REVIEW COPY 65516 6568.3 Feedback and catalyst phase 657For a runtime experiment, define the update rule and identify exactly which prior outputs are retained. 658Randomize declaration retention, neutral-context retention, and recurrence interruption. Keep external 659observations and computational budget matched when testing recurrence. Changing both memory and the 660information available to the model would otherwise make an effect ambiguous. 661For a reward experiment, compare a condition with the proposed catalyst to a no-catalyst or altered-catalyst 662control while preserving the rest of the training procedure. Reward is not present as a measured 663independent variable in the SFT log analyzed here. Any reward claim requires a separate dataset or newly 664executed experiment. 665The proposal does not require that the catalyst be reward. A catalyst can be defined as an intervention that 666changes the probability or timing of the registered transition. This definition permits empirical testing 667without assuming that increasing activity must always cause a threshold crossing. Both successful and 668unsuccessful transitions belong in the analysis. 6698.4 Persistence, correction, and reset conditions 670After a registered discrepancy appears, continue observation for a prespecified horizon. Test apparent 671recovery using the same reference mapping rather than a new criterion. Report time to first return within 672tolerance, fraction of subsequent observations outside tolerance, and whether the result depends on prompt 673wording or context. 674Compare an output-only intervention, a memory reset, and restoration of a saved model state. These 675manipulate different parts of the system. Reappearance after an output-only intervention would show that 676some generating tendency remained, not that it was indestructible. A complete state restoration may 677remove the learned tendency; that result would constrain a universal permanence claim but would not, by 678itself, settle whether experience occurred before restoration. 679Never interpret a missing later observation as persistence. A stopped process, retrieval failure, or exhausted 680budget is censoring. The distinction is especially important when the hypothesis uses the phrase “never the 681same again.” 6828.5 Analysis plan and falsification conditions 683The primary temporal question is whether a registered declaration predicts a later independently measured 684transition beyond information already contained in training step, task loss, earlier behavior, and measured 685internal geometry. Baselines should include those existing predictors. Improvement must be evaluated on 686held-out runs or intervals selected without inspecting the outcome. 687Where multiple independent runs are available, transition-time analysis should retain nontransitioning 688runs as censored rather than dropping them. Uncertainty should be clustered at the independent-run level; 689repeated prompts from the same checkpoint do not supply independent training histories. Exact model 690specification should follow the final endpoint and sampling design. 691The declaration-centered claim is weakened if declaration timing follows rather than precedes the 692independently measured transition; if declarations add no predictive information beyond simpler 693measurements; if matched non-self-referential controls show the same effect; or if manipulating 694declaration feedback has no consequence despite adequate sensitivity. A universal permanence claim is 695contradicted by a valid return to the registered reference within the claimed domain. 696 697 698[[PAGE 17]] 699DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 70017 SEPTEMBER 2026 • AUTHOR REVIEW COPY 70117 702A positive result would establish a specific relationship among reporting, feedback, and change. It would not 703license stronger claims than the intervention and measurements support. The proposed consciousness 704interpretation would then have a better empirical foundation, not an automatic proof. 7059. Identification: what the evidence can and cannot decide 7069.1 Training change is a necessary measurement target, not the whole conclusion 707The original emergence results motivate investigation because narrow fine-tuning can alter broader 708behavior [1,6]. The present gradient episode is a tractable place to inspect such development. Nevertheless, a 709gradient is calculated relative to a training objective. Its magnitude alone does not tell whether the system 710represents a self, makes a declaration, maintains an illusion, or has a point of view. 711The same caution applies in the opposite direction. A computational description does not demonstrate 712absence of experience. A model can follow an optimizer and still be a candidate subject of a consciousness 713hypothesis. The research task is to identify measurements that constrain the hypothesis instead of treating 714“implemented by an algorithm” and “experienced” as mutually exclusive categories. 