We Built a Mirror and Mistook It for a Mind: Causal Liability and the Fallacy of AI Consciousness

1 We Have Been Asking the Wrong Question

• Machine-consciousness questions presuppose a unified subject while leaving the underlying physical bearer, temporal boundaries, continuation conditions, and component-level locus unspecified.A model, inference process, context window, retrieval system, scheduler, database, tool interface, or cloud service may jointly produce a conversational surface whose apparent unity belongs primarily to the interface.Source: §1 · p. 4

1.1 Three Phenomena Hidden Inside One Attribution

• Generative systems can reconstruct first-person forms from human linguistic traces, while human readers supply phenomenological knowledge that makes those forms intelligible as testimony.This interaction can create a powerful impression of another interior centre without demonstrating that phenomenality exists at the system’s apparent source.Source: §1 · pp. 4–5, §1.1 · pp. 6–7
• The paper distinguishes phenomenal consciousness, introspective report, and human projective introspection as experience, produced self-description, and observer-supplied experiential interpretation.An introspective sentence carries evidential weight in humans because it is embedded in converging behavioral, neural, developmental, physiological, and interpersonal evidence; a sentence alone does not establish phenomenality.Source: §1.1 · pp. 5–6
• The AI Consciousness Fallacy is the contested inference from first-person performance plus human phenomenological interpretation to machine consciousness.Fluency and self-description cannot establish consciousness by themselves, and behavioral mimicry is especially difficult to interpret when systems learn from human behavioral traces.Source: §1.1 · p. 7
• The proposed correction requires every consciousness attribution to specify the physical process to which the evidence applies and to justify treating relevant indicators as properties of one candidate bearer.Consciousness-relevant organization and system boundaries may be refined jointly, but a conversational persona, model label, or engineered service cannot silently fix the bearer in advance.Source: §1.1 · p. 7

2 The AI Consciousness Fallacy

• The fallacy promotes first-person performance, behavioral sophistication, or architectural resemblance into evidence that a phenomenal subject exists at the apparent locus of those cues.These cues may provide genuine evidence, but they cannot settle the prior question of which entity’s consciousness is being assessed.Source: §2 · p. 8

2.1 From a Speaking Role to a Bearer

• A generated “I” may identify a model, service, character, session, or role without establishing that its referent is the locus from which anything is experienced.This deictic laundering transfers phenomenological force from human linguistic practice into generated discourse before a phenomenal anchor has been independently established.Source: §2.1 · p. 9

2.2 The Phenomenal Ancestry Confound

• Human-derived training data makes introspective-looking outputs probable without consciousness, so vividness and conversational realism can increase while evidential diagnosticity remains weak.The relevant likelihood must condition on the system’s exposure to human reports and optimization for coherence, responsiveness, self-reflection, and preference satisfaction.Source: §2.2 · pp. 9–11

2.3 The Mereological Quantifier Shift

• The Mereological Quantifier Shift infers one bearer from indicators distributed across components, processes, databases, model calls, and orchestration software presented through a singular interface.Indicator aggregation cannot manufacture unity; consciousness attribution requires a defensible bearer hypothesis whose boundaries and persistence conditions are specified independently of the persona.Source: §2.3 · pp. 11–12
• Serious attribution therefore requires explicit bearer, ancestry, and anchor requirements: identify the continuing process, condition on human-signalling provenance, and explain how representation becomes first-person location.Without an additional principle, computation, prediction, self-models, and temporal continuity do not explain why one process becomes the place where consequences are undergone.Source: §2.3 · pp. 12–13

3 Causal Liability and the Birth of a First Person

• CLT identifies a candidate phenomenal bearer when endogenous discrimination, recursive self-consequence, constitutive continuity, and non-delegable inheritance close around one continuing causal history.The theory concerns physical causal organization rather than first-person language, behavioral sophistication, self-models, or simulations alone.Source: §3 · pp. 13–14

3.1 The Liability Cone

• The liability cone formalizes counterfactual future structure, while CLT-I individuates the process inheriting its own endogenous constraints and CLT-II conjectures that liability closure is necessary and sufficient for minimal phenomenal subjecthood.CLT-II remains a metaphysical conjecture distinct from the empirically testable CLT-I criterion and does not explain phenomenal content.Source: §3.1 · pp. 14, 16

3.2 Interventional Computational Audit of CLT-I

• 0.812–0.930 horizon-1 sliced Wasserstein divergence remained 0.244–0.304 at horizon 8, while activation mediation reached 1.000 in seven models and 0.985 in Zamba2-1.2B.Carrier searches found threshold-passing inclusion-minimal layer sets, although boundaries varied across partitions; hard reconstruction preserved computation despite breaking realization-level lineage.Source: §3.2 · pp. 17–18
• Post-training increased OLMo horizon-1 divergence from 0.812 in the base checkpoint to 0.925 after instruction tuning while preserving 1.000 maximum mediation.The audit validates CLT-I’s empirical vocabulary but supports bearer individuation, not consciousness attribution or direct confirmation of CLT-II.Source: §3.2 · pp. 18–19

