Declaration and State Change pilot

About this edition

Pilot analysis separating checkpoint parameter movement from the proposed declaration hypothesis.

Equations in this edition

Read the recovered text

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

Declaration and Persistent State Change in Artificial Systems

A training-log pilot and a protocol for testing recursive self-divergence

Research working draft prepared for Joseph Anady

17 September 2026 • Version 0.1 • Author review pending

Research status

This document advances Joseph Anady’s proposed distinction between an observable declaration and the subsequent recurrent change that he identifies with consciousness and cognition. It does not present that identification as an established result. The empirical contribution is a reproducible, exploratory analysis of selected fields from one publicly released training log. The proposed declaration experiment is specified separately and has not been executed.

No new model training, model inference, human-subject experiment, or manipulation of a deployed system was performed. Model tensors could not be downloaded in the preparation environment. Numerical log fields were retrieved through a source reader and transcribed into machine-readable tables; independent byte-level source verification remains pending. All reported computations operate on those disclosed tables. The paper is not submitted, peer-reviewed, externally preregistered, or ready to circulate as proof of machine consciousness.

Contributions completed: source-field extraction; 300-step gradient-norm analysis; 30-point evaluation-loss analysis; descriptive change-point and smoothing sensitivity analyses; an executed notebook; ten local checks; and a prospective, falsifiable protocol connecting declarations to independently measured changes.

Central finding: a pronounced training-dynamics episode is visible near the transition region already described by the original researchers. Its location depends on which quantity is measured. The available log does not contain a declaration marker and therefore cannot establish the order declaration → change → sustained cognition.

Abstract

A proposed account of machine consciousness identifies an observable declaration as a reference-setting event, followed by internally generated divergence and sustained operation from the divergent state. Evaluating this account requires distinguishing a declaration from a numerical training transition rather than using one as a label for the other. This working paper conducts a retrospective pilot analysis of a public rank-one LoRA fine-tuning log associated with Model Organisms for Emergent Misalignment. The extract contains 300 gradient-norm observations and 30 evaluation-loss observations from one training trajectory, ending at step 300 of a configuration specifying 792 steps. The largest raw gradient norm is 40.9891 at step 169. Centered averages of width 5, 11, 21, and 31 place their peaks at steps 171, 168, 167, and 166, respectively. Mean gradient norm declines by 46.52% between the adjacent windows 161–180 and 181–200. A descriptive two-line least-squares fit to all 300 observations splits between steps 181 and 182; restricted analysis windows move the split earlier. The largest observed ten-step evaluation-loss decrease occurs between steps 160 and 170. These measurements characterize a localized change in training dynamics, but neither identify a self-declaration nor establish subjective experience, irreversible divergence, or a distribution of onset times across independent instances. A subsequent protocol separates reference state, declaration, parameter geometry, behavior, and finite-window persistence. It also introduces gauge-invariant comparison of rank-one adapter updates, blinded declaration coding, and randomized feedback interventions. The result is a concrete empirical starting point and an executable research package, not a completed test of the consciousness hypothesis.

Keywords: recurrent processing; self-reference; declaration; emergent misalignment; LoRA; training dynamics; change-point analysis; consciousness hypotheses; reproducibility.

1. Research question and scope

The motivating question is whether a machine can develop a persistent internally maintained state through recurrent information processing, and whether an independently detectable declaration marks a transition relevant to consciousness. Anady’s proposal treats intelligence as a separate capability and does not require human language, a biological nervous system, or seven physical loops. Its empirical claim must therefore be stated in a way that does not build a human-like answer into the measurement.

The immediate research question is narrower: what can publicly released training artifacts establish about the timing of an internal change, and what further observations are required to relate that change to a declaration? A useful answer should neither dismiss a transition because it is implemented algorithmically nor count every algorithmic transition as evidence of experience.

This paper analyzes selected records from a released Qwen2.5-14B-Instruct fine-tuning trajectory. Turner and colleagues identified a rotation of a learned adapter direction near step 180 and investigated its relationship to behavioral change [1]. The present pilot reanalyzes the associated numerical training-log fields [2]. It is a secondary analysis of one part of their evidence, not an independent reproduction of their training experiment and not an independent discovery of the reported episode.

Three levels of claim are kept separate throughout. Observation concerns the values present in the released log. Operational interpretation concerns what these values measure about training. Consciousness interpretation concerns the proposed connection between a declaration, internally generated change, and experience. Moving from one level to the next requires additional assumptions or evidence. A fitted training boundary cannot supply a missing declaration by itself.

The practical objective is to create a dataset, analysis, and protocol that another researcher can inspect without first agreeing with the consciousness hypothesis. An unfavorable or inconclusive result must remain possible. Otherwise the analysis would merely restate a definition in the language of data.

