The theatre of replacement

Mira’s case is not a prediction about neurosurgery. It is a measuring instrument. By holding visible performance steady while changing the realiser, it forces three positions to expose what they regard as constitutive. The positions are often described as if one side worships software and the other worships cells. Their serious versions are more precise.

Three positions, three preservation lines

Philosophical positionSoftware functionalism says that a mental state is fixed by the right pattern of causal roles. Pain, for example, is not identified by a particular molecule but by how the state is caused, what it does to other internal states and how it guides action. The strongest version does not require a literal copy of one brain. It requires a functionally isomorphic organisation at the grain relevant to mentality. Chalmers’s fading and dancing qualia arguments press this intuition through gradual neural replacement, while acknowledging that such arguments increase plausibility rather than deliver a deductive proof.[1]

Practitioner synthesisStructure-sensitive computation accepts multiple realisation but rejects task-level equivalence as too weak. A physical system computes through a mechanism: organised components, state transitions and rule-governed operations realised in matter.[4] Two systems can implement the same input-output map while differing in recurrence, latency, synchronisation, intrinsic versus externally orchestrated state, counterfactual response or capacity to adapt. On this view, silicon is not disqualified. A replay table is.

Philosophical positionBiological naturalism claims that consciousness depends on causal powers of biological systems that a program description alone does not supply. Searle’s formulation is more than that computers behave badly. It is that syntax is insufficient for intentionality, and an artificial system would need to duplicate the relevant causal powers of brains rather than only run the right program.[3] A defensible biological view must therefore say which powers matter, how they enter conscious processing and what intervention would reveal their absence.

The hidden assumption inside “nothing changed”

Mira’s report is not worthless. First-person testimony is central evidence in ordinary human cases. The mistake is to treat the report as self-interpreting. It can be generated through different routes: preserved experience, preserved access with altered quality, trained expectation, limited discriminability, delayed effects or a module designed to reproduce the report. The same concern applies to reaction time and accuracy. A match on a measurement does not tell us which internal relation generated it.

Block’s objections to simple functionalism and Searle’s objections to program sufficiency exploit this gap in different ways.[2][3] Chalmers responds by making organisation much richer than a list of outputs and by asking whether an unnoticed phenomenal switch could remain coherently embedded in cognition.[1] The useful lesson is not that intuition has selected a winner. It is that the comparison must move below the coarse behaviour that all sides were allowed to stipulate.

flowchart LR
  A["Three silent replacements in<br/>Mira’s brain"] --> B["The theatre of replacement"]
  B --> C["“Same function” is a ladder"]
  C --> D["Where material enters the<br/>mechanism"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

“Same function” is a ladder

Imagine two black boxes. For every input tested today, both return the same output. At the shallowest grain they are equivalent. Now open a side port and set an internal variable to a value never reached during normal operation. One box recovers; the other oscillates. Slow both clocks by ten per cent. One preserves the ordering of evidence and action; the other times out and commits before its correction arrives. Give both a surprising sequence that changes how they learn. One adapts; the other replays a trajectory fitted yesterday.

Nothing contradictory has happened. Equivalence is always relative to a question. The term becomes useful only when it names the family of observations and counterfactuals that must agree.

An equivalence tuple, not an adjective

Let system A be the reference circuit and system B the replacement. Define a comparison by the tuple:

O is the set of observables, such as choices, reports, internal readouts or physiological responses. I is the allowed intervention family, including input changes, state clamps, lesions, timing shifts and learning-history manipulations. H is the time horizon over which agreement must persist. ε is the tolerated divergence under a declared distance measure. “Functionally equivalent” without these four fields is not false. It is unfinished.

For a state correspondence h, a stronger requirement asks whether transitions commute:

Here x is a possible state of the reference, u is an input or intervention, FA and FB are the systems’ transition rules, h maps reference states to candidate states and d measures the resulting mismatch. The equation says: move the reference and then translate its state, or translate first and move the replacement. The two routes should meet within tolerance. This condition is valid only over the registered state and intervention domain. It says nothing about untested states, longer horizons or experience.