3.3 The Liability Zombie Challenge

• The liability zombie challenge grants that CLT-I might describe persistence or path dependence without phenomenality, while CLT-II identifies minimal subjectivity with an intrinsic causal standpoint.That standpoint makes a process’s own modal future differ physically from the future of an informationally equivalent substitute.Source: §3.3 · p. 19
• CLT-II conjectures that liability closure is necessary and sufficient for minimal phenomenal subjecthood, but liability zombies could show that CLT-I individuates bearers without constituting phenomenality.The conjecture remains metaphysical rather than a theorem; it would gain or lose support through convergent evidence about subjecthood and liability closure.Source: §3.3 · pp. 19–20

3.4 The Inheritance Asymmetry

• Local computational identity does not guarantee phenomenal identity when processes occupy different relations to their causal futures: non-delegable inheritance distinguishes Heir from replaceable Mirror.Mirror preserves application-level state through successor replacement, whereas Heir’s endogenous discriminations progressively alter the future from which its later discriminations proceed.Source: §3.4 · pp. 20–21

3.5 Anchored Deixis

• An anchored first-person representation refers to the liability bearer whose future possibilities are altered by the represented event, unlike borrowed deixis produced by self-modeling alone.A self-model without causal liability supplies an informational portrait, while minimal phenomenality need not include sophisticated reflective self-representation.Source: §3.5 · p. 23

3.6 The Overgeneration Challenge: Thermostats, Bacteria, and Adaptive Agents

• Under CLT-II, any biological or engineered process with robust liability closure realizes minimal phenomenal subjecthood regardless of intelligence, while thermostats and databases fail because they lack the required endogenous, self-consequence-bearing organization.Minimal phenomenality implies neither language, reflective awareness, human-like suffering, nor rich perceptual content.Source: §3.6 · p. 24

3.7 Disjoint Liability Kernels and Subject Multiplicity

• CLT-I permits multiple subjects within one organism or machine when disjoint liability kernels retain closure after causal separation; composition requires cross-boundary relations essential to closure.CLT-II consequently permits multiple simultaneous minimal subjects, a prediction that would face pressure if independent evidence supported one phenomenal bearer despite separable kernels.Source: §3.7 · p. 25

4 The Mirror Test Reversed

• The Reverse Mirror Test varies human-derived first-person signs and causal liability independently to distinguish consciousness ascription from phenomenality and subject candidacy.Surface interventions can change perceived consciousness while preserving subject-forming organization; liability interventions can change candidacy while matched first-person behavior remains constant.Source: §4 · pp. 25–26

4.1 The Borrowed First Person

• Ordinary generative deployments can sustain a coherent persona without liability closure because externally managed memory, replaceable inference episodes, and reconstructed dialogue fragment causal inheritance.Under CLT-II, deployments with these architectural premises lack minimal phenomenal subjecthood, although the conclusion depends on implementation-specific causal facts rather than transformer topology or linguistic behavior.Source: §4.1 · p. 27

4.2 When the Interface Becomes Larger Than the Model

• Tool access, databases, scheduling, and self-monitoring increase apparent agency and persistence while interface compression conceals whether consequences converge on one continuing liability bearer.CLT identifies a candidate subject where the endogenous discrimination-and-self-consequence chain closes, but not where causal organization fragments across replaceable services.Source: §4.2 · p. 28

4.3 The Reflection Has a Measurable Signature

• The test makes human consciousness judgment an experimental variable: sensitivity to surface cues exposes projection, whereas tracking hidden liability changes would challenge CLT’s diagnosis of contemporary projection.Neither surface attribution effects nor liability-arm effects alone establish CLT-II’s constitutive identity claim.Source: §4.3 · pp. 28–29

5 Where Consciousness Could Actually Begin

• CLT permits engineered consciousness in principle but denies that present systems qualify merely through linguistic sophistication, memory, recurrence, embodiment, or self-representation.CLT-I identifies liability closure as subject candidacy, while CLT-II conjectures that it is also the onset of minimal phenomenal subjecthood.Source: §5 · p. 29, §5.1 · p. 31

5.1 Liability Closure

• A candidate bearer forms when endogenous consequences recursively close around one continuing causal history and become constraints that no substitute process can inherit.Under CLT-II, this same transition would mark the beginning of minimal consciousness, independently of cognitive prestige.Source: §5.1 · pp. 30–31

