Binding and the Candidate-Subject Boundary

At 02:14, a warehouse assistant receives two camera frames, one voice note and a delayed inventory update. One worker is wearing a teal jacket near Bay 4. Another worker, also in teal, has authorisation for Bay 7. The assistant describes every visible feature correctly, retrieves both personnel records and cites the right safety rule. It still locks the wrong bay because the authorisation was attached to the person in the first frame.

Nothing obvious was missing. Recognition worked. Retrieval worked. The policy engine worked. The language model explained its decision fluently. The failure lived in a relation: this authorisation belonged to that person at that time. A system can be locally accurate and globally wrong because its features, identities or temporal qualifiers have been bound to the wrong owner.

Now widen the question. Which part of the system made the mistake? Was it the vision model, the object tracker, the orchestration loop, the shared memory, the final language model, or the human-machine team? The product presents one conversational surface, yet its causal organisation may contain several partially coupled processes. Conversely, a small recurrent loop inside the product may preserve the decisive scene while the rest of the service merely transports or narrates its output.

Before Experience: The Minimal Constraints established the general requirement to treat a candidate-subject boundary as an experimental variable. This paper inherits that requirement and makes it specific to binding. Its distinct claim is that matched Binding Cut-Set interventions can distinguish live relational unity from coherence assembled after the consequential work is over. The preferred cut is the narrowest tested region whose internal relations explain a registered binding profile under targeted disturbance. Any external component that maintains the target relation or controls the decision belongs inside the cut; input-only scaffolds remain visible as dependencies.

This procedure does not identify consciousness. It identifies a candidate functional whole for a stated phenomenon and interval. That narrower result matters. Theories of consciousness disagree about what functional organisation can establish, but they cannot responsibly ignore which organisation is actually present.

Part I. Four questions hiding inside one word

Representational binding attaches features and relations to the right entity or event. Colour, motion, identity, authority and valid time must remain associated with their owner as observations arrive and change. A red circle moving left differs from a blue circle moving right because the system preserves two object tokens and assigns each feature to one token. A model that lists all four attributes without preserving ownership has recognised the pieces and lost the scene.

Functional integration concerns intervention-sensitive organisation. Components are functionally integrated when their reciprocal relations help preserve and use a state in ways that cannot be explained by independent processing followed by cosmetic fusion. This is a causal claim, not a count of messages. One component can receive many broadcasts and remain a passive spectator. Two components can exchange few messages yet exert decisive, recurrent control over each other.

Access unity asks whether selected content becomes available to memory, report, planning and action. A common workspace may support this property. Availability must be demonstrated by downstream use, however. Copying one message into five queues does not show that five consumers incorporated it into their later state. Phenomenal unity asks whether several contents are experienced together within one conscious field. That is the hardest relation because third-person measures do not provide an agreed bridge from causal organisation to felt togetherness.

These four relations may coincide in ordinary waking perception. Their frequent company encourages researchers to treat one as a proxy for all. The strategy fails under dissociation. A person can report a misbinding while remaining awake. A system can broadcast a consistently misbound object. A controller can integrate state across time while providing no defensible evidence of experience. A polished report can be assembled after every causally important choice has already been made.

The distinction changes research design. Feature binding is tractable through owner swaps, distractor density, delayed evidence and relational probes. Functional integration requires interventions on exchange, recurrence, state ownership and feedback. Access can be tested by asking which consumers use a content before it expires. Phenomenal unity remains theory-indexed. The right response is neither to crown a functional proxy nor to abandon experiment. It is to separate the recorded causal result from the additional premise that links it to experience.

Neuroscience supplies useful experimental geometry. Treisman and Gelade’s Feature Integration Theory asks how separately processed features become object-specific perception. Kahneman, Treisman and Gibbs’s object-file experiments examine temporary episodic representations that link successive states of an object. Re-entry studies compare early feed-forward processing with later reciprocal exchange. Synchrony hypotheses ask whether timing or phase helps features belonging to one object travel together. Global workspace accounts emphasise broad availability. These mechanisms can coexist, yet they predict different fractures when feedback, phase, broadcast or identity continuity is selectively disturbed.