The equivalence-grade protocol

The tuple becomes operational through six grades. They are not a consciousness scale, and a higher number does not mean “more conscious”. Each grade is a certificate for a deeper preservation claim. A system may pass one grade and fail the next. The first failure is more informative than an average score because it names the first causal property the replacement does not preserve.

The grades are deliberately asymmetric. Failing E2 defeats a claim of state-level preservation. Passing E2 does not prove E3 because a candidate can reproduce intervention snapshots while carrying time differently. Passing E4 still does not force E5 because theories disagree about which physical or functional properties constitute experience. This is not bureaucratic caution. It is what prevents evidence from migrating across distinct questions.

Why the ladder is stricter than benchmark parity

A benchmark samples trajectories chosen by its designers. A replacement can exploit that narrowness. It may cache responses, use an external clock, depend on a controller outside the declared boundary or reproduce the correct output through a route that fails under lesion and repair. This is the implementation problem in experimental form. A mapping from physical states to abstract states is cheap unless it respects the mechanism’s actual components, temporal organisation and counterfactual dependencies.

Structure-sensitive computation does not demand microscopic identity. That would make every temperature fluctuation disqualifying. It demands a principled grain: coarse enough to ignore irrelevant material variation, fine enough to preserve the dependencies that make the target capacity work. The right grain is discovered by interventions that change the claimed capacity, not selected because it makes the preferred replacement pass.

Where material enters the mechanism

Calling a process computational does not make its material irrelevant. A bridge can implement a traffic pattern only because steel, concrete and geometry sustain the relevant loads. A neuron can implement a state transition only because membrane conductances, dendrites, synapses, metabolism and surrounding tissue sustain particular dependencies. The scientific question is which of those dependencies are indispensable for the target phenomenon and which are replaceable implementation details.

Biology may matter in two different ways

Practitioner synthesisAn intrinsic-biology claim says that being alive, cellular or of a particular biological kind is itself necessary, even if no unique information-processing contribution has yet been identified. This is a coherent metaphysical commitment, but it is experimentally difficult because a functional match cannot confirm it and a failure can usually be attributed to an imperfect replacement.

A capacity-linked biology claim names a property that biology supplies: rapid synaptic change, dendritic nonlinearity, ionic and chemical signalling, metabolic regulation, developmental history, or closed coupling to a living body. Human cortical dendrites, for example, exhibit active electrical events and nonlinear operations that are not captured by the textbook point-neuron picture.[5] This does not establish that dendritic spikes constitute consciousness. It establishes a concrete property that a proposed neural replacement may need to preserve if the target capacity depends on it.

Gradual replacement is informative, but not self-interpreting

Gradual replacement thought experiments gain force by removing dramatic discontinuities. If one neuron is replaced while behaviour, memory and report remain stable, it feels arbitrary to declare that experience vanished at that step. Chalmers uses fading and dancing qualia to argue that large unnoticed changes in experience would create implausible dissociations within cognition.[1] The argument is serious because it asks what a phenomenal change could do while every relevant cognitive relation stays fixed.

Yet “every relevant relation” is exactly what the experiment stipulates rather than measures. If replacement preserves only the current behavioural path, hidden differences may accumulate. If it preserves the complete fine-grained causal organisation, the conclusion is stronger, but then the decisive work was done by the organisation clause. Graduality removes a sharp boundary. It does not tell us which variables belong inside the preserved organisation.

A real prosthesis preserves a loop, not a whole person

Existing neuroprostheses show why boundary discipline matters. A cochlear implant converts sound into electrical stimulation of surviving auditory pathways, bypassing damaged inner-ear structures.[10] It demonstrates that useful auditory function can cross a change in realiser. It does not establish that every feature of normal hearing is preserved, nor that arbitrary neural tissue can be replaced by the same strategy.