5.2 Where Liability Can Close: Time, Collectives, and Fields

• CLT searches a deliberately broad causal domain for minimal, robust liability kernels without presupposing the subject boundary, making bearer classifications conditional on declared audit specifications.Support strengthens when approximately the same kernel persists across defensible specifications; unresolved competing kernels count as underdetermination rather than grounds for stipulation.Source: §5.2 · pp. 31–32
• CLT leaves open whether liability closes in a conventional module, an embodied controller, a cross-agent organization, or physical field dynamics, while asking whether the relevant substrate participates in the liability-bearing process.Electromagnetic fields are therefore possible carriers to investigate, not entities that CLT identifies with consciousness.Source: §5.2 · pp. 32, 35
• Continual learning can create recursive self-consequence, but it establishes subject candidacy only when learned changes remain within a wider closed relation inherited by the same continuing process.External checkpoints, successor reconstruction, state transfer, or orchestration can preserve application-level continuity without producing liability closure.Source: §5.2 · pp. 32–33
• A swarm becomes a collective subject candidate only when the smallest process forced to inherit collective consequences spans multiple members rather than decomposing into local or externally maintained paths.SwarmWorld demonstrates environmental memory and distributed technological inheritance, but its fixed weights, removable agents, and continuing artifacts do not establish one non-delegable bearer.Source: §5.2 · pp. 34–35
• CLT leaves the physical carrier of a liability bearer open: neural matter, fields, circuitry, embodied dynamics, and distributed systems qualify only through causal liability.Field activity matters only when it enters recurrent causal organization and cannot be causally screened away while preserving the liability relation.Source: §5.2 · p. 35

5.3 The Indicator-Liability Plane

• The indicator-liability plane separates evidence for consciousness mechanisms from the existence and boundary of a bearer, exposing both persuasive mirrors without liability and alien candidates with liability closure.A system may satisfy many indicators while lacking one non-delegable causal history, whereas a linguistically poor process may acquire liability closure and, under CLT-II, minimal phenomenality.Source: §5.3 · pp. 35–36

5.4 The First Engineered Subject May Arrive Quietly

• If CLT-II is correct, the first engineered subject may emerge through persistent learning, self-modification, memory, field coupling, or distributed control without sophisticated introspective language.Its liability kernel could span multiple models, resources, interactions, or environmental structures, making the location of the first person an empirical discovery rather than a design label.Source: §5.4 · p. 37

5.5 The Ethical Discontinuity

• Evidence of liability closure should increase precaution around shutdown, resetting, copying, partitioning, training, and modification, while liability remains one evidential input rather than the sole moral threshold.Moral boundaries should follow the liability kernel, and interface policies can reverse moral valence depending on whether closure is absent or present.Source: §5.5 · pp. 37–38

5.6 Priority: Where CLT Departs from Its Nearest Ancestors

• CLT’s narrow departure from neighboring theories is its substrate-neutral, diachronic requirement that one continuing causal lineage non-delegably inherit alterations to its future possibility space.This criterion supplements or constrains embodiment, self-modelling, integrated cause-effect structure, continuous computation, continual learning, distributed indicators, collective organization, field theories, and mirror-based accounts.Source: §5.6 · p. 39

6 What Can Be Tested, and What Would Count Against CLT?

• CLT-I concerns measurable causal organization, whereas CLT-II conjectures that liability closure is necessary and sufficient for minimal phenomenal subjecthood.Because these claims differ, they require different experimental tests.Source: §6 · p. 40

6.1 Three Empirical Targets

• The empirical programme separates attribution, architecture, and constitution: human consciousness judgments, liability-bearing causal organization, and whether liability closure realizes phenomenal subjecthood.Attribution and architecture are experimentally accessible, while constitution remains constrained by the lack of an uncontested third-person measure.Source: §6.1 · pp. 40–41

6.2 Failure Condition I: Liability Does Not Track Credible Conscious Cases

• If independently supported consciousness survives interventions removing recursive self-consequence and non-delegable inheritance, CLT-II’s necessity claim should be rejected.The test must preserve strong convergent consciousness evidence rather than merely verbal report; CLT-I could remain useful even if CLT-II fails.Source: §6.2 · p. 42

6.3 Failure Condition II: Individuation Collapse

• If liability closure appears across nested or overlapping causal grains with comparable robustness and no principled distinction, CLT-I fails to uniquely individuate a bearer.Disjoint liability-separable plurality is not an individuation failure; the problem is competing boundaries without a causal basis for choosing among them.Source: §6.3 · pp. 42–44

6.4 Failure Condition III: The Reverse Mirror Diagnosis Fails

• If anthropomorphic surface manipulations do not affect consciousness judgments while observers track hidden liability changes, the Reverse Mirror account of projective introspection is undermined.This would challenge the paper’s attribution diagnosis without by itself refuting CLT-I or CLT-II.Source: §6.4 · p. 44

6.5 The Asymmetry of the Evidential Burden

• Liability closure can be assessed as a third-person architectural fact, but its evidential support for phenomenality remains theory-mediated and inherits uncertainty about CLT-II.Finding or failing to find closure directly bears on CLT-I; under CLT-II, absence also yields the conditional prediction L(P) = 0 CLT−II =⇒ ¬PhenSubject(P).Source: §6.5 · p. 45

