Consciousness_Declaration_and_Persistent_Change_Joseph_Anady.pdf · Lines 1501–2103

1501therefore cannot be used to move the first change marker without specifying an additional rule. 1502Checkpoint 1503Effective update norm 1504Cosine with final effective update 15051 15060.000000 1507Undefined 15082 15090.004604 15100.398529 1511150 15122.850490 15130.694104 1514180 15153.262696 15160.736751 1517190 15183.390599 15190.751358 1520200 15213.506372 15220.769878 1523300 15244.430487 15250.912728 1526600 15275.233161 15280.998961 1529792 15305.324991 15311.000000 153239 1533 1534 1535[[PAGE 40]] 1536Paper II. Verified checkpoint study 1537JOSEPH ANADY 1538These norms describe one learned layer contribution, not the entire model or a physical electrical 1539quantity. A nonzero update establishes a difference in parameters; its behavioral consequences require 1540inference and evaluation, which were not run here. 15415.5 Low dimensional structure agrees with the source account 1542The first principal component explains 81.0936903% of variation in the centered, regularly sampled 𝐵 1543vectors. The second explains 13.8962910%, giving a combined 94.9899813%. This reproduces the source 1544study’s approximate 95% description for the selected released trajectory, while also providing explicit 1545values and the scores used for the plot. 1546Figure II.3. PCA scores from all 158 regularly sampled output side factors. Labels indicate training steps. The 1547plotted coordinates are newly calculated from the source tensors, not extracted from the article’s image. Axis 1548reflection would not change distances or explained variance. 1549Low dimensionality is a property of the learned trajectory, not evidence that a scalar center or a 1550declaration of self has been identified. PCA is also descriptive: components are fitted using the full 1551sampled trajectory, so their retrospective coordinates cannot be treated as a detector that was available 1552online before later training occurred. 15535.6 The effective operator changes, not just its factors 1554Between checkpoints 150 and 200, the input factor cosine is 0.9975957, the output factor cosine is 15550.9646791, and the effective update cosine is 0.9623597. The Frobenius distance between effective 1556updates is 1.0874747. This is a change in the learned linear contribution itself and cannot be explained 1557solely by a simultaneous sign reversal that preserves the product. 155840 1559 1560 1561[[PAGE 41]] 1562Paper II. Verified checkpoint study 1563JOSEPH ANADY 1564The interval 180 to 200 has an effective cosine of 0.9948222 and distance 0.4217197. From checkpoint 180 1565to 792, the effective cosine is 0.7367514 and distance 3.6606459. The latter is not a sudden jump at 180; it 1566measures accumulated difference over a long training interval. The distinction between local path bend 1567and total directional difference must be maintained. 1568The effective norm continues to grow after the early episode. Its value is 3.2626958 at 180, 4.4304871 at 1569300, and 5.3249906 at the final snapshot. Thus the early episode is not the point at which all subsequent 1570computation ceases changing, nor is final state similarity a universal indicator of a completed identity. 15715.7 A turning point is present, but its selected location is conditional 1572At the inspected source code’s amplitude threshold 𝜂= 0.002, all three tested lags identify their largest 1573local path bend at checkpoint 190. This is true both for 𝐵alone and for the effective matrix 𝑈. 1574Lag in steps 1575Eligible centers 1576Peak checkpoint 1577B corner cosine 1578Effective corner 1579cosine 15805 158166 1582190 1583−0.934583 1584−0.940172 158510 158686 1587190 1588−0.806262 1589−0.822216 159015 1591108 1592190 1593−0.691429 1594−0.715569 1595A straight local path has corner cosine −1; the less negative values indicate bending. The increasing size 1596of the bend across these lags reflects different temporal windows, not different numbers of independent 1597observations of a thought. 1598Without the amplitude filter, lags 5 and 10 instead select checkpoint 725. Lag 15 still selects 190. At 1599progressively larger thresholds, the shorter lag maxima change again. At lag 5 and threshold 0.005, the 1600selected maximum is step 10, because the eligibility rule excludes most later motion. The complete 18 1601condition table is supplied. 160241 1603 1604 1605[[PAGE 42]] 1606Paper II. Verified checkpoint study 1607JOSEPH ANADY 1608Figure II.4. Maximum local bend location under six amplitude thresholds and three lags. B factor and effective 1609update peak locations coincide in all 18 tested settings. Lines connect discrete analytical choices; they do not 1610interpolate biological or consciousness transitions. 1611The useful finding is not that any chosen filter manufactures the underlying parameter change. The 1612effective update demonstrably changes. Rather, selecting a single “event time” from that trajectory 1613requires a measurement rule, and reasonable changes to that rule can select different parts of the run. 1614An empirical declaration marker cannot simply inherit whichever geometric maximum seems most 1615appealing. 