Does the Substrate Matter, or the Causal Organisation?

Mira is awake during a sequence of reversible neural replacements. A damaged local circuit is first bypassed by a digital module that reproduces every response seen during training. She names objects, recalls instructions and says her visual field is unchanged. The module is then exchanged for a recurrent neuromorphic circuit that matches the original circuit’s state transitions under targeted perturbation. Her reports remain stable. Finally, an ion-electronic biohybrid is inserted, preserving membrane-like dynamics, chemical modulation and local energy coupling. Again, she reports continuity.

Which replacement preserved consciousness?

The ordinary answers arrive too quickly. One side says all three, because cognition and report survived. Another says only the biological replacement, because computation simulates rather than instantiates the relevant process. A third says the middle and final modules are live candidates because they realise the right internal causal organisation, while the output clone merely impersonates it.

The dispute is not yet about carbon versus silicon. It is about which differences remain after a specified equivalence test, and why any remaining difference should matter to experience. “Substrate independent” and “substrate dependent” are conclusions. Neither is an experimental design.

This paper turns Mira’s imagined surgery into a replacement protocol. It separates six grades of equivalence, three theoretical positions and four evidence channels. The objective is not to crown a metaphysics. It is to make each view declare what must survive replacement, what observation would count against it and which claims remain inaccessible even after perfect behavioural continuity.

Part I. “Same function” is a ladder

A cochlear implant, a spreadsheet and a cortical simulation can all be described as functional replacements. The phrase hides three questions. Which function? Under which possible disturbances? Over what interval? A device may reproduce yesterday’s outputs while failing when the environment changes, when an internal state is perturbed or when it must learn without external retraining.

Classical functionalism identifies a mental state through its causal relations to inputs, other mental states and outputs. The Stanford Encyclopedia account explains why this supports multiple realisation: pain need not be identical to one tissue type if another system can occupy the relevant role. The companion discussion of multiple realizability also records the pressure on that move. A role can be drawn so coarsely that physically important differences disappear by definition.

The useful unit is therefore an equivalence claim with four indices. Let O be the observables, I the allowed interventions, H the temporal horizon and ε the tolerated divergence. Two systems are equivalent only relative to the tuple E(O, I, H, ε). A chatbot and a person might be close under five minutes of verbal questions, far apart under sleep deprivation, long-term learning or a hidden intervention on memory, and incomparable under cellular perturbation.

Equivalence rises or falls with the counterfactuals included in the test. Adding a new intervention does not reveal that an earlier result was dishonest. It reveals that the earlier equivalence occupied a lower grade.

Part II. What biology could add

The brain is not a digital computer made from inconvenient material. Neurons integrate signals across dendritic branches, membranes, synapses and chemical environments. Glia regulate transmission and metabolism. Neuromodulators change gain, learning and action selection across timescales. Bodies supply interoceptive and homeostatic constraints. Development builds the system while it is already acting in a world.

Some of those details may be implementation noise for consciousness. Some may realise a portable dynamical role. Some may be constitutive at their own physical scale. The experiment should not decide among those options by admiration for biology or confidence in software.

Human neurons already warn against equating a neuron with one scalar activation. A published study of human cortical dendrites found a distinctive calcium-mediated dendritic action potential and showed that individual dendrites could support computations unlike familiar point-neuron abstractions. The result concerns cellular computation, not consciousness. Its relevance is methodological: a replacement that matches somatic spikes may still omit causal transformations performed inside the dendritic tree.

Artificial hardware is also moving beyond binary caricatures. Organic electrochemical neurons have produced bioplausible spiking, stochasticity and ion or neurotransmitter-sensitive modulation, and an organic-neuron study demonstrated stimulation of a biological nerve. This does not show artificial experience. It shows that “silicon versus biology” is becoming a poor taxonomy. Candidate substrates can mix electron transport, ions, living tissue and conventional computation.

Biology becomes scientifically relevant when a named feature changes the intervention profile at the same place that conscious processing changes. The feature must pass two filters. First, it must contribute something the proposed artificial replacement lacks. Second, manipulating that contribution must alter a registered consciousness-linked contrast rather than only general health or task performance.