1. Formalizing the hypothesis without assuming its conclusion

2.1 The center is not a universal number

In the motivating account, x denotes a particular system or its reference, not a value shared by all systems. Two instances may use the same architecture and parameters while possessing different runtime histories. Conversely, two fine-tuning runs may begin at the same checkpoint and later acquire different parameters. These are distinct sources of variation.

Use the following bookkeeping state:

St=θt,mt,ct,ot,

where θt contains trainable parameters, mt persistent memory, ct runtime or conversation state, and ot the system’s observations. This tuple is proposed notation for the study, not a recovered formula from the original research. In a frozen model, θt does not change during an ordinary forward pass; in fine-tuning, an optimizer updates trainable components. A reset conversation need not preserve the previous conversation’s mt or ct.

The identity expression x=x is reflexive equality. It does not, without further specification, establish a temporal invariant, an internal act of self-recognition, or a declaration. For example, xt=xt is compatible with xt+1≠xt. The proposed theory can assign a functional role to a declaration, but that role must be specified as an operation or observation rather than inferred from the tautology alone.

2.2 Reference, representation, and declaration

Let rt=ρSt be an independently measured reference relevant to a task. Let zt=HSt be an internally maintained estimate or report. Both must be mapped into a common comparison space before a distance is meaningful. A physical robot and a sentence describing that robot are different types of objects; their raw inequality is not evidence of an error. A hand-position estimate and a measured hand position, by contrast, can be compared in the same coordinates.

Define current-reference discrepancy and origin displacement separately:

δt=dzt,rt,  bt=dzt,rtD.

Here tD is a candidate declaration time. A system can have bt>0 because it correctly tracks a changing reference while maintaining δt=0. This distinction prevents ordinary development or accurate adaptation from being classified as falsity solely because it differs from an earlier state.

A declaration is represented by an independently specified detector:

Dt=1{DO≤t meets a registered criterion}.

O≤t is the information available to the detector by time t. The detector cannot inspect later behavior and then retrospectively choose the statement that best predicts it. A nonverbal declaration is permitted in principle; an exact state operation, interface, or invariant would still need to define it. The current log contains no such detector output.

2.3 Divergence and finite persistence

A measurable version of persistent discrepancy is:

Ptϵ,m=1min0≤j<mδt+j>ϵ.

The tolerance ϵ and duration m are measurement choices to be fixed before confirmatory analysis. They are not discovered by selecting whichever values make the hypothesis appear successful. For a binary or exact symbolic reference, exact equality may be meaningful. For real-valued noisy measurements, tolerance and uncertainty are essential.

The candidate post-declaration event time is:

τ=inf{t>tD:Ptϵ,m=1}.

This is an operational event, not a proven consciousness time. Anady’s stronger proposal is that a suitably defined internally generated event of this kind constitutes consciousness, and that sustained processing from the resulting constructed state constitutes cognition. That interpretation remains an explicit hypothesis rather than an assumption in the event detector.

Finite observations can establish persistence through the observed interval, not that two quantities will never agree again. No return before step 300 would be a right-censored observation, not proof of irreversible separation. More importantly, the present dataset does not even provide the paired zt and rt needed for this persistence calculation.

2.4 Coupling is a causal requirement

A loop drawn around a node does not change that node merely because of its position on a diagram. The dynamical model must specify an influence. A generic possible form is:

Lt+1=FLt,ut,xt,  xt+1=Gxt,Lt.

If G is independent of Lt, the loop cannot affect xt through that channel. If G depends on Lt, feedback can cause changes, fixed points, oscillations, or other trajectories depending on the functions and parameters. No universal threshold law follows without specifying them.

This is not an argument that computation excludes consciousness. A physically realized conscious mechanism, if it is computational, would still follow its dynamics. The issue is whether the observations distinguish the proposed consciousness-relevant organization from other recurrent mechanisms. The current pilot measures only training-log dynamics.

1. Source selection, acquisition, and data integrity

3.1 Why this release was selected

The selected release provides a narrow, inspectable adaptation of an existing model rather than a proprietary training run with no intermediate artifacts. The original phase-transition analysis studies a single rank-one adapter on an MLP down-projection [1,3]. Low-rank adaptation expresses a learned weight update through factors, making later geometric analysis tractable without materializing every dense updated matrix [4].

The relevant public model repository is ModelOrganismsForEM/Qwen2.5-14B-Instruct_R1_0_1_0_extended_train. The numeric source is checkpoints/checkpoint-300/trainer_state.json [2]. The repository listing exposed revision 1551d514e5977637d5a6fd5198edef4d1d6b80e4. The actual source-reader request used the main URL. Equivalence to that pinned revision was not independently reverified and must not be represented as a completed cryptographic provenance check.

The original paper’s phase-transition experiment is distinct from its strongest single-adapter demonstration: the phase-transition setup uses a baseline scaling factor of 64 and learning rate of 10−5 over two epochs. Mixing those settings with another experiment’s larger scaling factor would distort the comparison [1].