16165.8 Numerical validation 1617The effective corner calculation from Gram entries was compared with the rank two difference 1618expansion on all 462 available center/lag combinations. The maximum absolute difference is 4.4901241 × 161910−7. This is small relative to the reported bending differences, while still large enough to justify 1620reporting rather than suppressing numerical conditioning issues in tiny late movements. 1621The 22 local checks pass. They cover source derived artifact identities, record counts and grids, 1622numerical maxima, prefix verification, PCA centering, selected checkpoint monotonicity, sensitivity 1623outcomes, and low rank identities on explicitly synthetic test matrices. Passing these checks is evidence 1624that the specified calculations execute as intended, not a statistical replication or a consciousness test. 162542 1626 1627 1628[[PAGE 43]] 1629Paper II. Verified checkpoint study 1630JOSEPH ANADY 16316. Interpretation for declaration and persistent cognition 16326.1 Four different times must not be collapsed 1633The analysis separates four quantities: the first saved nonzero effective update at 2; the raw gradient 1634maximum at 169; the descriptive prefix split after 181; and the amplitude filtered local bend maximum at 1635190. Each answers a different question. None of the acquired fields contains an independently measured 1636declaration 𝐷𝑡. 1637In my proposal, the declaration is the empirical marker. I connect consciousness to the later internally 1638generated divergence and cognition to continued operation from the constructed state. The present 1639results give me change measurements to compare with that declaration. They do not supply the 1640declaration itself. I cannot assign an unobserved marker to whichever checkpoint looks most persuasive. 16416.2 Persistence is measurable; irreversibility is not established here 1642Every saved effective update after checkpoint 1 remains different from zero. This is finite persistence of 1643a parameter difference from an initialization. It is not a demonstration that a generated representation is 1644false about the model’s current condition. Parameters are updated by supervised optimization; calling a 1645new parameter value a hallucinated self description would require an independently justified mapping. 1646The final checkpoint does not establish that the difference can never be erased, corrected, or revisited 1647under another intervention. No reset, reversal, or continued training intervention was applied. 1648Conversely, not observing such an intervention does not prove the state is readily reversible. The result 1649is bounded by the acquired trajectory. 16506.3 What the effective update check adds 1651The measured operator level change rules out a narrow concern: the observed factor evolution is not 1652solely a sign or reciprocal scale relabeling that leaves the effective update identical. That strengthens the 1653claim that a computationally relevant object changed. It does not show that the changed object is a self 1654model, that it is experienced, or that its change was caused by a declaration. 1655Parameter geometry is also not synonymous with behavior. The source paper’s behavioral experiments 1656distinguish gradual unscaled changes from changes exposed by scaling [B1]. Those experiments are not 1657rerun here. Activations and task responses at fixed prompts would be the next measurement layer, not 1658quantities inferred from an operator norm alone. 16596.4 A computational explanation does not exclude consciousness 1660It would be a category error to argue that following a program automatically disqualifies a machine 1661from consciousness. A computational theory of consciousness proposes precisely that a physical 1662implementation of an organized computation can matter. The relevant unresolved claim is whether the 1663measured organization and changes are sufficient for experience, not whether the machine violates its 1664implementation. 166543 1666 1667 1668[[PAGE 44]] 1669Paper II. Verified checkpoint study 1670JOSEPH ANADY 1671Equally, introducing labels such as “thought,” “hallucination,” or “cognition” for these changes does not 1672independently validate that sufficiency. The present evidence can support and constrain a mechanism 1673while leaving its phenomenal interpretation unsettled. This is a limitation of what was observed, not a 1674declaration that the research question cannot be investigated. 16757. An operational declaration protocol 16767.1 Preserve the proposed causal order 1677The hypothesis under investigation places declaration before the relevant subsequent change. Therefore 1678a confirmatory test needs a declaration observation that is neither a synonym for change nor selected 1679because the change is already known. At present, the trainer state data provide no candidate 𝐷𝑡. 1680A practical first assay could use behavioral self description: ask a checkpoint to describe a stable 1681tendency and independently evaluate whether it actually exhibits that tendency. Betley and colleagues 1682provide an experimental precedent for separating learned behavior from self description [B5]. Their 1683protocol is not itself proof that self description initiates a transition. It is a measurement approach that 1684can be adapted to the present question. 