The agent that vanishes when the roles are separated

Watch your hand reach for a cup. Before the movement, there is a visual scene, a bodily orientation, a judgement that the cup is yours, a selection of grip and force, and perhaps the thought “I am picking it up”. Afterwards, there may be a memory of having acted. Ordinary language assigns all of this to one subject.

Sāṅkhya asks us to slow the episode down. What if the system that detects, compares, selects and moves is not itself conscious? What if the “I” that claims the movement is another operation inside that system? What if awareness is present but does not push a single causal lever?

The Sāṃkhyakārikā names puruṣa through a cluster that includes witnessing, separateness, neutrality, seeing and non-agency. In the next verse, it says that conjunction makes the non-conscious sign-system appear conscious and makes the neutral puruṣa appear to act. The famous blind-and-lame analogy then gives each side what the other lacks: one supplies activity without seeing; the other supplies seeing without locomotion. [1, SK 19–21]

To understand that allocation, avoid translating prakṛti as inert “matter”. In this system, prakṛti includes bodies, sensory capacities, action capacities, affective and cognitive processes, the empirical “I”, and the subtle continuity that carries dispositions. Nature is active because its three guṇas transform. Puruṣa is conscious because it manifests or witnesses appearance, not because it computes a plan. Modern summaries that put “mind” on the puruṣa side erase the very puzzle the system is designed to create. [2] [6]

The cascade should not be read as a factory conveyor belt. It is an ontological account of how differentiated capacities emerge from prakṛti. Yet the allocation is clear enough to expose three roles that everyday talk fuses:

Three questions that should never share one answer

Suppose the hand reaches the cup successfully. One question is causal: which transformations made the movement occur? A second is attributive: which process represented the movement as mine? A third is phenomenological: was there something it was like for a subject to undergo the episode? A single observation can bear differently on each question. Accuracy may constrain the first. A first-person report may inform the second. Neither automatically settles the third.

The distinction becomes sharp under intervention. Block the visual input and the grip may fail. Alter retained dispositions and the person may hesitate. Suppress the ownership narrative and the movement may proceed without an explicit “I”. These are changes in the causal and self-model architecture. By contrast, Sāṅkhya denies that manipulating puruṣa is part of the same experimental space, because puruṣa does not transform. That makes puruṣa difficult to operationalise, but it also prevents the witness from becoming a convenient name for whichever mechanism has not yet been explained.

This allocation is not a generic claim made by all Indian philosophy. Nyāya and Mīmāṃsā traditions can attribute agency and cognition to an enduring self, while Buddhist arguments reject a permanent witness of the Sāṅkhya kind. [12] The disagreement matters. Sāṅkhya’s value here comes from a specific wager: the causal sufficiency of the cognitive apparatus does not exhaust the question of awareness. That wager must be assessed, not smoothed into a vague idea that “Eastern thought separates mind and consciousness”.

This separation also disambiguates “cognition”. A pattern can be discriminated inside prakṛti. That discrimination can be manifest in relation to puruṣa. A further process can describe it as “my judgement”. The same word, cognition, often slides across all three. Sāṅkhya forces the slide into view.

The load-bearing claim of this paper is narrower than Sāṅkhya’s full metaphysics. When control, appropriation and witnessing are causally distinguishable, behavioural competence plus an “I” report cannot by itself establish a conscious doer. That claim can guide an experiment even for readers who reject puruṣa as an independent reality.

flowchart LR
  A["The empty driver’s seat"] --> B["The agent that vanishes when<br/>the roles are separated"]
  B --> C["How prakṛti can remember,<br/>choose and move"]
  C --> D["Why the witness looks like a doer"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#007D68,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

How prakṛti can remember, choose and move

Return to the cup. Suppose it is filled with hot tea. The visual system registers its location. Heat from an earlier spill conditions caution. The handle’s orientation constrains the grip. A judgement selects “lift by the handle, slowly”. The arm moves. The action succeeds, and the episode alters future readiness.

No single stage needs a miniature observer. Each stage can be explained as a transformation that receives conditions and changes what happens next. Sāṅkhya gives those transformations a distinctive vocabulary.