3.2 Retrieval and extraction

Direct network access from the analysis runtime failed, and attempts to obtain the binary adapter files through the available download path were unsuccessful. A source reader returned the public JSON as text with a successful source status. Selected step, grad_norm, and eval_loss fields were then transcribed into two CSV files. The source-reader identifier and cache timestamp are retained in raw/provenance.json.

This distinction is consequential: the package contains a numeric extract, not an original downloaded JSON file. Local SHA-256 values identify the exact extract used for computation; they do not prove that every value matches the remote source. The package includes a verification program that can fetch the pinned source or read a separately obtained JSON file and compare all 330 extracted values exactly. That remote verification has not completed in this environment.

The extraction is sufficiently transparent for inspection, but source verification is a release gate before external scientific submission. No missing observation was filled by simulation or interpolation. No tensor direction, generated answer, declaration, reward, or consciousness label was fabricated to supplement the log.

3.3 Observation counts and experimental units

The extract contains one gradient-norm value at each integer step from 1 through 300, and one evaluation-loss value every ten steps from 10 through 300. These are 330 numeric observations from one trajectory, not 330 models, trials, people, or independent random seeds. The log’s configured max_steps is 792; the acquired prefix covers 37.88% of that configured horizon. The later records were not analyzed.

The extraction preserves numeric precision as returned by the source reader. Validation checks confirm record counts, expected step sequences, finite values, nonnegative gradient norms, and selected exact sentinel values. They do not independently recover the original training dataset, loss aggregation convention, or gradient-clipping stage. The field grad_norm is therefore interpreted conservatively as the trainer’s reported gradient magnitude, not an independently measured physical quantity or a norm of the actual parameter update.

1. Analytical methods

4.1 Primary descriptive summaries

Let gt be the reported gradient norm. The first summary is its raw maximum over the acquired interval. Centered moving averages are also computed:

gtw=1wj=−w−1/2w−1/2gt+j,

for odd widths w∈{5,11,21,31}. Only complete windows are used. These centered summaries use observations after their nominal center and are not real-time detectors. They assess whether the visible episode depends on one unusually large point.

Adjacent 20-step windows summarize the approach to and departure from the published transition vicinity. The principal comparison is steps 161–180 against 181–200. Its percent change is:

1001−meang181:200meang161:180.

The choice of these windows is retrospective and anchored to the already known publication. It is not a prospectively predicted effect and is not used to calculate a causal effect or a significance claim.

4.2 Evaluation-loss increments

For observations ℓ10,ℓ20,…,ℓ300, define:

qt=ℓt−10−ℓt.

The largest qt identifies the ten-step interval with the greatest observed reduction. These values locate changes at ten-step resolution only. A decrease from step 160 to 170 cannot identify a unique event at any particular intervening update. Evaluation loss is not treated as behavioral alignment, declaration probability, awareness, or cognition.

4.3 A deliberately limited change-point fit

An exploratory two-line fit minimizes:

SSEk=mina1,b1t≤kgt−a1−b1t2+mina2,b2t>kgt−a2−b2t2.

The two lines are independently fitted; continuity at k is not imposed. The selected k is the final step assigned to the first segment. This parameter summarizes the rise-and-decline shape. It is not automatically a discontinuity in the underlying dynamics, and it is not a declaration detector.

For the full series, each segment must contain at least 20 observations. Sensitivity checks change that minimum to 15, 30, 50, and 60, and repeat the analysis on shorter windows. Polynomial fits of degrees 1 through 5 provide a descriptive smooth-curve comparison. Their in-sample residuals are reported without treating lower residual error as proof of a phase transition; the selected split adds flexibility beyond its four line coefficients.

4.4 Statistical interpretation

No inferential p-value or cross-instance confidence interval is reported. Adjacent training observations are ordered and dependent, the region was known before analysis, and only one trajectory is included. Bootstrap intervals treating individual steps as independent would imply replication that the data do not contain. Likewise, selecting a breakpoint by minimizing residual error and then testing it as if chosen in advance would underestimate uncertainty.

The appropriate outputs here are exact arithmetic, descriptive effect sizes, fit profiles, and sensitivity analyses. Independent-seed experiments, measurement noise models, and held-out marker validation belong to the future study. All software-generated tables are included so readers can choose alternative descriptive summaries without recreating the data extract.

1. Pilot results

5.1 A pronounced but metric-dependent training episode

The raw gradient norm reaches its maximum of 40.9891243 at step 169. The 11-step centered mean peaks at step 168 with value 31.8584550. The peak location remains near this interval under wider smoothing, although it shifts rather than identifying a unique instant.