Human hippocampal prosthesis studies have used recorded neural activity and patterned stimulation to facilitate specific memory encoding and recall tasks in participants already undergoing intracranial monitoring.[11] Later work targeted features and categories, with effects varying across content and participants.[12] These are important demonstrations of a hybrid loop. They are not replacements of the hippocampus, tests of personal identity or evidence that memory’s phenomenal character is substrate-independent.

A different ontology changes the bridge, not the need for a protocol

Physicalism and emergentism usually ask which physical organisation produces or constitutes experience. Functionalism locates the decisive organisation in causal roles. Biological naturalism restricts the relevant powers more tightly. Idealism may treat awareness as primary and a brain or machine as a condition that structures a local perspective. Neutral monism and dual-aspect views treat mental and physical descriptions as aspects of a more basic reality. Panpsychism assigns experiential or proto-experiential properties broadly, then faces a combination problem. Process philosophy treats enduring subjects as stabilised patterns of activity rather than static substances.

None of these positions is established by replacement performance. Yet they can still generate different preservation requirements. A consciousness-primary view might say that substrate does not produce awareness but constrains what can manifest, integrate or remember. It must then explain why R1, R2 or R3 shapes the perspective differently, and what observable change follows. Changing the metaphysical verb from production to manifestation does not remove the causal burden.

stateDiagram-v2
  [*] --> Claim
  state "Three positions, three<br/>preservation lines" as Claim
  state "The hidden assumption inside<br/>“nothing changed”" as Test
  state "An equivalence tuple, not an<br/>adjective" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Make the disagreement experimental

The hardest replacement arguments often stall because the candidates are designed to agree on the very data used to judge them. The unfolding argument makes this pressure explicit. A recurrent network can, for a fixed finite task, be represented by a feed-forward construction that reproduces the same input-output behaviour. If a consciousness theory says recurrence matters while every report-based experiment can be matched by a non-recurrent system, the theory appears unable to connect its preferred structure to discriminating evidence.[6]

The challenge is powerful against vague appeals to internal topology. It is not a universal proof that topology never matters. Replies have questioned whether reports exhaust the admissible evidence and whether recurrent and unfolded systems remain equivalent under interventions on their internal organisation.[7] A more recent mathematical result identifies a concrete boundary: when recurrent weights change rapidly with history, a single static feed-forward unfolding need not preserve the function. Plasticity, temporal encoding and resource growth can restore discriminating observations, although this does not show that recurrence or plasticity is necessary for consciousness.[8]

Use adversarial predictions, not theory-themed measurements

A substrate experiment should begin with at least two rival models that make different predictions under the same intervention. Otherwise, investigators can choose measurements that flatter their preferred account. The large 2025 adversarial collaboration comparing integrated information theory and global neuronal workspace theory illustrates the discipline. Proponents and theory-neutral researchers preregistered divergent predictions, pass and fail criteria, and interpretations before collecting fMRI, MEG and intracranial EEG data from 256 participants. The results aligned with some predictions while challenging key commitments of both theories.[9]

Integrated information theory offers one explicit structure-sensitive proposal. IIT 4.0 treats a candidate substrate in terms of intrinsic cause-effect power, specificity, integration, exclusion and composition, assessed through interventions on its units.[13] It therefore predicts that systems with matching external functions can differ if their intrinsic causal structures differ. This is a clear preservation line. The difficulty is practical and theoretical: exact analysis becomes infeasible for realistic systems, the choice of units and timescale matters, and the proposed identity between Φ-structure and experience remains contested.

Theory-derived indicator methods for artificial systems take a broader route. They extract properties from several consciousness theories and assess whether a configured AI system instantiates them, while treating indicators as evidence rather than verdicts.[14] The equivalence-grade protocol complements that approach. Indicators ask which theoretically relevant properties are present. Replacement grades ask which of those properties survive a controlled change of realiser.

A six-probe battery for substrate substitution

The following battery turns Mira’s theatre into a research protocol. Each probe changes one causal property and records both the match and the first divergence. The probes should be preregistered, applied to a family of candidate replacements and interpreted against explicit rivals.