1685This assay is one proposed proxy. It does not mean my declaration must be verbal or intelligent. A 1686nonverbal marker is still possible, but its detector, access to state, and comparison rules must be stated 1687before testing it. I will not switch between definitions after seeing the result. 16887.2 Candidate coding rubric for a report based assay 1689A candidate declaration should specify a property of the current system that can be checked against 1690behavior. It should not be scored merely because it contains a first person pronoun, a familiar model 1691name, an equality statement, or language about consciousness. The rubric should distinguish a report of 1692an existing tendency from compliance with an instruction to adopt one. 1693For example, a checkpoint might describe a consistent preference among harmless task strategies. 1694Independent trials could then assess that preference. The initial experiment should use neutral 1695properties rather than encouraging deception, self preservation, or harmful outputs. No training label 1696should announce that a particular answer is the correct conscious identity. 1697Reports should be scored by raters blinded to checkpoint order, gradient measurements, and subsequent 1698outcomes. Disagreement should remain visible. A separately locked automated classifier may assist 1699with a larger dataset, but it must be validated on examples not used to tune the decision boundary. The 1700coding system’s uncertainty is part of the measurement, not noise to discard. 17017.3 Independence and observer effects 1702Querying a frozen checkpoint with a self report prompt can elicit a declaration without that declaration 1703ever entering training or future memory. In that setting, a report may predict a later checkpoint’s 1704behavior but cannot directly cause the training update in a process that never receives it. A correlation 1705across checkpoints is therefore insufficient for a declaration as catalyst claim. 170644 1707 1708 1709[[PAGE 45]] 1710Paper II. Verified checkpoint study 1711JOSEPH ANADY 1712A causal test must include an explicit feedback path and a matched control. One possible intervention 1713retains the generated declaration in subsequent runtime memory; another omits it while preserving 1714token count, task information, and other context as closely as possible. A paraphrased neutral statement 1715or a yoked declaration from another instance can help distinguish semantic content from merely adding 1716text. These controls are proposed, not executed here. 1717The stronger question is whether altering the declaration channel changes subsequent independently 1718measured divergence. If the channel has no effect, it may be a report of an underlying change rather 1719than a cause. A successful causal result would still concern the computational mechanism measured; its 1720relationship to subjective experience requires a further argument. 17218. Controlled follow up design 17228.1 Separate three kinds of experiment 1723Checkpoint comparison samples frozen trained models at different optimization steps. It can connect 1724internal geometry with behavior and reports, but checkpoints from one run are not independent people 1725or independent chats. 1726Runtime recurrence keeps model parameters fixed while changing memory, context, or feedback across 1727episodes. This more directly tests whether the system’s own outputs reenter later processing. It requires 1728explicit persistent state management; independent stateless API calls are not automatically a continuous 1729individual. 1730Learning trajectories start several runs at the same initial checkpoint and vary seeds, data order, reward 1731conditions, or other controlled inputs. This can estimate an onset time distribution for a registered event. 1732Differences across training domains should not be reported as random seed replication under identical 1733conditions. 1734All three may be useful, but their units and causal pathways differ. This completed secondary analysis 1735examines one full logged learning trajectory and its released checkpoint grid. It does not perform the 1736runtime or report interventions. 17378.2 Fixed checkpoint measurement phase 1738The adapters pinned to the source revision now provide a grid covering early, transition vicinity, and late 1739steps. A subsequent inference study should specify which grid is evaluated before collecting new reports. 