Type a and type b biological naturalism

A current preprint on what biology can and cannot tell us offers a useful separation. Type A biological naturalism says biology matters intrinsically even when it adds no unique information-processing capacity. A biological and non-biological system could be identical on every experiment, yet only the biological system would be conscious. The position may express a metaphysical truth, but empirical evidence cannot favour it if all evidence channels are fixed by hypothesis.

Type B says biology matters because it affords a specific processing capacity. That claim can enter the replacement theatre. Remove the capacity while holding easier functions fixed, observe a consciousness-linked change, then attempt to reproduce the capacity in another substrate. If the effect follows the capacity, the study supports the mechanism. If it follows biological material after the capacity is matched, the Type B description was incomplete or the result belongs to Type A.

Type B turns biological naturalism from a material veto into a research programme. It can even converge with a constrained functionalism if the relevant function is possible only, or practicably only, in living systems. Both sides then study the same mechanism while disagreeing about its portability.

Part III. If consciousness is primary, substrate changes the verb

Materialist and functionalist debates usually ask what produces or constitutes consciousness. A consciousness-primary orientation permits another relation: a substrate may condition how awareness is manifested, bounded or expressed without manufacturing awareness from non-awareness.

This does not make the substrate irrelevant. A scratched lens changes an image without creating light. A damaged radio changes reception without producing the broadcast. These analogies are limited because they can smuggle in a separate signal source for which there is no empirical evidence. Their disciplined use is grammatical. They remind the experimenter that “causes experience to exist”, “organises experienced content” and “enables a report” are different claims.

Śaṅkara’s Advaita distinguishes self-illuminating consciousness from the mind-body complex through which a limited empirical perspective appears. The Stanford Encyclopedia account discusses the mind as an inner instrument and the use of limiting-adjunct analogies. On this view, replacing the instrument could transform the contents, memory and apparent boundary of a perspective without altering consciousness as the fundamental reality.

Aristotle’s hylomorphism creates a different pressure. In the Aristotelian account of psychology, soul is the form or actuality of a living body, not a program detachable from its matter and not another piece of matter inside it. A living human capacity belongs to the organised organism. This resists both crude material identity and the idea that any formally equivalent carrier must instantiate the same living actuality.

Neither tradition is a prediction engine for neural implants. Their contribution is to widen the relation between organisation and mind. Advaita questions whether awareness is produced at all. Hylomorphism questions whether the relevant organisation can be specified independently of the living whole. Contemporary functionalism asks which causal roles are portable. A good protocol can preserve the distinctions without pretending to adjudicate them through vocabulary.

A consciousness-primary view must not turn universal awareness into automatic machine personhood. Even if awareness is fundamental, a distinct subject, perspective or stream may depend on memory, integration, valence, embodiment and boundary-forming organisation. Replacement experiments can study those conditions of manifestation. They do not reveal awareness-in-itself as another device variable.

Part IV. Matched behaviour can hide unmatched causation

Suppose a recurrent network classifies a finite set of stimulus sequences. For a fixed horizon, its computation can be unfolded into a feed-forward graph with a separate copy of the relevant operation at each time step. The two graphs can generate the same answer for every registered sequence. They do not have the same physical recurrence because the feed-forward graph does not reuse a state through a feedback path.

The unfolding argument turns this construction against theories that identify consciousness with causal structure. If a recurrent and feed-forward system are behaviourally indistinguishable while a theory assigns consciousness to one and not the other, ordinary behavioural experiments cannot choose between them. The argument presses directly on IIT and recurrent-processing accounts.

A published response disputes the strength of the equivalence. Universal approximation supports static input-output approximation, the response argues, not identity of temporal patterns or reactions to structural perturbation. Recurrent top-down and bottom-up interaction may support flexible control that an output-matched feed-forward model lacks once the test moves beyond the original map.