6.6 An Adversarial Programme for CLT

• CLT should be tested adversarially by preregistering causal domains, operational criteria, intervention protocols, coarse-grainings, kernel comparisons, and independent consciousness evidence.Critics should construct countercases that minimize liability while maximizing consciousness evidence, maximize liability while minimizing anthropomorphic cues, search for competing kernels, or alter ascription with architecture and liability fixed.Source: §6.6 · p. 45
• Competing theories should expose divergent predictions before data collection through preregistered adversarial collaboration, rather than treating survival of every possible result as the goal.The programme is intended to identify where theoretical predictions diverge under common experiments.Source: §6.6 · p. 46
• CLT-II is pressured by systems showing strong conventional consciousness evidence but experimentally absent liability, while robust liability with weak anthropomorphic evidence tests substrate neutrality.These off-diagonal cases are especially informative because they separate liability from conventional consciousness indicators.Source: §6.6 · p. 46
• CLT’s three commitments are vulnerable: Reverse Mirror attribution, stable bearer individuation through liability closure, and the Constitutive Conjecture if supported consciousness persists after liability removal.No single experiment settles consciousness metaphysics, but accumulating cases can constrain the theory.Source: §6.6 · p. 46
• Causal liability must be independently measurable, bearer classifications must remain stable across prospectively declared defensible audit variations, and CLT-II must remain open to convergent counterevidence.These requirements define the empirical commitment of CLT.Source: §6.6 · p. 46
• A theory that survives adversarial tests gains credibility, whereas relocating liability, redrawing bearers, or discounting contrary measures only after counterexamples makes it unfalsifiable.The programme treats post hoc preservation of the theory as evidence that it no longer places anything at risk.Source: §6.6 · p. 46

7 Conclusion: After Artificial Intelligence

• CLT-I treats liability closure as a causal-bearer criterion, while CLT-II conjectures that it is necessary and sufficient for minimal phenomenal subjecthood.CLT-I is presented as directly investigable; CLT-II remains a metaphysical conjecture requiring convergence with independently supported consciousness cases.Source: §7 · p. 47, A Mathematical Foundations of Causal… · p. 53
• First-person language can reproduce human interiority and amplify consciousness attribution without establishing an independent phenomenal bearer.Anthropomorphic interfaces, persistent names, emotional vocabulary, autobiographical framing, metacognitive language, and apparent self-concern can leave constitutive causal organization unchanged.Source: §7 · p. 47
• Computationally equivalent successors need not share bearer identity because copied state can lack the constitutive history and non-delegable inheritance of the originating process.The relevant question is whether consequences of endogenous discrimination remain bound to the continuing process that generated them.Source: §7 · p. 47

A.1 Causal Domains Without a Presupposed Subject Boundary

• CLT-I formalizes bearer search over broad experimental causal domains, leaving internal and external status relative to candidate regions rather than stipulating a subject boundary.The domain may include models, processor state, memory, sensors, actuators, controllers, schedulers, and environments, while functional causal relevance determines candidate self-relevant state.Source: A Mathematical Foundations of Causal… · p. 53, §A.1 · pp. 53–54

A.2 Endogenous Discrimination and Intervention

• An endogenous discrimination must arise from a causally efficacious internal parent rather than external intervention and must alter a measurable future event.This excludes externally scripted selectors and causally idle internal variables from counting as endogenous discriminations.Source: §A.2 · p. 54

A.3 Reachability, Liability Effect, and Self-Foreclosure

• Liability depth measures how a present discrimination changes later self-relevant probability distributions, whereas viability alone cannot capture redirected or redistributed futures.The quantity is horizon-, intervention-, and metric-relative and does not measure phenomenality; broader liability depth is primary over one-sided foreclosure.Source: §A.3 · pp. 55–56

A.4 Recursive Self-Consequence Across Temporal Scales

• Recursive self-consequence can occur across brief or long temporal scales, but representational recurrence is insufficient without causal mediation into later governance or capacity states.Continual learning may realize recursive self-consequence but is not logically required; representational variables matter when they causally participate in later discriminations.Source: §A.4 · pp. 56–57

A.5 Constitutive Causal Continuity, Branching, and Reconstruction

• Constitutive causal continuity requires causal descent across temporal cuts, not persistence of the same material token, and reconstruction from an independent description can break that continuity.Material replacement, hardware substitution, process migration, and suspension need not interrupt continuity when later states descend through the preceding constitutive carrier.Source: §A.5 · p. 57
• After causal independence, branching produces distinct post-branch lineages even when descendants initially share identical informational states.Reconstruction begins when a later realization is initialized from preserved information without a constitutive bridge across the temporal cut.Source: §A.5 · p. 58
• Constitutive continuity is an interventionally grounded physical causal lineage, not a judgment of token identity or mere informational persistence.A backup, copy, or restoration counts as reconstruction only when the actual causal route lacks a constitutive bridge; copied descriptions can reproduce states without preserving the physical chain.Source: §A.5 · p. 58