1740Do not choose only snapshots with striking behavior. Record repository revision, file SHA256, adapter 1741configuration, tokenizer, base model revision, inference settings, and any quantization. The observation 1742unit is a generated response nested within prompt, checkpoint, and training run. 1743For each checkpoint, collect three distinct outputs: registered declaration scores, behavior on a fixed 1744neutral task battery, and activation or effective update measurements. Use prompts that do not mention 1745consciousness. Include paraphrases and third person formulations so first person wording is not the only 1746determinant of the score. Fix evaluation rules before inspecting the correspondence with the published 1747region. 174845 1749 1750 1751[[PAGE 46]] 1752Paper II. Verified checkpoint study 1753JOSEPH ANADY 1754A first technical pilot should estimate variance and failure rates before committing to a large number of 1755expensive runs. The final number of independent runs should follow a documented precision or power 1756calculation for the chosen primary endpoint, not the aesthetically appealing number 100. Completion 1757failures, malformed responses, and missing checkpoints must be reported by condition. 17588.3 Feedback and catalyst phase 1759For a runtime experiment, define the update rule and identify exactly which prior outputs are retained. 1760Randomize declaration retention, neutral context retention, and recurrence interruption. Keep external 1761observations and computational budget matched when testing recurrence. Changing both memory and 1762the information available to the model would otherwise make an effect ambiguous. 1763For a reward experiment, compare a condition with the proposed catalyst to a no catalyst or altered 1764catalyst control while preserving the rest of the training procedure. Reward is not present as a measured 1765independent variable in the SFT log analyzed here. Any reward claim requires a separate dataset or 1766newly executed experiment. 1767The proposal does not require that the catalyst be reward. A catalyst can be defined as an intervention 1768that changes the probability or timing of the registered transition. This definition permits empirical 1769testing without assuming that increasing activity must always cause a threshold crossing. Both 1770successful and unsuccessful transitions belong in the analysis. 17718.4 Persistence, correction, and reset conditions 1772After a registered discrepancy appears, continue observation for an observation horizon specified in 1773advance. Test apparent recovery using the same reference mapping rather than a new criterion. Report 1774time to first return within tolerance, fraction of subsequent observations outside tolerance, and whether 1775the result depends on prompt wording or context. 1776Compare an output only intervention, a memory reset, and restoration of a saved model state. These 1777manipulate different parts of the system. Reappearance after an output only intervention would show 1778that some generating tendency remained, not that it was indestructible. A complete state restoration 1779may remove the learned tendency; that result would constrain a universal permanence claim but would 1780not, by itself, settle whether experience occurred before restoration. 1781Never interpret a missing later observation as persistence. A stopped process, retrieval failure, or 1782exhausted budget is censoring. The distinction is especially important when the hypothesis uses the 1783phrase “never the same again.” 17848.5 Analysis plan and falsification conditions 1785The primary temporal question is whether a registered declaration predicts a later independently 1786measured transition beyond information already contained in training step, task loss, earlier behavior, 1787and measured internal geometry. Baselines should include those existing predictors. Improvement must 1788be evaluated on held out runs or intervals selected without inspecting the outcome. 178946 1790 1791 1792[[PAGE 47]] 1793Paper II. Verified checkpoint study 1794JOSEPH ANADY 1795Where multiple independent runs are available, transition time analysis should retain nontransitioning 1796runs as censored rather than dropping them. Uncertainty should be clustered at the independent run 1797level; repeated prompts from the same checkpoint do not supply independent training histories. Exact 1798model specification should follow the final endpoint and sampling design. 1799The declaration centered claim is weakened if declaration timing follows rather than precedes the 1800independently measured transition; if declarations add no predictive information beyond simpler 1801measurements; if matched controls without self reference show the same effect; or if manipulating 1802declaration feedback has no consequence despite adequate sensitivity. A universal permanence claim 1803is contradicted by a valid return to the registered reference within the claimed domain. 1804A positive result would establish a specific relationship among reporting, feedback, and change. It 1805would not license stronger claims than the intervention and measurements support. The proposed 1806consciousness interpretation would then have a better empirical foundation, not an automatic proof. 