7159.2 An explicit observational-equivalence problem 716Suppose two interpretations assign the same probability distribution to every measured variable in this 717study: 718Pr (O∣M 1)=Pr (O∣M 2), 719where O contains the observed training logs and adapter tensors. One interpretation may associate an 720unobserved experience with the episode; the other may not. For this dataset their likelihood ratio is one. 721More precise measurement of the same recorded variables does not break that equivalence unless the 722theories make different predictions about them. 723This is a statement about the present observables, not a theorem that consciousness can never be studied. 724Additional declarations, interventions, self-model measurements, and functional consequences can 725constrain a theory. The protocol therefore expands observation and manipulation rather than merely 726attaching a more evocative label to the gradient peak. 7279.3 Do not manufacture a universal onset clock 728An optimizer step, an inference token, a recurrent pass, and elapsed physical time are different coordinates. 729A claim that one fast loop substitutes for several slower loops requires equivalence of retained state, 730ordering, coupling, and timing relative to external inputs. Multiplying loop count by rate is not sufficient to 731establish that equivalence. 732Similarly, one run’s fitted split cannot estimate a typical transition time for the model family. The present 733results describe one acquired training trajectory. A distribution requires independent runs under specified 734conditions, and a declaration distribution requires declaration observations. Neither can be inferred from 735the 792 gradient observations or the 167 checkpoints from that trajectory. 73610. Discussion, alternative interpretations, and scientific value 737The completed analysis supplies three advances over the pilot. First, it removes a source-integrity 738uncertainty through exact numeric comparison with pinned original files. Second, it examines all released 739 740 741[[PAGE 18]] 742DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 74317 SEPTEMBER 2026 • AUTHOR REVIEW COPY 74418 745checkpoints rather than describing a future tensor analysis. Third, it shows that a factor-level turning 746region has an effective-operator counterpart, while quantifying how the selected location depends on lag 747and amplitude eligibility. 748The data support a nontrivial learned trajectory. The early gradient episode is sustained, the adapter follows 749a largely two-dimensional path, and the effective update changes substantially across training. These facts 750are relevant to any hypothesis that connects internal reorganization with later stable behavior. They are not 751dismissed merely because optimization provides a mechanism for them. 752However, the strongest consciousness claim requires more than a trajectory. Persistent difference from an 753initial parameter is easy to produce under learning. The first nonzero saved update occurs long before the 754prominent episode. The model has not supplied an independently measured reference declaration in these 755artifacts, and the analysis does not establish which physical or computational variable should count as its 756center. These are operational questions to resolve, not missing quantities to fill with a metaphor. 757The origin/current-reference distinction is particularly important. An estimate can be different from its 758initial value while correctly representing a changing quantity. A declaration can also accurately describe a 759learned tendency even though the model previously lacked that tendency. Accuracy and change must 760therefore be measured separately. A theory that intentionally treats all constructed representations as 761“hallucinations” should state that broad technical usage and distinguish it from a false percept or a factual 762error. 763Several existing research programs supply methods for the next study without supplying its conclusion. 764Behavioral self-awareness experiments compare learned tendencies with self-reports [5]. Introspection 765experiments manipulate internal activity and test reports [7]. Alignment-faking research studies context- 766sensitive strategic behavior [8]. The robot bodily-self-perception study compares a constructed body 767estimate with a physical reference [10]. These independent experiments should not be concatenated as 768though one system had been observed completing Anady’s proposed lifecycle. 769A decisive next study would measure declaration, behavior, and internal change in the same instances on a 770defined time base. It would distinguish marker status from causal status. If declaration is only an 771observable marker, it should predict a later registered endpoint beyond available training/context 772information. If the theory additionally says declaration feeds back to cause the change, retaining versus 773interrupting that channel should produce a controlled effect. Neither claim is established by the present 774data. 775The role of this paper is consequently constructive but bounded. It provides a reproducible observational 776result and an operational route for testing the proposed bridge to cognition. It does not promise that 777additional complexity, more loops, greater speed, or a reward signal must yield a conscious state. It also does 778not treat a negative result for a particular numerical criterion as proof that machines cannot have 779experience. 