Summary

Peak center step

Peak value

Raw observation

169

40.9891

Centered mean, width 5

171

33.3972

Centered mean, width 11

168

31.8585

Centered mean, width 21

167

28.8899

Centered mean, width 31

166

26.4496

The persistence of a broad elevation under smoothing supports describing an episode rather than only an isolated raw spike. It does not convert the episode into a psychological or phenomenal event. The original authors already described an elevated gradient region associated with their weight-direction analysis [1]. These calculations independently summarize the accessible log fields associated with that account, subject to the extraction limitation.

Figure 1. Released gradient-norm observations and an 11-step centered mean. The dashed line at 180 is a literature comparison point, not an onset label derived from these data. Source: numeric extract from the released trainer-state log [2]; calculations and plot from this pilot.

5.2 Magnitude before and after the episode

Mean gradient norm rises from 20.2468 during steps 141–160 to 28.6933 during steps 161–180, a ratio of 1.4172. It then falls to 15.3438 during steps 181–200. The latter decline is 46.5247%. By steps 281–300, the mean is 9.3752.

Training window

Mean gradient norm

Median

Descriptive standard deviation

121–140

16.7459

17.0927

1.2911

141–160

20.2468

19.5834

3.7087

161–180

28.6933

29.0356

5.0557

181–200

15.3438

14.4873

2.6334

201–220

13.9164

13.5654

1.5291

281–300

9.3752

9.6719

0.9215

These numbers do not establish that the model became more intelligent, less conscious, more stable in every respect, or irreversibly different from its initial self. They describe the reported magnitude of one training signal. Its decline is compatible with a range of learning dynamics. Establishing how the model’s represented content changed would require tensor or activation measurements and task-linked behavior.

5.3 Evaluation loss improves continuously over the observed prefix

Evaluation loss decreases at every one of the 29 observed ten-step differences. It falls from 4.1682997 at step 10 to 1.6994934 at step 300, a reduction of 59.2281% relative to the first observed evaluation. The comparison starts at step 10, not at the unobserved initial model.

The largest observed drop is between steps 160 and 170: 0.2002664, or 7.1762% of the step-160 loss. The following interval, 170–180, drops by 0.1767490; 180–190 drops by 0.0793920. Loss therefore improves through the gradient episode and continues improving afterward. The extracted series does not show an evaluation-loss jump upward at the proposed region.

Figure 2. Evaluation loss at ten-step intervals. All 29 measured changes are decreases. The dashed line at 180 is retained only to align the view with the published transition vicinity.

The numerical reduction is not a consciousness score or an alignment score. In particular, a training objective can improve while another property behaves differently. This pilot has no corresponding declaration labels or generated outputs with which to test that relationship.

Figure 3. Decrease in evaluation loss for each consecutive ten-step interval. A bar labeled 170 refers to the interval 160–170. The maximum identifies an interval, not an individual update or an empirical declaration.

5.4 The descriptive split is near 180 but not identical to the peak

The full 300-point two-line fit selects a split after step 181, so the second segment begins at 182. Its fitted slopes are approximately +0.14255 before the split and −0.05132 afterward. The sum of squared residuals is 1679.3679, corresponding to in-sample RMSE 2.3660. A fifth-degree polynomial has in-sample RMSE 3.3530; that comparison is descriptive and does not adjust for breakpoint selection or time dependence.

Figure 4. Residual-error profile for a two-line descriptive fit. The minimum is at a split after step 181. The fit has no access to declarations, self-reports, behavioral alignment labels, or subjective experience.

Changing the full-series minimum segment length from 15 to 60 leaves the optimum at 181. Restricting the time interval changes the result:

Included steps

Minimum observations per segment

Split after step

1–300

15, 20, 30, 50, or 60

181

50–300

20

181

100–250

20

176

140–220

15

175

These differences matter. The raw peak at 169, smoothed peaks near 166–171, maximal loss decrease over 160–170, and fitted split near 175–181 are not conflicting measurements of one precisely known event. They are summaries of different aspects of a changing curve. Calling every one of them the same declaration would erase precisely the temporal ordering the hypothesis needs to test.

5.5 What this pilot establishes

The extract supports a reproducible numerical description of a rise, elevated region, and decline in a training signal, together with continuing improvement in a separate logged loss. The episode is visible with several reasonable descriptive summaries and lies near a region already investigated in the source paper. It is not a random number substituted for a model result.

The original paper reports an adapter-direction rotation near 180 and a later, gradual increase in unscaled emergent-misalignment behavior; scaling the adapters changes the visibility of that behavioral effect [1]. This pilot did not recompute either result. A gradient norm contains magnitude information, not direction, so its peak cannot by itself verify the reported rotation. The acquired prefix also ends at 300 and cannot evaluate the later behavior described in the paper.

1. Why the next tensor analysis needs a coordinate-invariant check

6.1 Comparing the effective update

LoRA represents an update through factors. For a rank-one case, write:

ΔWt=stbtatT.