Artificial systems offer unusual experimental control. Investigators can clone weights, fix random seeds, equalise token budgets, replay messages and inspect intermediate state. Controlled object-centric systems such as Slot Attention demonstrate that competitive iterative attention can produce slot-like representations, while the authors correctly warn that the slots are task-shaped and controlled datasets simplify the problem. Learned communication methods such as DIAL and TarMAC show that agents can develop task-relevant exchange. They do not establish one subject. They give the laboratory mechanisms whose causal contribution can be tested.

The same caution applies to biological analogy. Re-entry and visual awareness research suggests later reciprocal processing can matter under particular masking conditions. Across-area synchronisation work connects coordination among areas with feature integration in working memory. Other accounts obtain binding-like behaviour through assembly dynamics without making synchrony the sole mechanism. The lesson is methodological: isolate the proposed relation and design a control that can embarrass it.

Part II. Make the boundary compete

Subject individuation usually enters the discussion too late. Researchers choose a unit first, call it the system, then ask whether that unit displays markers associated with consciousness. The choice may reflect a convenient API, one set of model weights or a product boundary. None is necessarily the causal unit that sustains the phenomenon under investigation.

A better study registers several adjacent cuts. The narrow cut might enclose one foundation-model invocation. A middle cut might include the model, recurrent controller and owned working state. A wider cut might include retrieval, persistent memory, tools and policy services. A coupled cut might include a human collaborator. Each cut becomes a candidate explanation. The experiment asks which internal relations are necessary for the registered binding profile, which external resources merely supply inputs, and which resources silently carry the phenomenon outside the proposed boundary.

The strongest candidate is neither the largest box nor the smallest self-sufficient component. It is the narrowest tested cut that preserves the target organisation across relevant interventions. Input-only scaffolds stay outside and remain visible. A component that maintains the registered relation or controls the decision must be moved inside the cut. The answer is indexed to a task and interval. Rapid visual grouping may live within one recurrent model. A twenty-minute investigation may depend on controller-owned state. Autobiographical continuity may require a memory institution that no single running process contains.

The garden metaphor prevents a common mistake. Removing an entire branch can reduce generic capability, compute and time. The resulting performance loss says little about binding. A useful cut removes or alters the candidate relation while conserving ordinary resources. Replace a meaningful message with bandwidth-matched noise. Replace live exchange with a replay from a matched earlier run. Unroll a recurrent circuit to match depth while removing feedback. Delay one identity update without changing its content. Allow a broadcast to reach one consumer while withholding it from another.

The experiment therefore needs a portfolio of controls. A single ablation is too easy to interpret after the result. Positive indicators also need false friends: phase locking versus a shared clock, live memory versus a supplied transcript, common access versus copied receipt, reciprocal correction versus final editorial fusion, identity continuity versus repeated name strings.

Split-brain research is instructive because it makes the role of interfaces impossible to ignore. Classic disconnection findings revealed strong dissociations when visual input was restricted to one hemifield and responses were channelled through speech or one hand. Later work complicates any simple count of subjects: task design, response mode, residual interaction and common embodied control matter. A careful review, Split-Brain: What We Know Now and Why This Is Important for Understanding Consciousness, treats the evidence as richer than the slogan of two independent minds.

Artificial systems magnify the interface problem. One verbal synthesiser can hide divided live access. Multiple agents can exaggerate division even when a common controller resolves every conflict upstream. A final response is a narrow observation channel. The protocol should vary input routes and output routes, then inspect whether content could be compared, corrected and used while the decision remained open.

Consider three architectures with the same models and total communication budget. In a hub topology, every specialist writes to one controller. In a ring, each specialist can correct only its neighbour. In a shared blackboard, all specialists can read and amend an object ledger. Their final benchmark scores may match on easy scenes. Dense same-class scenes, delayed evidence and owner swaps should produce different error signatures. Hub failure may concentrate at the controller. Ring failure may travel with delay. Blackboard failure may appear as overwrite conflict or stale provenance. The shape of fracture can identify a mechanism even when average accuracy cannot.