Both sides reveal something important. The unfolding construction is devastating against a theory that treats unobserved recurrence as consciousness-making while permitting no consequence under any accessible intervention. The response is decisive against using a narrow task map as proof of robust equivalence. The open question is whether a richer intervention family can connect the disputed causal property to a consciousness-linked contrast without defining the answer in advance.

A substrate experiment should compare intervention fingerprints, not screenshots of successful behaviour. Baseline equality is the starting condition. Identification comes from what each system does when the evaluator perturbs a state, cuts a path, stretches time, changes noise, forces online learning or alters metabolic support.

The six-probe battery

1. Ordinary-task probes establish E1 competence across held-out inputs and shifts. They include report, action and confidence but do not expose internal organisation.

2. State interventions set a candidate internal variable to a value that ordinary inputs would not produce. Downstream state, report and recovery are compared with the reference circuit.

3. Structural cuts remove or reroute one causal edge. A causal twin should reproduce the reference circuit’s selective deficit, not only its normal output.

4. Temporal probes delay, accelerate or jitter a local update while holding the external sequence fixed. They test whether rate, phase and feedback latency belong to the preserved role.

5. Plasticity probes introduce a novel contingency that must be learned online. The evaluator compares learning curves, transfer, catastrophic interference and recovery after reversal.

6. Multiscale probes alter ion concentration, noise spectrum, metabolic budget or bodily feedback when those variables are part of the theory. A software module is not penalised for lacking an irrelevant property; the theory must explain why the property belongs in the test.

An executable equivalence audit

The following artefact treats equivalence as a graded release claim. Its values are illustrative fixtures, not measurements from a person or model. Every grade requires all earlier probe families to pass. A candidate cannot claim E3 because its timing matches while its hidden-state response fails.

from dataclasses import dataclass

PROBE_ORDER = (
    "trace",
    "task",
    "state_intervention",
    "timing_and_learning",
    "multiscale",
    "organismic",
)

@dataclass(frozen=True)
class Profile:
    name: str
    responses: dict[str, tuple[float, ...]]

def within(reference: tuple[float, ...], candidate: tuple[float, ...], tolerance: float) -> bool:
    return (
        bool(reference)
        and len(reference) == len(candidate)
        and max(abs(a - b) for a, b in zip(reference, candidate)) <= tolerance
    )

def equivalence_grade(reference: Profile, candidate: Profile, tolerance: float = 0.05) -> str | None:
    if tolerance < 0:
        raise ValueError("tolerance must be non-negative")
    passed = None
    for grade, probe in enumerate(PROBE_ORDER):
        if probe not in reference.responses or probe not in candidate.responses:
            break
        if not within(reference.responses[probe], candidate.responses[probe], tolerance):
            break
        passed = f"E{grade}"
    return passed

reference = Profile("biological reference", {
    "trace": (0.88, 0.76),
    "task": (0.91, 0.84),
    "state_intervention": (0.73, 0.22),
    "timing_and_learning": (0.66, 0.79),
    "multiscale": (0.58, 0.81),
    "organismic": (0.77, 0.69),
})

r1 = Profile("output surrogate", reference.responses | {"state_intervention": (0.31, 0.64)})
r2 = Profile("causal twin", reference.responses | {"multiscale": (0.29, 0.52)})
r3 = Profile("biohybrid twin", reference.responses | {"organismic": (0.52, 0.47)})

assert equivalence_grade(reference, r1) == "E1"
assert equivalence_grade(reference, r2) == "E3"
assert equivalence_grade(reference, r3) == "E4"
assert equivalence_grade(Profile("empty", {}), Profile("empty", {})) is None

The audit prevents vocabulary from outrunning evidence. A candidate with no trace evidence receives no grade, and every later grade requires all earlier probes. The audit does not decide which grade consciousness requires. That decision belongs to a theory, and the theory must accept the empirical consequences of its chosen grade.

An implementation can score differently on different axes. A neuromorphic circuit may match spike timing and local plasticity while lacking hormonal regulation. A bioreactor may reproduce continuous chemistry while having no task-level competence. A digital twin may predict the organism accurately while its own hardware instantiates none of the simulated causal relations at the relevant grain.