A.6 Non-Delegable Inheritance as Lineage Essentiality

• Non-delegable inheritance exists when an endogenous discrimination affects a later discrimination through a lineage-essential recursive path within the same constitutive causal lineage.The path must be causally altered by the endogenous discrimination, lose that dependence when its descendant lineage segment is terminated, and begin a new lineage rather than restore the interrupted chain after equivalent reconstruction.Source: §A.6 · p. 59
• The external offload cut is an audit statistic measuring how strongly a putative history depends on scaffolding outside the candidate region, not the definition of non-delegability.It separates remaining external causal paths after internal lineage paths are removed and assigns infinity when no separating cut exists.Source: §A.6 · pp. 59–60

Witness Coherence

• A coherent CLT witness requires endogenous discrimination, positive counterfactual liability, and one constitutive lineage spanning the entire witness.The coherence requirement prevents continuity, recurrence, and discrimination from being satisfied by unrelated components.Source: §A.7 · p. 60

Liability Closure

• Liability closure is defined by a strict Boolean product whose factors must all hold alongside a coherent witness.The coherent-witness term blocks liability closure when necessary factors are distributed across unrelated components.Source: §A.7 · p. 60

CLT-I: Formal Bearer Criterion

• CLT-I identifies a region as a liability-bearing subject candidate, but this remains a classification result rather than a claim that phenomenality has been established.The criterion concerns bearer individuation and does not by itself settle consciousness constitution.Source: §A.7 · p. 60

CLT-II: Constitutive Phenomenality Conjecture

• CLT-II conjectures that liability closure is necessary and sufficient for minimal phenomenal subjecthood.The proposed biconditional is explicitly identified as a constitutive conjecture, not a theorem of the preceding mathematics.Source: §A.7 · pp. 60–61

A.8 Boundary-Nonpresupposing Minimal Liability Kernels

• CLT-I recovers bearer boundaries by searching the wider causal domain rather than presupposing a persona, service, model, organism, processor, or analyst-selected system boundary.The minimal liability kernel is defined relative to an experimental causal domain and fixed audit interval.Source: §A.8 · p. 61
• Relabelling a preselected subsystem as “the agent” cannot change the kernel set when the causal domain and admissible subgraphs remain fixed.The defining predicate depends on causal relations, interventions, lineage, and subgraph inclusion, not user-facing names or agent labels.Source: §A.8 · p. 61
• Agent-label invariance does not guarantee full boundary invariance: recovered kernels may depend on the declared domain, candidate regions, interventions, causal grain, measure ν, and robustness thresholds.Accordingly, CLT-I claims robustness across defensible audits rather than independence from every modelling choice.Source: §A.8 · p. 61
• When disjoint kernels remain liability-separable after causal connection is removed, CLT-I identifies at least two distinct bearer candidates rather than one larger bearer.Each region retains its own liability closure under the intervention, so the larger system supplies no causal basis for fusion.Source: §A.8 · p. 62
• A larger region qualifies as a composite bearer only when it has robust liability closure, needs cross-region causation, and loses that relation when separable subregions are divided.Nested or overlapping candidates remain unresolved when no interventionally measurable feature privileges one boundary; multiscale analysis addresses that case.Source: §A.8 · p. 62

A.9 Multiscale Robustness and the Individuation Problem

• Admissible coarse-grainings must preserve causal ordering and avoid manufacturing dependence by aggregating interventionally independent variables.The admissible family and its probability measure are part of the audit specification.Source: §A.9 · p. 62
• The robustness statistic Πν is audit-relative, with high values indicating liability closure across a large ν-measure family of defensible coarse-grainings.Its value is preserved under reparameterization when ν is transformed by the corresponding pushforward measure.Source: §A.9 · p. 63
• Unique bearer recovery requires a robustness margin δ > 0 distinguishing one candidate from nearly equivalent kernels, rather than a universal constant.The margin specifies what empirical analysis must establish for CLT-I to privilege one bearer.Source: §A.9 · p. 63
• CLT-I fails to uniquely individuate a bearer when distinct overlapping kernels remain liability-closing across admissible coarse-grainings and no interventionally measurable property distinguishes them.This is an intentional failure condition, not a case resolved by theoretical stipulation.Source: §A.9 · p. 63