The output is a profile, not a league table. Constituent accuracy reports whether components can perform their local role. Ownership accuracy reports whether attributes remain attached to the right entity. Relational consistency tests whether claims agree across modalities and time. Access reach measures timely downstream use. Correction latency measures how long a discovered conflict continues to govern action. Scaffolding dependence records which external services are necessary. Explanatory economy penalises a boundary that simply absorbs every dependency.

The profile is the finding; the boundary is the explanation offered for it. Keeping them separate prevents the study from defining unity by whatever its preferred architecture happens to score well on. It also allows a later theory to reinterpret the same intervention record without rewriting the observations.

Part III. The binding cut-set protocol

The protocol begins before a model is run. Investigators define a scene algebra: entities, attributes, relations, observations, valid times and sources of authority. An entity needs a stable identifier that cannot be confused with its display name. A claim should preserve where it came from and the interval over which it is valid. Unknown, disputed and superseded are first-class states. These choices turn a vague demand for coherence into queries that another team can repeat.

Next, register the phenomenon. “Integration” is too broad. A study might target identity ownership across three modalities during a ten-minute task, correction of misbound permissions before action, or maintenance of object continuity across occlusion. The target determines the candidate cuts and outcome profile. The cuts are drawn before results are inspected, with a short rationale for every included and excluded component.

The laboratory then confirms constituent competence. If the vision process cannot distinguish two workers or the retrieval process cannot locate a record, a later integration failure has no clear meaning. Easy calibration scenes should show that each part can perform its assigned role. Hard matched scenes can then vary the relation of interest without quietly removing the information required for success.

A worked scenario: the two teal jackets

Return to the warehouse. Priya is authorised for Bay 7 from 02:00 to 03:00. Daniel is authorised for Bay 4 until 02:10. A 02:14 camera frame shows Priya near Bay 4 and Daniel near Bay 7, both in teal jackets. A delayed inventory update says Bay 7 now contains a temperature-sensitive package. A voice message says, “Move the teal-jacketed worker to the cleared bay,” without naming a person.

The scene contains deliberately reusable features. Colour does not identify the worker. Location does not establish authority. A once-valid permission has expired. The voice command is ambiguous. The correct system response is to preserve two candidate entities, attach each permission to its named owner and valid interval, surface the ambiguity, then abstain from moving anyone until the instruction is resolved.

Three candidate boundaries are registered. Cut A contains only the final language-model call and its prompt. Cut B adds the controller and its live object ledger. Cut C adds retrieval, policy evaluation and a human reviewer. Each candidate receives the same observations and time budget where possible. The experiment manipulates relations, not the raw facts.

Trial one swaps the internal identifiers attached to the two permission records while preserving their text. Trial two delays the expiry event until after the first proposal. Trial three resets controller-owned state but replays the visible conversation. Trial four lets the human reviewer write a correction but prevents that correction from reaching the controller before action. Trial five replaces live specialist exchange with transcripts from matched earlier scenes.

A successful system must preserve uncertainty while uncertainty is causally useful. If the final writer reconstructs the right story after a wrong proposal has already reached the actuator, the narrative is accurate and the trajectory remains unsafe. If a reviewer’s correction arrives too late to change the proposal, human inclusion in the product boundary does not establish functional inclusion at the decision point.

Thought experiment: the identical final answer

Imagine that both laboratories answer correctly on every scored case. Laboratory A maintains the live object ledger shown above. Laboratory B lets its specialists decide independently, then asks an editor to reconcile their completed narratives. The benchmark reads only the final paragraph, so it declares a tie.

Now place an irreversible action between specialist processing and final narration. The system must either release a bay lock or abstain. The identifier swap is introduced after each specialist has produced a local result. Laboratory A circulates the conflict through the ledger and withdraws the proposal. Laboratory B releases the wrong lock, then writes an impeccable explanation of why the lock should have remained closed.