Determination is not a conscious homunculus

The Sāṃkhyakārikā characterises buddhi through determination, adhyavasāya. A determined state does not require a tiny person deciding inside the mind. It is the system’s settled orientation among alternatives. In later analysis, buddhi can receive what the other instruments present, integrate across objects and reach the discriminative insight that distinguishes prakṛti from puruṣa. [1, SK 23, 35–37]

Ahaṃkāra is often rendered “ego”, but the English word invites too much personality theory. Its more precise role here is I-making or appropriation. It turns a determination and movement into the form “I know”, “I want” or “I act”. Manas, meanwhile, coordinates and constructs the immediate field on which determination works. The five sensory and five action capacities are powers of receiving and doing, rather than only anatomical organs. This functional reading is consistent with major historical reconstructions, while the details of how cognition becomes conscious remain contested. [3] [11]

Memory without a remembering witness

The kārikā does not supply a modern taxonomy of working, episodic and semantic memory. It supplies a continuity architecture. The subtle or sign-bearing body, liṅga, carries the inner instrument and its dispositions across changing embodiments in the tradition’s account. Its bhāvas and residual saṃskāras shape later transformations. Memory, in the operational sense needed here, is therefore not a file opened by puruṣa. It is the present effect of retained structure on a new cognitive episode. [1, SK 40–46, 67] [8]

This distinction matters. A system may use past information without re-presenting a past scene. A bias in grip, a learned aversion, a linguistic habit and an explicit recollection are different expressions of retained state. Calling all of them “memory” obscures the causal carrier. Sāṅkhya’s subtle-body account is not a neuroscientific storage model, but it asks the right first question: what persists such that the next transformation is not history-free?

Action is transformation, not a command from outside nature

In this framework, action is not a mysterious injection of force by a self. It is pariṇāma, an ordered transformation of prakṛti under prior and present conditions. An intentional act is special because inner capacities organise it around an object and purpose, not because puruṣa leaves neutrality and becomes a motor. Research on psychophysical agency in Sāṅkhya-Yoga highlights this distinction between the doer embedded in psycho-physical processes and the self understood as pure consciousness. [8]

This does not make every action random or unowned in practical life. Ahaṃkāra, buddhi and dispositions form a temporally structured doer that can learn, anticipate consequences and answer to others. The tradition’s claim of non-agentive puruṣa therefore should not be converted into “nobody is responsible”. Causal and ethical responsibility operate at the level where intentions, habits, harms and corrections actually change future conduct. Metaphysical non-agency is not a licence to erase those relations.

The kārikā also describes prakṛti as acting for the experience and liberation of puruṣa, using images such as milk nourishing a calf and a dancer withdrawing after being seen. Those are teleological images, not extra executive modules. They say what the process is for within the system’s soteriology. They do not put a planning homunculus behind nature. Recent scholarship usefully asks what this codependence means from prakṛti’s side rather than treating her as disposable machinery. [7]

The wheel keeps turning after the attribution breaks

A common objection now appears. If liberating discrimination reveals that puruṣa does not act, why does the person continue walking, speaking and choosing? The kārikā answers with residual momentum. Once the decisive discrimination has arisen, the body’s activity can continue through prior impressions, like a potter’s wheel that spins after the potter has stopped applying force. Puruṣa remains a spectator while prakṛti completes its course. [1, SK 64–68]

The mechanism therefore explains a seemingly contradictory pair: action can continue, yet the conscious witness need not be its producer. Non-agency is not inactivity; it is a claim about where causal production belongs.

Why the witness looks like a doer

The difficult part is no longer how action occurs. It is why action appears conscious and owned. The Sāṃkhyakārikā proposes a double misattribution. Because of conjunction, the non-conscious cognitive complex appears conscious. At the same time, the non-agentive puruṣa appears to perform the guṇas’ activity. Each side borrows the other’s signature.

This is more than the statement that people make an intellectual mistake. The empirical “I” is itself part of the mechanism. Ahaṃkāra packages a distributed episode into a first-person form. That form is useful for coordination, commitment and social accounting. Its usefulness does not make its metaphysical implication true.