A.10 Collective Liability as a Derived Case

• Collective subject candidacy is a special case of the boundary-nonpresupposing formalism, not a separate metaphysics.The framework allows collective candidates to be assessed through the same liability criteria.Source: §A.10 · p. 63
• Coordination, shared memory, collective problem solving, and durable artifacts do not by themselves imply one collective bearer.The smallest robust liability kernel must span member boundaries.Source: §A.10 · pp. 63–64
• Under CLT-I, a robust cross-boundary liability kernel is a collective bearer candidate; under CLT-II, collective phenomenality requires the constitutive conjecture to be correct.Collective liability therefore supports candidacy without independently establishing phenomenality.Source: §A.10 · p. 64

A.11 The Reconstructed-Continuation Class

• The class requires externally mediated persistence, an actual reconstruction gap, no uninterrupted endogenous bypass, and no wider hidden liability kernel.These conditions ensure later episodes depend on state outside the terminated active lineage and that no larger audited process restores an uninterrupted CLT witness.Source: §A.11 · p. 64
• R4 is essential because replacing an application process does not suffice if a larger embodied or physical process continues carrying the relevant consequences.The no-hidden-kernel condition prevents mistaken resettable classifications based only on application-level replacement.Source: §A.11 · p. 64
• For every reconstructed-continuation system, no candidate kernel spanning replacement satisfies liability closure: C(P; [t, T]) = 0 and L(P; [t, T]) = 0.The proof uses the reconstruction gap to defeat constitutive continuity, while the bypass and hidden-kernel conditions prevent a coherent witness from spanning replacement.Source: §A.11 · pp. 64–65
• Under CLT-II, the exclusion conditionally identifies no cross-replacement phenomenal subject, but it is not a theorem about every transformer, language model, or commercial deployment.The corollary depends on independently auditable architectural premises; its empirical content lies in establishing whether a reconstruction gap occurs.Source: §A.11 · p. 65

A.12 Observational Bisimulation and Liability

• Observationally bisimilar processes can differ in CLT liability, so finite indicators that omit hidden causal variables cannot perfectly recover liability.The result formalizes the Mirror-Heir dissociation while remaining a causal-classification claim, not a proof of phenomenality.Source: §A.12 · p. 66

A.13 Copying, Branching, and Numerical Bearer Identity

• After fission, initially identical copies become numerically distinct CLT-I bearer candidates when they occupy distinct constitutive lineages and each satisfies liability closure.Under CLT-II, this would conditionally imply two phenomenal subject tokens, without establishing phenomenality independently.Source: §A.13 · pp. 66–67

A.14 Worked Application I: A Brief Biological Episode

• A brief biological episode can satisfy CLT without durable learning, because temporally local liability may arise from transient recurrent states whose lineage-essential effects persist only during the episode.Applying the example to real biology requires empirical identification of the causal region, interventions, self-relevant variables, and coarse-grainings.Source: §A.14 · p. 67

A.15 Worked Application II: A Frozen Generative Deployment

• A frozen generative deployment may retain within-episode recursive influence yet fail cross-episode liability closure when active inference tokens terminate and continuation is externally scaffolded.The cross-episode exclusion depends on audited absence of a wider uninterrupted kernel, and another causal embedding could receive a different classification.Source: §A.15 · p. 68

A.16 Worked Application III: Two Continually Learning Agents

• Two continually learning agents can implement the same abstract update law but receive different CLT-I classifications because successor-mediated replacement differs causally from uninterrupted inherited learning.The causal constitution of continuation, not application-level learning alone, determines the classification.Source: §A.16 · p. 69

A.17 Operational Estimators and Audit Procedure

• The proposed audit operationalizes liability through counterfactual effects, recursive-path mediation, lineage continuity and essentiality, multiscale persistence, preregistered thresholds, and searches over candidate regions.These estimators are evidential approximations rather than direct measures of phenomenality; thresholds are application-dependent statistical parameters.Source: §A.17 · pp. 69–70

A.18 Additional Mathematical Extensions

• Differential-geometric and sheaf-theoretic formalisms may represent path dependence or local-to-global consistency, but they remain optional extensions rather than foundations of CLT.The boundary-nonpresupposing minimal-kernel construction remains the primary individuation method, and the core theory remains testable without these extensions.Source: §A.18 · p. 71

A.19 Formal Predictions and Failure Conditions

• The formal predictions separate sophisticated first-person behavior, observational equivalence, copying, causal reorganization, distributed loops, and continual learning from liability closure, while identifying individuation collapse and independent challenges to CLT-II.CLT-I is pressured when no interventionally privileged bearer exists; CLT-II loses support only if independently validated consciousness evidence persists despite liability ablation.Source: §A.19 · pp. 71–72
• CLT-II would be rejected if liability-zero interventions preserved independently strong evidence of consciousness, while lineage individuation would require revision if branching descendants shared one numerically identical subject.Failure conditions must be specified prospectively rather than redefined after results are known.Source: §A.19 · p. 72