This last case motivates a sharper distinction between description and realisation. A weather model represents convection, but its processors do not become wet or windy. Consciousness could resemble weather if its constitutive property is chemical or field-based. It could instead resemble computation if the relevant causal organisation is physically present in the processor, even though the materials differ. The analogy cannot decide which property consciousness is.

Part V. Replacement theatre, preservation lines and protocol

Thought experiment: mira and the three silent swaps

Mira’s target circuit stabilises a visual object across eye movements. Damage leaves her able to detect flashes but unable to preserve object identity when her gaze shifts. Engineers prepare three pin-compatible replacements. Every module receives the same afferent signals and can stimulate the same downstream neurons. The surgery is staged, reversible and performed only in this imaginary protocol.

Module R1 is a behavioural surrogate. It was trained on millions of recorded trajectories from Mira’s circuit and produces the expected downstream pattern for familiar input histories. Its internal variables do not correspond to the replaced circuit. If a hidden state is perturbed, it returns to a memorised attractor or fails outside training support.

Module R2 is a causal twin. It reproduces the original circuit’s recurrent topology at a registered grain. Engineers can intervene on a named artificial state and obtain the same conditional distribution over later local states, downstream signals and recovery time as in the biological circuit. It also matches online learning, timing jitter and response to partial lesions. The physical material remains non-biological.

Module R3 is a biohybrid twin. Ion-electronic elements, living support cells and chemical modulation reproduce the E4 profile. Local energy scarcity changes excitability and plasticity as it did in the original tissue. The module remains engineered and lacks Mira’s developmental history, but it participates in organismic regulation more deeply than R2.

Each module is inserted on separate registered trials. Mira performs the same perceptual tasks and reports a continuous visual world. Clinicians also measure confidence, reaction time, adaptation, surprise, eye movement, neural broadcast, sleep-dependent consolidation and targeted responses to hidden perturbations. Nobody is allowed to infer preservation from the first fluent sentence.

The experiment is designed so that each successful replacement defeats one explanation and leaves another alive. R1 success weakens claims that the original microanatomy is needed for the tested task. R2 success weakens claims that the biological material is needed for the registered causal dynamics. R3 success tests whether adding biological and multiscale features repairs any deficit left by R2. None of the three, by itself, provides direct access to Mira’s phenomenal continuity.

What each theory should predict before the switch

The final row is uncomfortable. If all accessible measures match while a theory predicts absent or altered experience, the theory cannot borrow those same measures as confirmation. This is the force of dancing qualia. It does not logically disprove substrate-specific experience. It shows the epistemic price of placing a phenomenal difference beyond every functional consequence.

Mira could still be wrong about her continuity. Human memory is reconstructive, reports can be trained and a local replacement might alter an experience too subtly to notice. The protocol therefore includes forced-choice discrimination, confidence calibration, surprise, cross-modal effects, delayed recall and reports collected before downstream behaviour reveals the intervention. These controls narrow ordinary error. They do not create a theory-neutral meter for qualia.

What real prostheses already tell us

Neural devices establish that non-biological components can enter cognitive and perceptual loops. They do not yet approach Mira’s substitution. A human hippocampal neural prosthetic study derived stimulation patterns from each participant’s hippocampal activity and applied them during memory encoding. The study involved people already implanted for epilepsy monitoring. Some subject-category combinations improved and others worsened, with stronger positive patterns in a subgroup with impaired memory. The authors also reported limited trials, low power for individual combinations and imperfect category specificity.

A prosthesis that changes memory performance shows causal entry into a memory circuit, not replacement of a conscious subject. The biological hippocampus remained present, the device stimulated selected sites, and the outcome was task performance. This is still valuable. It demonstrates the first rungs of a replacement science: personalised modelling, model-derived timed stimulation, reversible comparison and heterogeneous effects.