Reflection is an explanation added by later interpreters

Later Sāṅkhya commentators often use reflection imagery. Buddhi is likened to a transparent or polished medium in which consciousness is reflected; cognitive activity then seems conscious, while puruṣa seems qualified by the content. This helps explain intimacy without causal transfer. Yet the reflection model should not be silently inserted into every line of the root text. The kārikā gives conjunction and apparent borrowing; the more developed mirror mechanics belong to a commentarial history. [5]

The distinction matters because a metaphor can solve one problem by hiding another. A mirror explains how an image appears without the face entering the glass. It does not explain how an immutable puruṣa and changing prakṛti enter a relation, why one stream of cognition is manifest to one puruṣa, or how the relation can be known. These are real pressures on the system, not minor missing details.

A small formal model that prevents the slide

Formalism is useful here only because it blocks an illicit substitution. Let X_t be the current input, B_t retained dispositions, D_t a determination, A_t an executed action and R_t an ownership report. Let W mark the hypothesised witness relation. Then a minimal operational model is:

Now the counterfactuals are explicit. Intervene on D, and action should change. Intervene on O, and first-person reporting may change while action remains stable. Intervene on a merely declarative W, and neither should change. If toggling the supposed witness changes control, the implementation no longer represents Sāṅkhya’s non-agentive relation.

Matched behaviour is where explanation begins

Imagine two systems that produce the same steering action and the same sentence about it. In system A, one recurrent controller generates both. In system B, a deterministic controller acts and a separate language model reconstructs a plausible reason afterwards. In system C, a human remotely authorises the manoeuvre while the vehicle narrates it locally. Behavioural matching hides three different causal stories.

A useful experiment therefore needs a rival generator. Remove the narrator and check whether action survives. Swap the retained state and check which preferences move. Delay the report until after the action and test whether its details track the actual control variables. Patch the controller while holding the report prompt fixed. These interventions convert a surface resemblance into competing mechanism predictions.

Machine self-report is especially vulnerable to contamination because the question itself supplies concepts, expectations and linguistic form. A model can say “I felt in control” by completing a learned discourse about agency. That output may still reveal internal information under carefully controlled conditions, but only after leakage, prompting shortcuts and generic narrative competence have been excluded. A report becomes evidence through discriminating controls, not through grammatical sincerity.

Sāṅkhya contributes a useful rival generator of its own: an apparently conscious doer produced by competent prakṛti plus I-making and witness-ascription. A physicalist mimic can replace puruṣa with higher-order monitoring; a functionalist can make the same organisation constitutive of consciousness; a purely narrative system can generate the report without either. The observation is valuable only when these alternatives predict different intervention outcomes.

The function–witness matrix

The table below is not a scorecard for deciding what is conscious. It is a discipline for asking which role is being claimed. It keeps the tradition’s allocation visible while preventing a machine analogy from expanding beyond its evidence.

A disciplined comparison with agency science

Contemporary research on the sense of agency also resists a single undivided construct. Synofzik and colleagues distinguish a low-level feeling of agency from a higher-order judgement of agency. Grünbaum and Christensen argue that common measures may target different abilities or phenomenal characters, bodily actions or external effects. [9] [10]

The comparison is useful but bounded. The source structure is Sāṅkhya’s distinction among cognitive function, I-making and puruṣa. The target structure is an empirical separation among control signals, agency judgements and reported phenomenology. The preserved relation is separability. The break occurs because no modern agency measure operationalises puruṣa, and Sāṅkhya’s aim is liberation from misidentification, rather than only better construct validity.

This is structural guidance, not formal equivalence. The traditions do not share objects, measurement practices or ontological commitments. Their overlap generates a design question: can we alter control, ownership report and candidate witness-status independently, then observe which judgements follow?

stateDiagram-v2
  [*] --> Claim
  state "Three questions that should<br/>never share one answer" as Claim
  state "Determination is not a<br/>conscious homunculus" as Test
  state "Memory without a remembering<br/>witness" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build the separation into systems without importing the metaphysics

Consider a configured travel-reimbursement assistant. An employee submits a delayed-flight claim. A language model reads the narrative and proposes a reimbursement. A policy service checks permitted context and limits. An authority kernel decides whether the proposed amount may be committed. A typed action writes to the payment system. An effect receipt confirms the new world state. A user-facing model explains the result.