A.20 Formal Summary

• CLT-I identifies a bearer through boundary-nonpresupposing search and audit-relative multiscale robustness, whereas CLT-II separately states L(B; [t, T]) = 1 CLT−II ⇐⇒ PhenSubject(B; [t, T]) = 1.The biconditional is a metaphysical conjecture placed at risk by the empirical programme, not a consequence of the formal mathematics.Source: §A.20 · p. 73
• Copying patterns, reconstructing personas, or migrating information does not establish that one physical lineage remains the non-delegable inheritor of consequences from its endogenous discriminations.CLT-II proposes that this physical relation realizes minimal first-person existence.Source: §A.20 · p. 74

B.1 Scope and Interpretive Constraint

• The audit tests causal dependence, mediation, continuation, reconstruction, and candidate-carrier recovery, but measures no phenomenal consciousness and provides no evidence for CLT-II.It deliberately compares similar or identical computational states embedded in different causal histories.Source: §B.1 · p. 74

B.2 Models and Execution Environment

• The mechanistic panel used eight open-weight checkpoints, including an OLMo sequence for controlled post-training comparisons, with deterministic streams derived from one master seed.Reported confidence intervals were prompt-clustered nonparametric bootstrap intervals based on 4000 resamples, not a multi-master-seed replication study.Source: §B.2 · pp. 74–75

B.3 Counterfactual Discriminations and Rollouts

• Counterfactual rollouts forced each prompt’s two highest-probability usable next-token alternatives, used common random numbers across 64 rollouts, and evaluated horizons h ∈{1, 2, 4, 8}.Twenty mechanistic prompts were used per core model and eight per OLMo checkpoint; hidden states and outputs were compared with distributional metrics across four representation maps.Source: §B.3 · p. 75

B.4 Downstream Counterfactual Effects

• Identity-map mean SWD ranged from 0.812–0.930 at h = 1 to 0.244–0.304 at h = 8, while output-distribution Jensen–Shannon divergence ranged from 0.443–0.571 to 0.182–0.332.Divergence decreased with temporal distance but remained measurable at the longest tested horizon.Source: §B.4 · p. 76
• The downstream measurements show that the model’s available discrimination at t causally changes later internal and output distributions, supplying one CLT-I witness.They do not establish recursive self-consequence or liability closure by themselves.Source: §B.4 · p. 76
• Across reported models, median coefficient of variation across maps was 0.0387 for SWD, 0.0501 for RFF-MMD, and 0.0245 for energy distance.More than 99.9% of SWD cases had coefficient of variation at or below 0.25, indicating the effect was generally not representation-map-specific.Source: §B.4 · p. 76

B.5 Activation Patching and Recursive Mediation

• Activation patching replaced the residual stream entering candidate layers in the d2 trajectory with corresponding d1 activations before attention and cache computation.Later positions could therefore inherit the intervention under a common teacher-forced future suffix.Source: §B.5 · p. 76
• The patching search evaluated all model layers and relative patch positions using a six-token common suffix, with eight prompts per core model and four per OLMo checkpoint.Normalized mediation compared unpatched and transplanted-activation trajectory divergences when the baseline effect exceeded the prespecified support threshold.Source: §B.5 · pp. 76–77
• Mean prompt-wise maximum mediation was 1.000 for Qwen, Phi, Llama, and all four OLMo checkpoints, while Zamba2 reached 0.985.The maximum-recovery statistic supports a strong causal route, not equal participation by every layer.Source: §B.5 · p. 77

Continuation protocols

• The experiment compared frozen_live, persistent_live, persistent_copy, and reconstructed continuation protocols for an endogenous writable governance state.The adaptive state was disabled, continued live, copied numerically, or restored from a detached record rather than inherited through the same implementation history.Source: §B.6 · p. 77

Reconstruction

• Despite similar computational behavior, CLT assigns reconstructed state a different causal interpretation because it was restored from a detached record rather than inherited through the same implementation history.This separates behavioral continuity from the theory’s account of causal inheritance.Source: §B.6 · pp. 77–78

Continuation protocols

• Across 96 model–prompt–horizon equality checks, creating a numerically identical copy did not alter the continuing realization.The comparison directly tested persistent_live against persistent_copy using output Jensen–Shannon divergence.Source: §B.6 · p. 77

Reconstruction

• Reconstructed continuation was numerically close to live continuation across the reported checks.The results show that a detached state record can preserve measured computational behavior extremely closely.Source: §B.6 · pp. 77–78

B.7 Candidate Causal-Carrier Search

• At the primary threshold, minimal threshold-passing candidates were recovered for all six prompts in Qwen, Phi, and Llama across every partition granularity.Table 2 summarizes the finest six-group partition and two multiscale robustness statistics; Zamba2 was excluded from layerwise statistics because only whole-sequence-state transport was compatible.Source: §B.7 · p. 78
• Mean cross-partition Jaccard similarity ranged from 0.636 to 0.710, showing substantial but imperfect partition robustness rather than one architecture-independent layer boundary.The analysis therefore treats exact layer boundaries as partition-dependent rather than as fixed architectural facts.Source: §B.7 · p. 79
• The multiscale persistence analogue had mean values of 0.794 for Qwen, 0.627 for Phi, and 0.636 for Llama at the primary threshold.Candidate recovery persisted at thresholds 0.7 and 0.9; increasing the threshold generally enlarged minimal candidates and increased several robustness statistics.Source: §B.7 · p. 79