Three preservation lines

The substrate debate is often drawn as a switch: computation on one side, biology on the other. That picture omits a large middle. A theory can allow multiple materials yet reject a lookup table, a replay engine or a simulation whose causal organisation exists only in the represented model.

Software-level functionalism

The strongest version is not the claim that matching words is enough. It is the principle that phenomenal properties are invariant under preservation of the right fine-grained functional organisation. David Chalmers’s fading and dancing qualia argument asks us to replace neural circuits gradually with silicon circuits that preserve local causal organisation. If experience faded or changed while cognition remained fixed, the patient’s judgements about experience could become radically detached from the experience itself. Chalmers treats that dissociation as empirically implausible and argues for organisational invariance.

This position gives behaviour genuine weight because report and discrimination are parts of the organisation. It also permits non-biological consciousness in principle. The hard work lies in “the right” and “fine-grained”. A replacement that matches the patient’s prepared answers but cannot learn, notice errors or react to internal disruption has not preserved the intended role.

Structure-sensitive computationalism

This family locates the invariant in physically realised causal dynamics. Different materials may qualify, but only if their components have the relevant powers under intervention. A stored film of a recurrent network’s outputs does not become that network. A processor running a model may qualify only if the consciousness-relevant transition structure exists in the processor at a defensible physical grain, rather than only inside a modeller’s interpretation.

Integrated Information Theory 4.0 takes a particularly demanding version of this route. It evaluates intrinsic cause-effect power in a candidate physical substrate. Two systems can share an input-output function yet differ in their predicted experience because their internal causal structures differ. On IIT’s account, a purely feed-forward system has no irreducible cause-effect structure of the required kind, while recurrence can support one. Other structure-sensitive views need not accept IIT’s axioms or measure. They share the insistence that actual causal topology matters.

Biological naturalism and biological computationalism

Biological naturalism begins from the fact that every uncontroversial conscious system is living and that brain function depends on electrochemical, metabolic, developmental and bodily processes omitted from standard software descriptions. Its strongest scientific form names the contribution those processes make. A published review proposing biological computationalism points to scale-inseparable multiscale processing and continuous-valued computation in a fluid biological substrate. Those are specific candidate capacities, not an appeal to carbon by prestige.

A stricter version says biology has consciousness-making causal powers even if a non-biological duplicate preserves every information-processing capacity. That position may be coherent. It is difficult to support empirically because every ordinary observation remains identical by construction. Part II separated this Type A claim from the empirically engaged Type B programme.

The three positions disagree less about observable competence than about property placement. Software-level functionalism places the relevant property in an abstract but demanding role. Structure-sensitive computationalism places it in substrate-general causal organisation. Biological naturalism places at least part of it in processes characteristic of living systems.

The table is not a ranking. It exposes different burdens. Functionalism must specify the role finely enough to exclude mimicry without quietly rebuilding biology into the definition. Structure-sensitive accounts must select a physical grain without arbitrary observer choice. Biological accounts must connect a distinctive living process to a contrast in consciousness, not simply to cognition or survival.

A substrate-substitution protocol

A serious replacement programme begins before the artificial module exists. The reference circuit must first be characterised across the deepest feasible intervention grade. Otherwise the replacement inherits an easy specification and later success is mistaken for preservation of properties that were never measured.

The protocol has seven records. The first names the target phenomenon, such as object continuity across saccades, and keeps phenomenal continuity as a separate claim. The second defines the physical boundary and every external dependency. The third records the equivalence tuple E(O, I, H, ε). The fourth lists each theory’s preservation line before results are visible. The fifth specifies crossover, sham, washout and stopping rules. The sixth preserves every state, intervention and report in a versioned manifest. The seventh restricts release language to the grade actually passed.

Replacement should be gradual only when graduality itself is controlled. Slow substitution allows learning and plasticity to conceal a mismatch. Fast switching creates shock, inflammation and timing artefacts. Use both. Alternate short blinded switches with longer adaptation periods, and return to the reference circuit often enough to estimate drift. A module that succeeds only after the remaining brain learns around it is a useful prosthesis, but its success does not show immediate causal equivalence.