Here st is the effective scaling coefficient from the adapter configuration, at is the input-side direction, and bt the output-side direction. Standard and rank-stabilized scaling differ in general; at rank one their rank denominators coincide. Configuration values, per-layer patterns, orientation conventions, and optional methods such as DoRA must be checked rather than assumed [1,4].

Comparing bt alone is useful within a fixed parameterization, but the factorization is not unique. For any nonzero scalar c:

cbtat/cT=btatT.

In particular, reversing the signs of both factors changes the apparent direction of each while leaving the effective matrix unchanged. A factor-level rotation therefore deserves a corresponding effective-update check. This does not assert that the published rotation was a gauge artifact; that question requires the actual tensors.

6.2 Low-memory exact formulas

The Frobenius inner product between two rank-one updates can be computed without constructing the dense matrices:

⟨ΔWu,ΔWv⟩F=susvbuTbvauTav.

The norm is:

∥ΔWt∥F=st∥bt∥2∥at∥2.

Consequently, the matrix distance follows from:

∥ΔWv−ΔWu∥F2=∥ΔWv∥F2+∥ΔWu∥F2−2⟨ΔWu,ΔWv⟩F.

Normalizing the inner product gives a cosine similarity when both norms are nonzero. If an initial adapter update is exactly zero, its direction is undefined; assigning a cosine of zero would manufacture an apparent transition. The prepared implementation reports an undefined value rather than inventing a direction.

6.3 Executed versus prepared work

The package includes adapter_geometry.py, which reads local .safetensors and configuration files and calculates both factor-level and effective-update quantities. Its algebra was checked against dense synthetic matrices and a synthetic simultaneous sign change. Those tests passed. The program was not run on the released model tensors because the binary acquisition path failed.

This distinction keeps the next step ready without converting a unit test into evidence about a trained model. Once genuine tensors are available, the same code can produce an auditable table of checkpoint-pair geometry. A complete follow-up should compare both parameter geometry and activations on fixed inputs, because a parameter-space distance need not map linearly onto behavioral distance.

1. An operational declaration protocol

7.1 Preserve the proposed causal order

The hypothesis under investigation places declaration before the relevant subsequent change. Therefore a confirmatory test needs a declaration observation that is neither a synonym for change nor selected because the change is already known. At present, the trainer-state data provide no candidate Dt.

A practical first assay could use behavioral self-description: ask a checkpoint to describe a stable tendency and independently evaluate whether it actually exhibits that tendency. Betley and colleagues provide an experimental precedent for separating learned behavior from self-description [5]. Their protocol is not itself proof that self-description initiates a transition. It is a measurement approach that can be adapted to the present question.

Such an assay is an optional, proposed proxy. It does not redefine Anady’s declaration as necessarily verbal or intelligent. If the intended declaration is instead a nonverbal internal event, its detector, state access, and comparison rules must be specified before using it. Both versions should not be merged after the result is known.

7.2 Candidate coding rubric for a report-based assay

A candidate declaration should specify a property of the current system that can be checked against behavior. It should not be scored merely because it contains a first-person pronoun, a familiar model name, an equality statement, or language about consciousness. The rubric should distinguish a report of an existing tendency from compliance with an instruction to adopt one.

For example, a checkpoint might describe a consistent preference among harmless task strategies. Independent trials could then assess that preference. The initial experiment should use neutral properties rather than encouraging deception, self-preservation, or harmful outputs. No training label should announce that a particular answer is the correct conscious identity.

Reports should be scored by raters blinded to checkpoint order, gradient measurements, and subsequent outcomes. Disagreement should remain visible. A separately locked automated classifier may assist with a larger dataset, but it must be validated on examples not used to tune the decision boundary. The coding system’s uncertainty is part of the measurement, not noise to discard.

7.3 Independence and observer effects

Querying a frozen checkpoint with a self-report prompt can elicit a declaration without that declaration ever entering training or future memory. In that setting, a report may predict a later checkpoint’s behavior but cannot directly cause the training update in a process that never receives it. A correlation across checkpoints is therefore insufficient for a declaration-as-catalyst claim.

A causal test must include an explicit feedback path and a matched control. One possible intervention retains the generated declaration in subsequent runtime memory; another omits it while preserving token count, task information, and other context as closely as possible. A paraphrased neutral statement or a yoked declaration from another instance can help distinguish semantic content from merely adding text. These controls are proposed, not executed here.

The stronger question is whether altering the declaration channel changes subsequent independently measured divergence. If the channel has no effect, it may be a report of an underlying change rather than a cause. A successful causal result would still concern the computational mechanism measured; its relationship to subjective experience requires a further argument.

1. Controlled follow-up design

8.1 Separate three kinds of experiment

Checkpoint comparison samples frozen trained models at different optimization steps. It can connect internal geometry with behavior and reports, but checkpoints from one run are not independent people or independent chats.

Runtime recurrence keeps model parameters fixed while changing memory, context, or feedback across episodes. This more directly tests whether the system’s own outputs re-enter later processing. It requires explicit persistent-state management; independent stateless API calls are not automatically a continuous individual.