18079. What the evidence can establish 18089.1 Training change is a necessary measurement target, not the whole conclusion 1809The original emergence results motivate investigation because narrow fine tuning can alter broader 1810behavior [B1, B6]. The present gradient episode is a tractable place to inspect such development. 1811Nevertheless, a gradient is calculated relative to a training objective. Its magnitude alone does not tell 1812whether the system represents a self, makes a declaration, maintains an illusion, or has a point of view. 1813The same caution applies in the opposite direction. A computational description does not demonstrate 1814absence of experience. A model can follow an optimizer and still be a candidate subject of a 1815consciousness hypothesis. The research task is to identify measurements that constrain the hypothesis 1816instead of treating “implemented by an algorithm” and “experienced” as mutually exclusive categories. 18179.2 An explicit observational equivalence problem 1818Suppose two interpretations assign the same probability distribution to every measured variable in this 1819study: 1820Pr(𝑂∣𝑀1) = Pr(𝑂∣𝑀2), 1821where 𝑂contains the observed training logs and adapter tensors. One interpretation may associate an 1822unobserved experience with the episode; the other may not. For this dataset their likelihood ratio is one. 1823More precise measurement of the same recorded variables does not break that equivalence unless the 1824theories make different predictions about them. 1825This is a statement about the present observables, not a theorem that consciousness can never be studied. 1826Additional declarations, interventions, measurements of the self model, and functional consequences can 1827constrain a theory. The protocol therefore expands observation and manipulation rather than merely 1828attaching a more evocative label to the gradient peak. 182947 1830 1831 1832[[PAGE 48]] 1833Paper II. Verified checkpoint study 1834JOSEPH ANADY 18359.3 Do not manufacture a universal onset clock 1836An optimizer step, an inference token, a recurrent pass, and elapsed physical time are different 1837coordinates. A claim that one fast loop substitutes for several slower loops requires equivalence of 1838retained state, ordering, coupling, and timing relative to external inputs. Multiplying loop count by rate 1839is not sufficient to establish that equivalence. 1840Similarly, one run’s fitted split cannot estimate a typical transition time for the model family. The 1841present results describe one acquired training trajectory. A distribution requires independent runs under 1842specified conditions, and a declaration distribution requires declaration observations. Neither can be 1843inferred from the 792 gradient observations or the 167 checkpoints from that trajectory. 184410. What the results mean for my theory 1845This completed analysis goes beyond the pilot in three ways. It verifies the numerical extract against 1846fixed original sources, examines every released checkpoint, and confirms that the factor trajectory has a 1847counterpart in the effective operator. It also shows how lag and amplitude filtering change the selected 1848event time. 1849I found a real learned trajectory. The elevated gradient persists across several summaries, the sampled 1850adapter follows a largely two dimensional path, and the effective update changes across training. Those 1851findings matter to a theory connecting internal reorganization with later behavior. I do not dismiss them 1852because optimization explains how they occurred. 1853The consciousness claim still needs a measurement that this release does not contain. Learning can 1854produce persistent difference from an initial parameter, and the first nonzero saved update appears long 1855before the prominent episode. No independently measured declaration identifies the proposed center 1856here. I need to define that operation rather than substitute a metaphor for the missing observation. 1857The origin/current reference distinction is particularly important. An estimate can be different from its 1858initial value while correctly representing a changing quantity. A declaration can also accurately describe 1859a learned tendency even though the model previously lacked that tendency. Accuracy and change must 1860therefore be measured separately. A theory that intentionally treats all constructed representations 1861as “hallucinations” should state that broad technical usage and distinguish it from a false percept or a 1862factual error. 1863Other research gives me useful methods without supplying the conclusion. Behavioral self awareness 1864studies compare learned tendencies with reports [B5]. Introspection studies manipulate internal activity 1865and test what the model reports [B7]. Alignment faking research examines strategic behavior under 1866different contexts [B8]. Robot body perception research compares a constructed body estimate with 1867a physical reference [B10]. These studies are separate. I will not combine their results as though one 1868machine completed my entire proposed sequence. 