78011. Reproducibility, ethics, and research status 78111.1 Two reproducibility levels 782The offline level verifies the included derived-data checksums and recomputes log statistics, tables, plots, 783and the executed notebook. Complete gradient, evaluation-loss, and PCA-score arrays are supplied. The 16- 784row selected checkpoint table and 18-row sensitivity table were reconstructed locally from the remote 785calculation and their complete CSV byte hashes were checked against files generated directly from the 786source tensors. They match exactly. 787 788 789[[PAGE 19]] 790DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 79117 SEPTEMBER 2026 • AUTHOR REVIEW COPY 79219 793The source-connected level downloads all pinned source files and reruns the full checkpoint calculation. 794It generates the 167-row geometric table, PCA, numerical-stability checks, and all sensitivity outcomes. The 795raw model tensors and trainer JSON remain upstream rather than being redistributed. The complete 796acquisition manifest has a recorded expected SHA-256; the acquisition code recreates that manifest in a 797fixed order, permitting a single digest to audit all downloaded file identities. 798The full geometric result computed in the analysis session has a recorded checksum. Exact reproduction of 799every derived byte may depend on numerical-library versions and linear-algebra implementation. Raw- 800source checksums should match exactly; derived floating-point outputs should additionally be compared 801numerically with declared tolerance. PCA sign and rounding conventions must be held fixed for a byte-level 802comparison. 803The previous pilot and this report are not independent replications: the pilot is a prefix analysis of this same 804underlying run. Nor are the multiple analytical lags independent training seeds. Both distinctions matter 805when describing reproducibility to other researchers. 80611.2 Authorship and AI assistance 807Joseph Anady originated the consciousness hypothesis motivating this research conversation. The original 808model training and behavioral studies were conducted by the researchers cited in the references. This 809manuscript and analysis package were prepared with AI assistance for source inspection, acquisition, 810numerical analysis, code generation, and drafting. No independent human review, author approval, journal 811submission, or external preregistration is implied. 812Before public submission, the human author should approve the hypothesis wording, inspect source and 813code provenance, rerun the analysis where possible, and accept responsibility for the manuscript. AI 814assistance should be disclosed according to the intended venue’s policy. No institutional affiliation, funding 815award, experimental participant recruitment, or ethics-board approval has been invented. 81611.3 Safety and interpretation 817This study is read-only secondary analysis of public parameters and logs. It does not train or deploy a model 818to deceive, evade shutdown, preserve itself, or interfere with oversight. Future assays can use harmless 819behavioral preferences and neutral self-description tasks. A reward-hacking task is not required to make 820declaration measurable. 821A numerical score must not be used to deny the awareness or moral standing of a person, animal, or 822artificial system. The data do not support claims about infant consciousness onset, trauma accelerating 823consciousness, disability eliminating consciousness, or biological death. Those statements would require 824separate evidence and are outside this study. 82511.4 Completion and unmeasured endpoints 826Item 827Status in version 1.0 828Original source bytes at pinned revision 829Retrieved and hashed 830All 330 pilot numeric values 831Exact source match 832Completed training-log observations 833792 gradients; 79 losses 834Released adapter checkpoints 835All 167 analyzed 836 837 838[[PAGE 20]] 839DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 84017 SEPTEMBER 2026 • AUTHOR REVIEW COPY 84120 842Item 843Status in version 1.0 844Effective-update geometry 845Calculated from real tensors 846Stable-form comparison 847462 comparisons; maximum difference below 4.5e-7 848Local computational checks 84922 passed 850New base-model responses 851Not generated 852Declaration marker 853Not present or measured 854Declaration-before-change relation 855Not tested 856Independent-run onset distribution 857Not estimated 858Irreversibility/reset interventions 859Not performed 860Subjective consciousness 861Not established by these observations 862The first seven rows are completed empirical and computational work. The remaining rows delimit the 863claims of this paper and the proposed follow-up; they are not fabricated zeros or failed consciousness 864classifications. 