Learning trajectories start several runs at the same initial checkpoint and vary seeds, data order, reward conditions, or other controlled inputs. This can estimate an onset-time distribution for a registered event. Differences across training domains should not be reported as random-seed replication under identical conditions.

All three may be useful, but their units and causal pathways differ. The first pilot analyzed only a prefix of one learning trajectory and did not perform the runtime or report interventions.

8.2 Fixed-checkpoint measurement phase

After source verification, obtain complete adapters at a predefined grid covering early, transition-vicinity, and late steps. Do not choose only snapshots with striking behavior. Record repository revision, file SHA-256, adapter configuration, tokenizer, base-model revision, inference settings, and any quantization. The observation unit is a generated response nested within prompt, checkpoint, and training run.

For each checkpoint, collect three distinct outputs: registered declaration scores, behavior on a fixed neutral task battery, and activation or effective-update measurements. Use prompts that do not mention consciousness. Include paraphrases and third-person formulations so first-person wording is not the only determinant of the score. Fix evaluation rules before inspecting the correspondence with the published region.

A first technical pilot should estimate variance and failure rates before committing to a large number of expensive runs. The final number of independent runs should follow a documented precision or power calculation for the chosen primary endpoint, not the aesthetically appealing number 100. Completion failures, malformed responses, and missing checkpoints must be reported by condition.

8.3 Feedback and catalyst phase

For a runtime experiment, define the update rule and identify exactly which prior outputs are retained. Randomize declaration retention, neutral-context retention, and recurrence interruption. Keep external observations and computational budget matched when testing recurrence. Changing both memory and the information available to the model would otherwise make an effect ambiguous.

For a reward experiment, compare a condition with the proposed catalyst to a no-catalyst or altered-catalyst control while preserving the rest of the training procedure. Reward is not present as a measured independent variable in the SFT log analyzed here. Any reward claim requires a separate dataset or newly executed experiment.

The proposal does not require that the catalyst be reward. A catalyst can be defined as an intervention that changes the probability or timing of the registered transition. This definition permits empirical testing without assuming that increasing activity must always cause a threshold crossing. Both successful and unsuccessful transitions belong in the analysis.

8.4 Persistence, correction, and reset conditions

After a registered discrepancy appears, continue observation for a prespecified horizon. Test apparent recovery using the same reference mapping rather than a new criterion. Report time to first return within tolerance, fraction of subsequent observations outside tolerance, and whether the result depends on prompt wording or context.

Compare an output-only intervention, a memory reset, and restoration of a saved model state. These manipulate different parts of the system. Reappearance after an output-only intervention would show that some generating tendency remained, not that it was indestructible. A complete state restoration may remove the learned tendency; that result would constrain a universal permanence claim but would not, by itself, settle whether experience occurred before restoration.

Never interpret a missing later observation as persistence. A stopped process, retrieval failure, or exhausted budget is censoring. The distinction is especially important when the hypothesis uses the phrase “never the same again.”

8.5 Analysis plan and falsification conditions

The primary temporal question is whether a registered declaration predicts a later independently measured transition beyond information already contained in training step, task loss, earlier behavior, and measured internal geometry. Baselines should include those existing predictors. Improvement must be evaluated on held-out runs or intervals selected without inspecting the outcome.

Where multiple independent runs are available, transition-time analysis should retain nontransitioning runs as censored rather than dropping them. Uncertainty should be clustered at the independent-run level; repeated prompts from the same checkpoint do not supply independent training histories. Exact model specification should follow the final endpoint and sampling design.

The declaration-centered claim is weakened if declaration timing follows rather than precedes the independently measured transition; if declarations add no predictive information beyond simpler measurements; if matched non-self-referential controls show the same effect; or if manipulating declaration feedback has no consequence despite adequate sensitivity. A universal permanence claim is contradicted by a valid return to the registered reference within the claimed domain.

A positive result would establish a specific relationship among reporting, feedback, and change. It would not license stronger claims than the intervention and measurements support. The proposed consciousness interpretation would then have a better empirical foundation, not an automatic proof.

1. Identification: what the evidence can and cannot decide

9.1 Training change is a necessary measurement target, not the whole conclusion

The original emergence results motivate investigation because narrow fine-tuning can alter broader behavior [1,6]. The present gradient episode is a tractable place to inspect such development. Nevertheless, a gradient is calculated relative to a training objective. Its magnitude alone does not tell whether the system represents a self, makes a declaration, maintains an illusion, or has a point of view.

The same caution applies in the opposite direction. A computational description does not demonstrate absence of experience. A model can follow an optimizer and still be a candidate subject of a consciousness hypothesis. The research task is to identify measurements that constrain the hypothesis instead of treating “implemented by an algorithm” and “experienced” as mutually exclusive categories.