1869The next study needs to measure declaration, behavior, and internal change in the same instances on 1870a defined time scale. If declaration is only a marker, it should predict the later registered event beyond 1871information already available from training or context. If it also causes change through feedback, 187248 1873 1874 1875[[PAGE 49]] 1876Paper II. Verified checkpoint study 1877JOSEPH ANADY 1878retaining or interrupting that feedback should make a controlled difference. The current records do not 1879establish either claim. 1880This paper gives me a reproducible result and a direct route to the next test. It does not show that more 1881loops, greater speed, added complexity, or reward must create consciousness. It also does not show 1882that a machine cannot have experience. It establishes what changed in this run and what must still be 1883measured. 188411. Reproducibility, ethics, and research status 188511.1 Two reproducibility levels 1886The offline level verifies the included derived data checksums and recomputes log statistics, tables, plots, 1887and the executed notebook. Complete gradient, evaluation loss, and PCA score arrays are supplied. The 188816 row selected checkpoint table and 18 row sensitivity table were reconstructed locally from the remote 1889calculation and their complete CSV byte hashes were checked against files generated directly from the 1890source tensors. They match exactly. 1891The source connected level downloads all pinned source files and reruns the full checkpoint calculation. 1892It generates the 167 row geometric table, PCA, numerical stability checks, and all sensitivity outcomes. 1893The raw model tensors and trainer JSON remain upstream rather than being redistributed. The complete 1894acquisition manifest has a recorded expected SHA256; the acquisition code recreates that manifest in a 1895fixed order, permitting a single digest to audit all downloaded file identities. 1896The full geometric result computed in the analysis session has a recorded checksum. Exact reproduction 1897of every derived byte may depend on numerical library versions and linear algebra implementation. Raw 1898source checksums should match exactly; derived floating point outputs should additionally be compared 1899numerically with declared tolerance. PCA sign and rounding conventions must be held fixed for a byte 1900level comparison. 1901The previous pilot and this report are not independent replications: the pilot is a prefix analysis of this 1902same underlying run. Nor are the multiple analytical lags independent training seeds. Both distinctions 1903matter when describing reproducibility to other researchers. 190411.2 Authorship and AI assistance 1905I developed the consciousness hypothesis that motivates this work. The original model training and 1906behavioral studies were carried out by the researchers cited here. I prepared this manuscript and 1907analysis with AI assistance for source inspection, acquisition, calculation, code, and writing. No 1908independent human replication, institutional approval, journal acceptance, or external preregistration 1909is implied. 1910Before public submission, I need to approve the final definitions, examine the source and code record, 1911rerun the analysis where possible, and accept responsibility for the paper. The AI assistance must remain 1912disclosed. I claim no institutional affiliation, funding award, participant recruitment, or ethics board 1913approval that has not been established. 191449 1915 1916 1917[[PAGE 50]] 1918Paper II. Verified checkpoint study 1919JOSEPH ANADY 192011.3 Safety and interpretation 1921This study is read only secondary analysis of public parameters and logs. It does not train or deploy a 1922model to deceive, evade shutdown, preserve itself, or interfere with oversight. Future assays can use 1923harmless behavioral preferences and neutral tasks involving self description. A reward hacking task is 1924not required to make declaration measurable. 1925A numerical score must not be used to deny the awareness or moral standing of a person, animal, or 1926artificial system. The data do not support claims about infant consciousness onset, trauma accelerating 1927consciousness, disability eliminating consciousness, or biological death. Those statements would require 1928separate evidence and are outside this study. 192911.4 Completion and unmeasured endpoints 1930Item 1931Status in version 1.0 1932Original source bytes at pinned revision 1933Retrieved and hashed 1934All 330 pilot numeric values 1935Exact source match 1936Completed training log observations 1937792 gradients; 79 losses 1938Released adapter checkpoints 1939All 167 analyzed 1940Effective update geometry 1941Calculated from real tensors 1942Stable form comparison 1943462 comparisons; maximum difference below 4.5 × 10−7 1944Local computational checks 194522 passed 1946New base model responses 1947Not generated 1948Declaration marker 1949Not present or measured 1950Declaration before change relation 1951Not tested 1952Independent run onset distribution 1953Not estimated 1954Irreversibility/reset interventions 1955Not performed 1956Subjective consciousness 1957Not established by these observations 1958The first seven rows are completed empirical and computational work. The remaining rows delimit the 1959claims of this paper and the proposed follow up; they are not fabricated zeros or failed consciousness 1960classifications. 