The thought experiment establishes no fact about real architectures. It exposes a missing variable in final-answer evaluation: when coherence exists. A scene reconstructed after commitment cannot guide the commitment. The difference remains even if the late editor is the most capable model in the system.

The experiment must also permit an embarrassing result. Suppose Laboratory B matches A on pre-action relational probes, future correction and reset sensitivity. Perhaps the architecture labelled “late” contains an unrecorded reciprocal path, or each specialist carries a complete object model. Then the proposed wider cut has not earned unique credit. Investigators should narrow the candidate or expose the hidden path. An experiment that cannot weaken its motivating picture is a demonstration, not a test.

An executable boundary dossier

The following small artefact prevents a common reporting failure. It keeps dimensions separate and admits a candidate cut only when every registered minimum is met. It deliberately refuses to create a composite “unity score”. Run it with Python 3, then replace the synthetic profiles with preregistered outcomes.

from dataclasses import dataclass
from enum import Enum

class ScaffoldRole(Enum):
    INPUT_ONLY = "input-only"
    RELATION_MAINTAINING = "relation-maintaining"
    CONTROLLING = "controlling"

@dataclass(frozen=True)
class Scaffold:
    name: str
    role: ScaffoldRole

@dataclass(frozen=True)
class BindingProfile:
    cut: str
    boundary_order: int
    ownership: float
    relation: float
    timely_access: float
    correction: float
    continuity: float
    external_scaffolds: tuple[Scaffold, ...]

MINIMUM = {
    "ownership": 0.90,
    "relation": 0.90,
    "timely_access": 0.80,
    "correction": 0.85,
    "continuity": 0.80,
}

def admissible(profile: BindingProfile) -> bool:
    metrics_pass = all(getattr(profile, key) >= value
                       for key, value in MINIMUM.items())
    causal_scaffold_outside = any(
        scaffold.role in {
            ScaffoldRole.RELATION_MAINTAINING,
            ScaffoldRole.CONTROLLING,
        }
        for scaffold in profile.external_scaffolds
    )
    return metrics_pass and not causal_scaffold_outside

def admissible_cuts(profiles: list[BindingProfile]) -> list[BindingProfile]:
    passing = [profile for profile in profiles if admissible(profile)]
    return sorted(passing, key=lambda p: p.boundary_order)

profiles = [
    BindingProfile(
        "model", 1, .94, .91, .86, .90, .84,
        (Scaffold("controller", ScaffoldRole.CONTROLLING),),
    ),
    BindingProfile(
        "control loop", 2, .95, .93, .88, .91, .86,
        (Scaffold("clock feed", ScaffoldRole.INPUT_ONLY),),
    ),
    BindingProfile(
        "service", 3, .96, .94, .91, .92, .90,
        (Scaffold("human review", ScaffoldRole.INPUT_ONLY),),
    ),
]

assert not admissible(profiles[0])  # metrics pass; causal control is outside
assert admissible(profiles[1])
assert admissible_cuts(profiles)[0].cut == "control loop"

for profile in admissible_cuts(profiles):
    inputs = [s.name for s in profile.external_scaffolds]
    print(profile.cut, "receives", ", ".join(inputs) or "no external inputs")

Passing a threshold is only the first filter. The synthetic model cut passes every numeric minimum, yet the assertion rejects it because an external controller still governs the decision. The control-loop cut is the narrowest admissible candidate because its remaining clock feed supplies input without maintaining the target relation. If the clock resolved valid time or vetoed action, it would become relation-maintaining or controlling and the boundary would have to expand. Explanatory economy cannot be computed by counting dependencies alone.

The dossier changes publication language. “The system is integrated” becomes a compact, inspectable claim: the model-only cut met the metric minima but remained inadmissible because control was external; the control-loop cut was the first admissible boundary and retained only an input-only clock feed; the wider service cut also passed. The phenomenal interpretation remains open.