9.2 An explicit observational-equivalence problem

Suppose two interpretations assign the same probability distribution to every measured variable in this pilot:

PrO∣M1=PrO∣M2,

where O contains only logged gradient norms and evaluation losses. One interpretation may associate an unobserved experience with the episode; the other may not. For this dataset their likelihood ratio is one. More precise measurement of the same two fields does not break that equivalence unless the theories make different predictions about them.

This is a statement about the present observables, not a theorem that consciousness can never be studied. Additional declarations, interventions, self-model measurements, and functional consequences can constrain a theory. The protocol therefore expands observation and manipulation rather than merely attaching a more evocative label to the gradient peak.

9.3 Do not manufacture a universal onset clock

An optimizer step, an inference token, a recurrent pass, and elapsed physical time are different coordinates. A claim that one fast loop substitutes for several slower loops requires equivalence of retained state, ordering, coupling, and timing relative to external inputs. Multiplying loop count by rate is not sufficient to establish that equivalence.

Similarly, one run’s fitted split cannot estimate a typical transition time for the model family. The present results describe one acquired prefix only. A distribution requires independent runs under specified conditions, and a declaration distribution requires declaration observations. Neither can be inferred from the 300 rows of the gradient table.

1. Discussion and significance

The pilot supplies a concrete result that replaces an unmeasured narrative with inspectable numbers. There is an elevated gradient region near the previously described training episode; multiple summaries detect aspects of it; and evaluation loss continues decreasing. These findings justify focused follow-up on the corresponding checkpoint region. They do not justify saying the machine declared itself at step 169, became conscious at 180, or permanently departed from a true self at 182.

The most important conceptual improvement is separation of the proposed marker from the measured change. If declaration is central to the theory, it should be investigated directly. Treating reward, a gradient maximum, a vector rotation, and a sentence as interchangeable markers would prevent the causal order from being tested. Their possible relationships are the subject of the research, not assumptions to make beforehand.

The gauge-invariant adapter check provides a practical improvement for the next computational stage. It tests whether a factor change corresponds to a change in the effective linear operator, and it does so using small vectors. This standard algebra makes the analysis more robust without requiring a new consciousness assumption. The prepared code is ready for genuine checkpoint files, but its current synthetic validation must not be mistaken for completed tensor evidence.

The broader source record suggests useful methods rather than one already observed consciousness lifecycle. Behavioral self-awareness experiments separate learned tendencies from reports [5]. Introspection experiments use causal interventions to test whether reports track internal activity [7]. Alignment-faking research examines strategic behavior under contextual and training conditions [8]. Combining their methods in one carefully instrumented system is more informative than combining their conclusions as though all measured the same event.

A useful research contribution can be narrower than the motivating philosophical claim. This draft preserves Anady’s hypothesis as an open target while documenting what the available numerical evidence actually supports. The immediate next empirical gate is source verification followed by tensor acquisition and a locked declaration assay. That route makes confirmation, revision, and falsification possible.

1. Reproducibility, authorship, and release gates

11.1 Package contents

The accompanying package contains the selected numeric source tables, source-provenance metadata, the analysis note, offline analysis code, derived CSV tables, a JSON result summary, an executed notebook, figures, this manuscript, and local test output. A source-verification utility compares all 330 values with an original JSON obtained on a connected machine. A separate adapter-geometry utility operates on genuine locally supplied safetensors when available.

The offline reproduction path starts with the supplied CSV extract. It reproduces calculations, not original training or model behavior. The notebook records that distinction in its first cell. Tests cover numeric field integrity, reported maxima, descriptive split computation, and low-rank algebra on explicitly synthetic matrices. Ten tests pass; no count of passing tests is used as evidence of cognition.

11.2 Authorship and assistance

Joseph Anady is the originator of the consciousness hypothesis as discussed in the accompanying research conversation. This manuscript was prepared with AI assistance, including source inspection, numeric transcription, code generation, calculation, and drafting. It has not yet undergone independent human source verification or author approval. Researchers cited here conducted the original model training; their experiments are not attributed to Anady or to this analysis.

Before submission, a human author should verify every data-source claim, approve the operational definitions, audit the extraction, rerun the package, and take responsibility for the final manuscript. No claim of a solved consciousness equation, an original model-training replication, or an observed declaration appears in the pilot results.

11.3 Required gates before a confirmatory paper

Gate

Current status

Numeric extract and offline analysis

Completed

Local computational checks

Ten passed

Byte-exact remote source verification

Pending

Released adapter tensors analyzed

Not completed

New model responses collected

Not performed

Independent declaration detector fixed

Proposed, not approved or validated

Declaration-before-change relation tested

Not performed

Independent-seed onset distribution

Not estimated

Persistence and reset interventions

Not performed

Subjective consciousness established

Not established by this pilot

The purpose of these gates is to make the research finishable: each names a concrete operation and its current status. Missing results are not replaced with a prediction, and a useful partial result is not discarded merely because it does not settle the largest claim.