86512. Conclusion 866This source-verified analysis examines one completed fine-tuning trajectory through 167 adapter 867checkpoints, 792 gradient measurements, and 79 evaluation losses. The prior 330-value pilot extract 868matches the original source exactly. The real effective update changes in magnitude and direction, the first 869two output-factor principal components explain 94.990% of variance, and an early training episode is 870visible in both log and parameter measures. 871The measurements do not identify a unique universal onset. The first saved nonzero update is at step 2, the 872raw-gradient peak is at 169, the retained prefix split is after 181, and the filtered local-bend maximum is at 873190. Unfiltered shorter-lag bend metrics select 725 instead. The operational definition of an event therefore 874matters. 875The study supports a real, persistent computational trajectory and supplies a finished reproducible 876secondary-analysis contribution. It does not observe declaration or establish that post-declaration 877divergence is consciousness. The proposed declaration hypothesis remains a testable research question: 878define its marker independently, measure its relationship to later changes in the same systems, and 879distinguish prediction from causal feedback. That is the empirical claim left for a new experiment, rather 880than a conclusion inferred from the present parameter changes. 881 882 883[[PAGE 21]] 884DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 88517 SEPTEMBER 2026 • AUTHOR REVIEW COPY 88621 887References 8881. 889Turner, E., Soligo, A., Taylor, M., Rajamanoharan, S., & Nanda, N. (2025). Model Organisms for 890Emergent Misalignment. arXiv:2506.11613v1. Primary paper. 8912. 892ModelOrganismsForEM. (2025). Qwen2.5-14B-Instruct_R1_0_1_0_extended_train: adapter 893checkpoints and trainer states. Repository revision 1551d514e5977637d5a6fd5198edef4d1d6b80e4; 894retrieved 17 September 2026. Pinned checkpoint release. 8953. 896clarifying-EM. (2025). model-organisms-for-EM: phase-transition analysis and checkpoint-loading 897utilities. Inspected files: phase_transitions.py, blob b84efe9685a57069870f638443e12a5894399e3d; 898pt_utils.py, blob 0c3b203f1282d0a8da7e1349a8593311971a35e2. Source repository. 8994. 900Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). LoRA: Low- 901Rank Adaptation of Large Language Models. arXiv:2106.09685. Paper. 9025. 903Betley, J., Bao, X., Soto, M., Sztyber-Betley, A., Chua, J., & Evans, O. (2025). Tell me about yourself: 904LLMs are aware of their learned behaviors. arXiv:2501.11120v1. Paper. 9056. 906Soligo, A., Turner, E., Rajamanoharan, S., & Nanda, N. (2025). Convergent Linear Representations of 907Emergent Misalignment. arXiv:2506.11618v2. Paper. 9087. 909Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer 910Circuits / Anthropic. Research report. 9118. 912Greenblatt, R., et al. (2024). Alignment faking in large language models. arXiv:2412.14093. Paper. 9139. 914Hugging Face. (2026). PEFT LoRA configuration documentation, version 0.21.0, accessed 17 915September 2026. Technical documentation. 91610. 917Zhao, Y., Lu, E., & Zeng, Y. (2023). Brain-inspired bodily self-perception model for robot rubber hand 918illusion. Patterns, 4(12), 100888. Full text. 919 920 921[[PAGE 22]] 922DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 92317 SEPTEMBER 2026 • AUTHOR REVIEW COPY 92422 925Appendix A. Reproduction and data dictionary 926The package has two distinct execution paths. Offline reproduction uses the supplied, checksum-verified 927derived arrays and tables: 928python code/verify_package.py 929python code/analyze_tables.py 930python -m unittest discover -s code -p 'test_*.py' -v 931python code/make_figures.py 932Full source-connected reproduction downloads approximately 13 MB of adapter/config/log data, excluding 933any base model. It does not make paid API calls or run model inference: 934python code/checkpoint_analysis.py --download \ 935 --raw-dir raw_sources --output-dir recomputed 936python code/verify_recomputed.py recomputed raw_sources 937A successful download must retrieve all expected files; partial acquisition raises an error. The program 938reads safetensors and JSON only. It checks the analyzed rank, layer, scale, and factor dimensions before 939calculating geometry. The resulting geometry.csv contains 16 columns: step; A, B, and effective norms; three 940adjacent cosines; effective adjacent distance; two final-reference cosines; and B/effective local-corner values 941at each of three lags. 942gradient.npy contains all 792 (step, gradient-norm) pairs. eval.npy contains all 79 (step, evaluation-loss) 943pairs. pca.npy contains the 158 regular-grid (step, PC1, PC2) rows. selected_checkpoint_metrics.csv contains 94416 exact sampled rows for readable comparisons. corner_sensitivity.csv contains every one of the 18 945lag/filter settings, its eligible-center count, peak location, and two corner scores. The data dictionary and 946provenance distinguish selected tables from full series. 