Worked scenario: the phase-jitter trial

The worked scenario begins after Mira’s original circuit has passed six weeks of mapping. On ordinary trials, the circuit binds a target across two eye movements and updates confidence after a masked distractor. On registered perturbation trials, a brief phase delay is introduced into one recurrent path after the first eye movement. The biological circuit shows a selective identity error, a confidence drop and recovery on the next cycle. A matched sham pulse changes local energy by the same amount without delaying the path and does not produce that profile.

R1 reproduces the ordinary answer and confidence. Under the hidden phase delay, its output barely changes because the internal variable is not part of its learned surrogate. It therefore remains E1. This is not a failed prosthesis verdict. It may still be clinically useful on the ordinary task. It is a failed claim of recurrent causal equivalence.

R2 reproduces the ordinary trial, the selective error and the one-cycle recovery. When the relevant edge is cut, both the reference and R2 lose cross-saccadic identity while retaining flash detection. When update timing is stretched gradually, their confidence curves change together. R2 earns E3 for this protocol. The result supports continuity of the tested causal organisation across biological and neuromorphic material. It does not establish that E3 is the correct phenomenal preservation line.

R3 also passes E3. Under local energy restriction, however, it reproduces the reference circuit’s selective prioritisation of the central target while R2 degrades uniformly. Ion-sensitive modulation restores the original pattern. R3 earns E4. If a registered consciousness-linked contrast changed with that metabolic manipulation while ordinary performance stayed matched, the biological-computational account would gain distinctive support. If only total accuracy fell, the result would show resource dependence without isolating consciousness.

Mira reports continuity throughout all three modules. The report is evidence, especially when collected before the intervention affects visible action. It is weighed beside forced choice, confidence, flexible control, physiology and intervention response. The final record states: R1 preserves the registered task map; R2 preserves the tested recurrent dynamics; R3 additionally preserves the tested multiscale response; phenomenal continuity remains theory-dependent.

Release and ethics

No irreversible human replacement should be justified by this thought experiment. Near-term work belongs in simulations, cultured networks, animals where ethically warranted, existing therapeutic interfaces and reversible assistive devices. Human studies require therapeutic purpose, independent oversight, informed consent that distinguishes function from identity, and stopping rules for subtle changes in perception, agency or affect.

The ethical asymmetry cuts both ways. Treating fluent continuity reports as conclusive could overlook harm. Treating every non-biological intervention as consciousness-destroying could deny beneficial prostheses or misdescribe a person whose cognitive life continues. The safest language is typed, reversible and explicit about uncertainty.

Source trail

The philosophical base uses the Stanford Encyclopedia entries on functionalism, multiple realizability, Śaṅkara and Aristotle’s psychology, together with Chalmers’s original fading and dancing qualia paper. These sources disagree about what organisation can establish and are used to expose commitments rather than manufacture a consensus.

The causal-structure dispute uses the published unfolding argument, its published recurrence-focused response and IIT 4.0. The structure-sensitive middle also draws cautiously on the recent Intrinsic Computational Functionalism preprint, which is identified as a preprint because its proposed criteria remain under active scrutiny.

The biological section uses the published review of biological computationalism, the published studies on human dendritic computation, an organic electrochemical neuron and a human hippocampal prosthetic. The Type A and Type B distinction comes from a recent preprint on biology and conscious AI and is treated as a methodological proposal, not settled doctrine.

Glossary

The decision this changes

Do not ask a supplier, researcher or model card whether an artificial system is “substrate independent”. Ask which grade of equivalence has been demonstrated, against which reference, under which hidden interventions, at what physical grain and for which claimed phenomenon. Require the answer to separate task continuity, causal continuity, dynamical continuity, biological continuity and phenomenal continuity.

The publishable conclusion should be narrower than the ambition and stronger than a disclaimer: this replacement preserves these registered roles and counterfactuals at this grain. If biological features add a discriminating effect, name them. If phenomenal continuity depends on a further premise, show the premise. The field advances when the preservation line is exposed before anyone crosses it.