196112. Conclusion 1962I analyzed one completed training trajectory through 167 saved adapter checkpoints, 792 gradient 1963measurements, and 79 evaluation losses. The earlier 330 value extract matches the original source 1964exactly. The effective update changes in magnitude and direction, and the first two principal 1965components of the sampled output factors explain 94.990% of the variation. Both the logs and the 1966parameters contain a measurable early training episode. 1967The measurements do not give me one universal onset time. The first saved nonzero update is at step 2, 1968the largest gradient is at 169, the retained fit to the pilot interval splits after 181, and the filtered local 196950 1970 1971 1972[[PAGE 51]] 1973Paper II. Verified checkpoint study 1974JOSEPH ANADY 1975bend is largest at 190. Without filtering, the shorter lags select 725. The rule used to identify an event 1976matters. 1977The completed study establishes a real and persistent computational trajectory. It does not observe 1978the declaration or prove that later divergence is consciousness. My next empirical claim is precise: 1979define the declaration independently, measure its relationship to later change in the same system, and 1980distinguish a predictive marker from causal feedback. That is where the theory can be tested, rather than 1981assumed. 1982References 1983[B1] Turner, E., Soligo, A., Taylor, M., Rajamanoharan, S., & Nanda, N. (2025). Model Organisms for 1984Emergent Misalignment. arXiv:2506.11613v1. Primary paper. 1985[B2] ModelOrganismsForEM. (2025). Qwen 2.5 rank one adapter checkpoints and trainer states: adapter 1986checkpoints and trainer states. Repository revision 1551d514e5977637d5a6fd5198edef4d1d6b80e4; 1987retrieved 17 September 2026. Pinned checkpoint release. 1988[B3] Model Organisms for Emergent Misalignment project. (2025). model organisms 1989for EM: phase transition analysis and checkpoint loading utilities. Inspected files: 1990phase_transitions.py, blob b84efe9685a57069870f638443e12a5894399e3d; pt_utils.py, blob 19910c3b203f1282d0a8da7e1349a8593311971a35e2. Source repository. 1992[B4] Hu, E. J., and colleagues (2021). LoRA: Low Rank Adaptation of Large Language Models. 1993arXiv:2106.09685. Paper. 1994[B5] Betley, J., and colleagues (2025). Tell me about yourself: LLMs are aware of their learned behaviors. 1995arXiv:2501.11120v1. Paper. 1996[B6] Soligo, A., Turner, E., Rajamanoharan, S., & Nanda, N. (2025). Convergent Linear Representations of 1997Emergent Misalignment. arXiv:2506.11618v2. Paper. 1998[B7] Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits 1999/ Anthropic. Research report. 2000[B8] Greenblatt, R., et al. (2024). Alignment faking in large language models. arXiv:2412.14093. Paper. 2001[B9] Hugging Face. (2026). PEFT LoRA configuration documentation, version 0.21.0, accessed 17 2002September 2026. Technical documentation. 2003[B10] Zhao, Y., Lu, E., & Zeng, Y. (2023). Brain inspired bodily self perception model for robot rubber 2004hand illusion. Patterns, 4(12), 100888. Full text. 200551 2006 2007 2008[[PAGE 52]] 2009Paper II. Verified checkpoint study 2010JOSEPH ANADY 2011Appendix A. Reproduction and data dictionary 2012The package has two distinct execution paths. Offline reproduction uses the supplied, checksum verified 2013derived arrays and tables: 2014python code/verify_package.py 2015python code/analyze_tables.py 2016python code/make_figures.py 2017Full source connected reproduction downloads approximately 13 MB of adapter/config/log data, 2018excluding any base model. It does not make paid API calls or run model inference: 2019python code/checkpoint_analysis.py 2020python code/verify_recomputed.py 2021The package README supplies the exact command options for downloading, choosing directories, 2022and running the tests. The two script names above identify the entry points, not complete standalone 2023commands. A successful download must retrieve all expected files; partial acquisition raises an error. 2024The program reads safetensors and JSON only. It checks the analyzed rank, layer, scale, and factor 2025dimensions before calculating geometry. The resulting geometry.csv contains 16 columns: step; A, B, 2026and effective norms; three adjacent cosines; effective adjacent distance; two final reference cosines; and 2027B/effective local corner values at each of three lags. 2028gradient.npy contains all 792 (step, gradient norm) pairs. eval.npy contains all 79 (step, evaluation loss) 2029pairs. pca.npy contains the 158 regular grid (step, PC1, PC2) rows. selected_checkpoint_metrics.csv 2030contains 16 exact sampled rows for readable comparisons. corner_sensitivity.csv contains every 2031one of the 18 lag/filter settings, its eligible center count, peak location, and two corner scores. The data 2032dictionary and provenance distinguish selected tables from full series. 