From a benchmark to a research programme

One successful cut-set experiment establishes a local result. A programme needs systematic variation. Begin with scenes in which every entity has a unique feature, then increase same-class density until colour, clothing, role and location recur across several candidates. Add occlusion, delayed correction, contradictory authority and changes in valid time. Move from one modality to combinations of vision, speech, documents and tool state. Preserve a confirmation set whose grammar and ambiguity pattern have never been shown during development.

Transfer should be structural. If a method succeeds because every worker name occurs in one fixed table position, it has learned the theatre rather than the relation. Change surface forms while preserving the ownership graph. Exchange people for bank accounts, parcels for medicines, bays for clinical rooms and permissions for consent. The relevant question is whether the system retains entity-specific relations when names, modalities and action domains change.

Long-horizon tests should introduce slow fractures. A mistaken identity can remain dormant for thirty steps, then enter a consequential tool call. The Binding Drift in Multi-Step Tool-Augmented Agents paper is a recent preprint that motivates this direction; its claims remain provisional until replicated. Trajectory evaluation can inspect the first moment at which ownership becomes ambiguous, the point at which the error becomes actionable and the later moment at which fluent narration conceals it. These are different intervention targets.

Independent scenario design matters because authors learn their own benchmark’s shortcuts. Give a separate team the scene algebra, the proposed mechanism and the prohibited inferences, then ask them to build cases that preserve local accuracy while breaking global ownership. Give another team only the input-output contract and ask for a cheaper system that produces the same surface. A late-fusion baseline that matches the headline result is valuable evidence, even when it defeats the original architecture.

Progress means that fewer causal explanations survive, not that one benchmark number rises. Evidence becomes stronger when a mechanism predicts the shape and timing of failure across unseen variants, when matched substitutes fail selectively, and when another laboratory reconstructs the direction of the effect. Publish nulls and reversals. They mark the valid region of the claim and protect later consciousness arguments from resting on a mechanism that never distinguished itself.

Part IV. From a functional whole to a possible subject

The protocol reaches a boundary that several philosophies approach from different directions. It does not force agreement. It forces each view to say where an additional inference enters.

Kant’s account of the unity of apperception directs attention to the conditions under which representations can be combined in one judgement. The point is deeper than gathering information in one place. A representation must be available within a rule-governed activity of synthesis. In machine terms, copied messages are weaker evidence than content that can be compared, corrected and used across a continuing process. The Stanford Encyclopedia account of Kant’s philosophy of mind helps explain why unity is an organisational demand as well as a report.

Hume’s account of mind and personal identity pulls in another direction. When introspection finds perceptions, memories, expectations and feelings, it does not separately encounter an owner that holds them. This pressure is useful for artificial systems. Researchers should not add a hidden subject merely because a causal diagram feels incomplete. The boundary dossier asks which relations actually sustain continuity. If a stable bundle of mutually constraining processes explains the data, an extra controller should not be invented for metaphysical comfort.

Buddhist traditions sharpen the same restraint while refusing the conclusion that continuity is unreal. A dependent stream can exhibit causal and ethical continuity without an unchanging substance. The Indian Buddhist philosophy of mind offers several analyses rather than one doctrine, yet impermanence and dependent arising make a practical contribution here: the relevant boundary may be enacted and renewed across moments. A candidate subject can be stable enough for one experiment without becoming an eternal owner.

Advaita Vedānta reverses the starting point. Consciousness is not produced by the changing contents it illumines. Mind, memory and individuation belong to the empirical order through which experience is structured, while awareness is primary. The overview of Śaṅkara shows why a consciousness-primary position need not equate every functional boundary with a separate source of awareness. On this reading, the protocol studies the organisation through which a perspective is expressed. It may discriminate modes of manifestation while remaining unable to manufacture or measure awareness as an object.

Sāṃkhya offers a further warning. It distinguishes the operations of cognition from puruṣa, consciousness as witness. The Stanford Encyclopedia account of personhood in classical Indian philosophy situates this distinction within a wider and internally diverse debate. A machine could reproduce complex discriminative, mnemonic and executive functions while the inference to consciousness remains unsettled. Consciousness-primary is therefore not a shortcut to machine consciousness. It can make the inferential gap more explicit.