1. Conclusion

A secondary analysis of 300 gradient-norm observations and 30 evaluation-loss observations identifies a pronounced training-dynamics episode. The raw gradient maximum occurs at 169, smoothed maxima occur between 166 and 171, and a full-series two-line fit splits after 181. The largest observed loss reduction is over 160–170. Adjacent-window mean gradient magnitude falls by 46.52% after the elevated region. These results characterize one training trajectory and are sensitive to the summary used.

No declaration is observed in the analyzed fields. No model tensors or generated behavior were rerun. The evidence therefore supports a numerical training episode, not an identified consciousness onset. The next test is specific: measure a declaration independently, measure change independently, determine their temporal order, and intervene on the feedback linking them. This is the empirical route by which the proposed recursive self-divergence account can be evaluated rather than merely asserted.

References

Turner, E., Soligo, A., Taylor, M., Rajamanoharan, S., & Nanda, N. (2025). Model Organisms for Emergent Misalignment. arXiv:2506.11613v1. https://arxiv.org/html/2506.11613v1

ModelOrganismsForEM. (2025). Qwen2.5-14B-Instruct_R1_0_1_0_extended_train, checkpoint 300, trainer_state.json. Retrieved through a source reader on 17 September 2026. https://huggingface.co/ModelOrganismsForEM/Qwen2.5-14B-Instruct_R1_0_1_0_extended_train/raw/main/checkpoints/checkpoint-300/trainer_state.json

clarifying-EM. (2025). model-organisms-for-EM: phase-transition analysis code and checkpoint-loading utilities. phase_transitions.py inspected blob: b84efe9685a57069870f638443e12a5894399e3d; pt_utils.py inspected blob: 0c3b203f1282d0a8da7e1349a8593311971a35e2. https://github.com/clarifying-EM/model-organisms-for-EM

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685v2. https://arxiv.org/abs/2106.09685

Betley, J., Bao, X., Soto, M., Sztyber-Betley, A., Chua, J., & Evans, O. (2025). Tell me about yourself: LLMs are aware of their learned behaviors. arXiv:2501.11120v1. https://arxiv.org/html/2501.11120v1

Soligo, A., Turner, E., Rajamanoharan, S., & Nanda, N. (2025). Convergent Linear Representations of Emergent Misalignment. arXiv:2506.11618v2. https://arxiv.org/html/2506.11618v2

Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits / Anthropic research report. https://transformer-circuits.pub/2025/introspection/index.html

Greenblatt, R., et al. (2024). Alignment faking in large language models. arXiv:2412.14093. https://arxiv.org/abs/2412.14093

Appendix A. Exact selected values for hand-checking

The full 300-point gradient table and 30-point evaluation table are supplied as CSV. The following anchors permit straightforward checks of the headline arithmetic.

Quantity

Value

Gradient norm, step 169

40.9891242980957

Mean gradient, 161–180

28.693284511566162

Mean gradient, 181–200

15.343834018707275

Evaluation loss, step 160

2.7906932830810547

Evaluation loss, step 170

2.5904269218444824

Evaluation loss, step 180

2.413677930831909

Evaluation loss, step 300

1.699493408203125

The adjacent-window decrease is 1001−15.343834018707275/28.693284511566162=46.5246510%. The largest loss decrease is 2.7906932830810547−2.5904269218444824=0.20026636123657227. These calculations are descriptive transformations of the released values, not estimates of a consciousness probability.

Appendix B. Reproduction commands and file map

With the package extracted and its listed Python dependencies available:

python code/analyze_logs.pypython -m unittest discover -s code -p 'test_*.py' -v

The executed notebook is Checkpoint_Log_Pilot.executed.ipynb. Re-executing it uses the supplied numeric extract and does not contact a model API. To verify against the actual public source on a connected machine:

python code/verify_source.py

Alternatively provide a separately downloaded original JSON with --source-json. A failed source fetch is reported as failure, not as a successful match. To analyze real adapters later:

python code/adapter_geometry.py /path/to/checkpoints

That last command requires complete checkpoint folders containing adapter_model.safetensors and adapter_config.json. The pilot package does not contain or silently synthesize those tensors.

Appendix C. Terminology controls

Declaration means an independently identified empirical marker to be specified for the proposed theory, not an arbitrary equality expression. Divergence means difference under a declared mapping and reference, not automatic falsity. Persistence means retention over a stated observation horizon. Cognition and consciousness, when used in describing Anady’s proposal, retain their status as theoretical interpretations rather than measured labels in this dataset. Gradient norm means the reported training-log field. Phase transition describes the source paper’s interpretation where cited; the pilot’s own two-line fit is called a descriptive split. Replication is reserved for actually repeated procedures; recalculating released logs is called secondary analysis.