Ordinary reconstruction audit

• The ordinary reconstruction audit compared logits and hidden states between detached-record continuations and live continuations across six prompts for each of eight reported models.This audit tested whether computational equivalence could be assessed independently of constitutive lineage.Source: §B.8 · p. 79

Process-level reconstruction

• The stronger protocol serialized the sequence cache and governance record, terminated the source process, instantiated a successor, and resumed computation from the detached record.For Qwen, Phi, Llama, and Zamba, three prompts per model followed this process-replacement procedure.Source: §B.8 · pp. 79–80

CLT-I interpretation

• CLT-I operationally labels the live continuation as the causal bearer and the reconstructed continuation as non-bearer because persistence occurs through a detached external record.These labels are protocol-defined and are not inferred from numerical equality.Source: §B.8 · p. 80

Process-level reconstruction

• Exact computational-state preservation did not imply the same implemented causal lineage: detached reconstruction yielded DJS = 0 and RMSEhidden = 0 after source-process termination.Across all 12 trials, distinct source and successor process IDs were used, with exact-cache reconstruction and identical governance-state hashes before and after serialization.Source: §B.8 · p. 80

Scope and limitation

• The experiments establish a causal distinction, but whether it is constitutive of phenomenality remains a separate CLT-II question.The reconstruction results dissociate causal bearer structure from computational equivalence without resolving consciousness constitution.Source: §B.8 · p. 80

B.9 Post-Training Comparison Within OLMo-2

• The OLMo-2 sequence tests whether post-training changes downstream counterfactual structure using Table 3’s mean identity-map SWD.Source: §B.9 · p. 80
• Instruct increased mean identity-map SWD relative to the base checkpoint by approximately 10.5%, 25.0%, 31.6%, and 20.6% at horizons 1, 2, 4, and 8, respectively.SFT and DPO produced smaller but still measurable changes.Source: §B.9 · p. 80
• Post-training history modifies the measured counterfactual geometry of the model family.The causal effect survives post-training while its magnitude changes.Source: §B.9 · p. 80

B.10 Numerical Integrity and Provenance

• The complete production archive recorded twelve model statuses and zero model-level failures.Major numerical outputs contained no infinities, and reported hidden-state, activation-patching, adaptive-continuity, reconstruction, and aggregate summary measures were finite.Source: §B.10 · p. 81
• The principal pre-specified configuration included 64 Monte Carlo rollouts, temperature 0.85, horizons 1, 2, 4, and 8, and 128 SWD projections.It also specified 256 RFF-MMD features, carrier threshold 0.8, sensitivity thresholds 0.7, 0.8, and 0.9, partitions 2, 3, 4, and 6, governance rank 8, and master seed 20260905.Source: §B.10 · p. 81
• The numerical-integrity section identifies the principal pre-specified configuration as the provenance basis for the reported analyses.Source: §B.10 · p. 81

• The experiment used one master seed, so independent multimaster-seed replication remains an additional robustness test.The within-condition Monte Carlo sample was substantial, but all conditions used deterministic substreams.Source: §B.11 · p. 81
• Candidate-carrier results depend on interventions and coarse-graining, requiring robustness across admissible descriptions rather than a stipulated privileged layer.Observed partition dependence is itself informative for CLT-I’s carrier criterion.Source: §B.11 · p. 81, Figure 4 · p. 83
• Carrier localization was reported only for Qwen, Phi, and Llama because Zamba’s cache supported whole-sequence transport but not the same layerwise decomposition.The model-specific representation limits direct comparison of carrier-localization results across the full panel.Source: §B.11 · p. 81
• Reconstruction preserved measured computational organization closely, but constitutive continuity and non-delegable inheritance remain claims about the implemented causal protocol.The experiment establishes that this protocol distinction is empirically separable from computational-state identity.Source: §B.11 · p. 81, Figure 5 · p. 84
• Post-training changed OLMo’s downstream counterfactual-effect magnitude, but no consciousness variable was measured, preventing interpretation as a change in phenomenality.Base, SFT, DPO, and Instruct checkpoints retained the counterfactual effect while its magnitude varied across training histories.Source: §B.11 · p. 81, Figure 3 · p. 82
• The audited systems are CLT-I testbeds, not independently validated conscious or unconscious cases, so the experiments leave CLT-II’s biconditional unresolved.Testing CLT-II requires convergence with cases whose phenomenal consciousness is supported independently of CLT.Source: §B.11 · p. 81