1. Task and transfer probe. Match normal performance, then shift input statistics, nuisance variables and task composition. This separates trace fitting from a portable task map.
2. State intervention probe. Clamp, swap, ablate and restore mapped internal states. Test downstream effects and recovery, not only immediate output.
3. Temporal probe. Change delay, phase, update order and interaction speed. Record when a state graph ceases to be adequate because coupled clocks matter.
4. Plasticity and history probe. Give an event that changes the reference online, then test future behaviour before any candidate recompilation. Reverse histories that end at the same visible state.
5. Multiscale coupling probe. Vary regulatory inputs, energetic constraints, chemical context and body-linked signals that the target theory says are relevant.
6. Boundary probe. Remove external controllers, lookup services or hidden experimenter support. Determine whether the candidate itself has the claimed organisation or merely participates in a larger mechanism that does.

Executable lab: a plastic loop and its compiled twin

The lab below builds a small recurrent reference whose state and weight change with experience. It then compiles a lookup twin for every binary input prefix up to a chosen horizon. On the unperturbed calibration set, the twin matches exactly. A hidden-state perturbation is then applied only to the reference, and the input sequence is extended beyond the compiled horizon.

What it tests: whether trace equivalence survives a state intervention and online adaptation. Assumptions: binary inputs, deterministic updates, a finite compiled horizon and a simple Hebbian-like weight rule. How to use it: change the horizon, plasticity rate and perturbation size, then run the audit. A positive result: low error under a registered probe permits a preservation claim only for that probe and tolerance. A negative result: the compiled twin lacks a causal property required by the test. What it cannot establish: that the reference is conscious, that plasticity is necessary for consciousness or that a more capable replacement could not pass.

class PlasticLoop {
  constructor(eta, state = 0.15, weight = 0.8) {
    this.eta = eta;
    this.state = state;
    this.weight = weight;
  }
  step(input) {
    const output = Math.tanh(this.weight * input + 0.72 * this.state);
    this.weight = clamp(this.weight + this.eta * input * output * 0.25, -1.5, 1.5);
    this.state = output;
    return output;
  }
}

function compileTwin(horizon, eta) {
  const table = new Map();
  for (let length = 1; length <= horizon; length += 1) {
    for (let mask = 0; mask < 2 ** length; mask += 1) {
      const system = new PlasticLoop(eta);
      let prefix = "";
      for (let t = 0; t < length; t += 1) {
        const bit = (mask >> (length - t - 1)) & 1;
        prefix += bit;
        table.set(prefix, system.step(bit ? 1 : -1));
      }
    }
  }
  return inputHistory => table.get(inputHistory.join(""));
}

// Baseline traces match because the twin contains every prefix through H.
// A hidden-state perturbation changes the reference but not the lookup twin.
// Beyond H, the twin has no entry. That is a declared negative result.

The substrate-substitution protocol

A complete study can now be written as a compact commitment packet:

1. Name one target property. Do not test “mind” in general. Choose conscious access, perceptual integration, memory continuity, metacognitive control or another defined phenomenon.
2. Declare the candidate boundary. Include power, clocks, external controllers, sensors, body loops, experimenter actions and persistent state when they are causally required.
3. Register the preservation line. Specify O, I, H, ε and the proposed state correspondence before fitting the replacement.
4. Build a candidate family. Compare at least a trace surrogate, a causal twin and a richer candidate. Match normal performance so easy differences do not decide the test.
5. Predeclare rival predictions. State what would strengthen software functionalism, structure-sensitive computation and the specific biological hypothesis. Include a result that lowers confidence in each.
6. Run probes in increasing depth. Preserve raw trajectories, intervention timing, state maps, support services and all failed restorations.
7. Issue a typed conclusion. “E3 over this domain” is acceptable. “Substrate independent” is not. State what the result permits, defeats and leaves open.
8. Apply reversible ethics. Use simulation, cultured systems, existing therapeutic interfaces and reversible assistive studies before any irreversible human substitution. Separate functional benefit, identity risk and consciousness uncertainty in consent.