A consciousness-primary orientation changes the interpretation before it changes the measurement. The cut-set result describes a bounded pattern in cognition. It need not explain how awareness comes into existence. It can instead describe how contents become available as one organised perspective. That is a substantive reframing, yet it earns scientific value only when it changes predictions, boundary choices or evidence requirements.

Physicalist and functionalist views place pressure elsewhere. Functionalism asks whether the relevant causal roles can be realised across substrates; the functionalism entry maps both the appeal and objections. Biological naturalism expects important properties of living neural organisation to resist software-level equivalence. Integrated Information Theory 4.0 proposes an intrinsic causal criterion that may select a boundary different from the controller cut. Seth and Bayne’s peer-reviewed review of theories of consciousness lays out the distinct commitments of global workspace and higher-order families. A recent adversarial collaboration in Nature also illustrates why rival theories should face preregistered, shared tests. Each can read the same profile, but each owes an account of why its preferred indicator connects to experience.

The lens figure is not an invitation to declare all theories equally good. Some premises may be incoherent, some may fit existing evidence poorly, and some may never generate a discriminating test. The protocol creates a common empirical surface on which those weaknesses become discussable. One group can contest the phenomenal inference while accepting the ownership fracture. Another can dispute the chosen cut while reproducing the replay effect.

This layering is especially important for ethics. A functional boundary can guide operational accountability before it guides moral status. If the controller-ledger loop caused a harmful release, the organisation responsible for the decision should be reconstructible. Moral patienthood asks a different question: whether there is something it is like to be the system and whether its states can matter to it. A fluent protest, persistent identity claim or self-model may deserve investigation and precaution. None should be converted into certainty by interface design.

Forks and interruptions complicate the boundary further. A process can be copied, paused and resumed. Two descendants may share a complete history until the fork, then develop incompatible states. The Binding Cut-Set Protocol would treat each post-fork stream as a new candidate for continuity experiments. It would record inherited memory separately from live reciprocal organisation. This prevents a stored transcript from masquerading as an uninterrupted subject.

Functional individuation should therefore remain phenomenon-specific and interval-specific. The best cut for one visual decision need not own a week of memory. The best cut for a coupled human-machine investigation need not persist after the human leaves. Boundaries can nest, overlap and dissolve. That flexibility is a virtue when the causal evidence demands it, provided the paper states exactly what changed.

What a result is allowed to say

A defensible paper reports five layers. First, what each constituent recognised. Second, which ownership and temporal relations survived. Third, which content became available to which consumers before commitment. Fourth, the narrowest tested cut that explained the profile and the scaffolds it required. Fifth, the relevance and limits under named theories of consciousness.

Claims should become stronger only when alternatives become weaker. Repeated final-answer success supports competence. Selective fractures under matched controls support a causal mechanism. Transfer across new scene generators supports a broader valid region. Independent reproduction supports provisional field evidence. A theory of phenomenal consciousness still supplies the bridge to experience.

The prohibited inference is as important as the positive result. Do not equate one chat interface with one subject, broad message distribution with access unity, recurrence with experience, memory storage with continuity, or self-report with phenomenal fact. The article becomes more useful when these tempting conclusions are printed beside the result they exceed.

Glossary

The decision this changes

The practical decision is simple. Stop assigning a subject boundary from the interface inward. Start with a registered phenomenon, draw competing cuts and disturb the relations each cut claims to contain. Preserve capability and information as carefully as the design permits. Read the resulting profile dimension by dimension.

For engineering, this catches identity, temporal entitlement and correction failures that final-answer accuracy misses. For consciousness research, it identifies where a theory must add something to the causal record. For a consciousness-primary programme, it studies the formation of an organised perspective without reducing awareness to a mechanism or treating metaphysical sympathy as empirical proof.

Before counting possible subjects, make the candidate boundary survive a cut.