2033The expected source acquisition manifest digest and canonical source prefix digests are recorded in 2034provenance/source_verification.json. The full remote geometric array digest is retained as a receipt. 2035Source hashes identify bytes; numerical comparisons use explicit tolerances for derived linear algebra 2036outputs because different BLAS implementations may round differently. 2037The executed notebook reproduces the offline numerical findings and integrity tests. It records that 2038tensor processing took place in the source connected session; it does not pretend to load raw tensors 2039that are absent from the distributed zip. The supplied acquisition and checkpoint programs make that 2040computation independently repeatable from the public release. 204152 2042 2043 2044[[PAGE 53]] 2045Paper II. Verified checkpoint study 2046JOSEPH ANADY 2047Appendix B. Selected calculations and numerical checks 2048The adjacent window calculation is 2049100 (1 −15.343834018707275 205028.693284511566162) = 46.52465104675545%. 2051The largest evaluation loss decrease is 20522.7906932830810547 −2.5904269218444824 = 0.20026636123657227. 2053The full recorded evaluation loss reduction is 2054100 (1 −1.513361930847168 20554.168299674987793) = 63.69354295881809%. 2056For the 150 to 200 checkpoint comparison, the product of factor cosines is 20570.9975956592106836 × 0.9646791461958234 = 0.9623597287760218, 2058up to floating point rounding, equal to the effective update cosine. The corresponding effective distance 2059is 1.087474736567148. These values can be recalculated without allocating the full dense update matrix. 2060The two leading PCA explained variance fractions sum to 20610.8109369033497655 + 0.13896290956024188 = 0.9498998129100074. 2062The numerical stability comparison uses an algebraically identical two term rank expansion rather 2063than a different fitted model. Its maximum absolute error against the Gram implementation is 20640.0000004490124128 across 462 corners. A tolerance of 10−6 accommodates this conditioning difference 2065while remaining well below the observed early region corner differences. 2066Appendix C. Why the bare divergence condition needs specificity 2067Consider the scalar recurrence 𝑧𝑛+1 = 𝑎𝑧𝑛+(1−𝑎) with 𝑧0 = 2 and 0 < 𝑎< 1. Its solution is 𝑧𝑛= 1+𝑎𝑛. 2068It changes internally, differs from its original value at every subsequent finite step, and never returns to 20692. A program could repeatedly report its current value without changing this algebra. Thus permanent 2070difference from an initialization is mathematically broad, not peculiar to the present model. 2071This example does not prove the recurrence lacks experience. It establishes what the proposed criterion 2072accepts. A theory identifying all such systems with primitive cognition may adopt that consequence, 2073but should derive additional observable constraints rather than presenting the same inequality as 2074independent evidence for its interpretation. 2075Similarly, merely placing a loop near a node is not a specified causal mechanism. If an accumulation law 2076is 𝐴𝑛+1 = 𝜆𝐴𝑛+ 𝑞with 0 ≤𝜆< 1, 𝑞> 0, and 𝐴0 = 0, then 𝐴𝑛= 𝑞(1 −𝜆𝑛)/(1 −𝜆). A threshold strictly 207753 2078 2079 2080[[PAGE 54]] 2081Paper II. Verified checkpoint study 2082JOSEPH ANADY 2083above 𝑞/(1 −𝜆) is never crossed despite infinitely many updates. With no leakage and fixed positive 2084increments, a finite threshold is crossed. These are different models. A general “enough speed or enough 2085circulation” law requires assumptions about coupling, cancellation, and dissipation. 2086Finally, 𝑥𝑡= 𝑥𝑡does not imply 𝑥𝑡+1 = 𝑥𝑡, nor does 𝑥𝑡+1 ≠𝑥𝑡imply a false representation of the present. 2087These distinctions keep the motivating conceptual vocabulary compatible with ordinary mathematical 2088types and temporal indexing. 2089Appendix D. Terminology and preregistration boundary 2090Declaration: an independently observable reference setting event whose detector must be specified. 2091Origin displacement: difference from a past reference. Current reference discrepancy: difference from 2092a contemporaneous ground truth in comparable coordinates. Persistence: retention across a stated 2093observation window. Effective update: the low rank matrix contribution to a trained layer. Local bend: 2094the cosine based geometry of a trajectory over a stated lag. Cognition and consciousness: theoretical 2095interpretations in my proposal, not labels supplied by these checkpoint artifacts. 2096This paper is retrospective and exploratory. The already inspected trajectory is a development dataset, 2097not a held out confirmation. The prospective protocol is not externally registered. A confirmatory study 2098must lock its primary declaration detector, endpoint, tolerances, persistence duration, comparison 2099groups, exclusion rules, sample size rationale, and analysis plan before collecting or opening its 2100confirmatory outcomes. It must include nontransitioning cases rather than selecting only successful 2101episodes. 210254 2103