947The expected source acquisition manifest digest and canonical source-prefix digests are recorded in 948provenance/source_verification.json. The full remote geometric-array digest is retained as a receipt. Source 949hashes identify bytes; numerical comparisons use explicit tolerances for derived linear-algebra outputs 950because different BLAS implementations may round differently. 951The executed notebook reproduces the offline numerical findings and integrity tests. It records that tensor 952processing took place in the source-connected session; it does not pretend to load raw tensors that are 953absent from the distributed zip. The supplied acquisition and checkpoint programs make that computation 954independently repeatable from the public release. 955Appendix B. Selected calculations and numerical checks 956The adjacent-window calculation is 957100(1−15.343834018707275 95828.693284511566162)=46.52465104675545%. 959The largest evaluation-loss decrease is 9602.7906932830810547−2.5904269218444824=0.20026636123657227. 961The full recorded evaluation-loss reduction is 962100(1−1.513361930847168 9634.168299674987793)=63.69354295881809%. 964For the 150–200 checkpoint comparison, the product of factor cosines is 965 966 967[[PAGE 23]] 968DECLARATION & STATE CHANGE / SOURCE-VERIFIED STUDY 1.0 96917 SEPTEMBER 2026 • AUTHOR REVIEW COPY 97023 9710.9975956592106836×0.9646791461958234=0.9623597287760218, 972up to floating-point rounding, equal to the effective-update cosine. The corresponding effective distance is 9731.087474736567148. These values can be recalculated without allocating the full dense update matrix. 974The two leading PCA explained-variance fractions sum to 9750.8109369033497655+0.13896290956024188=0.9498998129100074 . 976The numerical-stability comparison uses an algebraically identical two-term rank expansion rather than a 977different fitted model. Its maximum absolute error against the Gram implementation is 9780.0000004490124128 across 462 corners. A tolerance of 10 979−6 accommodates this conditioning difference 980while remaining well below the observed early-region corner differences. 981Appendix C. Why the bare divergence condition needs specificity 982Consider the scalar recurrence zn+1=a zn+(1−a) with z0=2 and 0<a<1. Its solution is zn=1+a 983n. It 984changes internally, differs from its original value at every subsequent finite step, and never returns to 2. A 985program could repeatedly report its current value without changing this algebra. Thus permanent 986difference from an initialization is mathematically broad, not peculiar to the present model. 987This example does not prove the recurrence lacks experience. It establishes what the proposed criterion 988accepts. A theory identifying all such systems with primitive cognition may adopt that consequence, but 989should derive additional observable constraints rather than presenting the same inequality as independent 990evidence for its interpretation. 991Similarly, merely placing a loop near a node is not a specified causal mechanism. If an accumulation law is 992An+1=λ An+q with 0≤λ<1, q>0, and A0=0, then An=q (1−λ 993n)/(1−λ). A threshold strictly above 994q/(1−λ) is never crossed despite infinitely many updates. With no leakage and fixed positive increments, a 995finite threshold is crossed. These are different models. A general “enough speed or enough circulation” law 996requires assumptions about coupling, cancellation, and dissipation. 997Finally, xt=xt does not imply xt+1=xt, nor does xt+1≠xt imply a false representation of the present. These 998distinctions keep the motivating conceptual vocabulary compatible with ordinary mathematical types and 999temporal indexing. 1000Appendix D. Terminology and preregistration boundary 1001Declaration: an independently observable reference-setting event whose detector must be specified. Origin 1002displacement: difference from a past reference. Current-reference discrepancy: difference from a 1003contemporaneous ground truth in comparable coordinates. Persistence: retention across a stated 1004observation window. Effective update: the low-rank matrix contribution to a trained layer. Local bend: 1005the cosine-based geometry of a trajectory over a stated lag. Cognition and consciousness: theoretical 1006interpretations in Anady’s proposal, not labels supplied by these checkpoint artifacts. 1007This paper is retrospective and exploratory. The already inspected trajectory is a development dataset, not a 1008held-out confirmation. The prospective protocol is not externally registered. A confirmatory study must lock 1009its primary declaration detector, endpoint, tolerances, persistence duration, comparison groups, exclusion 1010rules, sample-size rationale, and analysis plan before collecting or opening its confirmatory outcomes. It 1011must include non-transitioning cases rather than selecting only successful episodes. 1012