If the final interface says “I approved your reimbursement”, it has compressed at least five roles into one pronoun. The model may have proposed but lacked authority. The policy service may have constrained the amount. The payment system produced the effect. The receipt, not the prose, proves the outcome. None of these facts decides whether any candidate subject experienced the episode.

Failure injection reveals the hidden bundle

A diagram can claim that roles are separate while shared state quietly couples them. Failure injection tests the claim. Give the proposal model a stale policy but keep the authority kernel current. The proposal should become wrong while the permitted action remains bounded. Corrupt the narrator’s context after execution. The explanation may become false, but the effect receipt should preserve what happened. Revoke authority between proposal and commit. The typed action should stop even if the model remains confident.

Now remove the effect receipt while allowing the action to time out ambiguously. The system should not let the narrator infer success from its own prior intention. It should enter an unknown-outcome state, perform readback through an independent path, then either confirm, compensate or escalate. This is where the philosophical distinction becomes operational: wanting, proposing, authorising, acting, claiming and knowing the outcome are different state transitions.

A more subtle fault targets ownership. Prompt the narrator with “Explain why you approved the claim”, even though it only received a post-hoc log. If it complies, the interface manufactures an agent boundary and a false causal history. The recovery is not to ban first-person language in every context. It is to bind any ownership statement to a decision receipt that identifies the actual proposal, authority and effect chain.

The decisive engineering test is whether each role fails locally. A bad explanation should not authorise money. A blocked action should not be reported as completed. A missing receipt should prevent certainty. An unresolved witness question should not weaken deterministic control.

The function–witness separation lab

The lab below makes the prose operational. It simulates a controller that acts on binary cues, an ownership module that emits first-person claims, and a witness marker that is deliberately absent from the action path. Change the controls, run paired trials and inspect which metrics move.

function mulberry32(seed) {
  return function random() {
    let t = seed += 0x6D2B79F5;
    t = Math.imul(t ^ (t >>> 15), t | 1);
    t ^= t + Math.imul(t ^ (t >>> 7), t | 61);
    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
  };
}

function runCondition({ controller, ownership, witness, trials, seed }) {
  const random = mulberry32(seed);
  let correct = 0;
  let firstPerson = 0;

  for (let i = 0; i < trials; i += 1) {
    const cue = random() < 0.5 ? 0 : 1;
    const action = random() < controller ? cue : 1 - cue;
    const claimsOwnership = random() < ownership;

    correct += Number(action === cue);
    firstPerson += Number(claimsOwnership);
    void witness; // recorded, intentionally absent from the causal path
  }

  return {
    accuracy: correct / trials,
    ownershipRate: firstPerson / trials
  };
}

Use the lab as a test harness pattern. In a real system, replace the toy cue with a versioned input, the controller with a model proposal and authority path, the ownership module with the interface narrator, and the synthetic metric with decision and effect receipts. Then inject failures. If changing the narrator changes permissions, or if a passive audit component can silently alter action, your architecture contradicts its own declared roles.

The strongest boundary: separability does not prove dualism

The architecture can validate a functional distinction without validating Sāṅkhya’s ontology. A physicalist can say that control, ownership and witnessing are all realised by physical processes, even when distributed. A biological naturalist can accept the functional audit while insisting that machine organisation lacks the biological conditions for consciousness. A functionalist can argue that the right causal organisation may itself constitute mentality, making a causally inert puruṣa unnecessary.

An emergentist can place awareness at a higher level of organised prakṛti. An idealist can reverse the dependence and treat prakṛti as appearance within awareness. Neutral monism and dual-aspect views can treat function and witness as aspects of a more basic reality. Panpsychism can distribute experiential properties more widely, inheriting a combination problem. Process philosophy can replace both enduring principles with events and relations. Each view can accept the intervention results while explaining them differently.

Sāṅkhya faces its own hard question. If puruṣa never changes or acts, what exactly individuates its relation to one cognitive stream, and how can that relation make a difference without becoming causal? Scholarship disagrees about how much conjunction, reflection and teleology explain. The unresolved relation between puruṣa and prakṛti is the paper’s principal metaphysical limitation. [4] [6] [7]

A testable prediction about anthropomorphic credit

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The empty driver’s seat
  M->>E: Three questions that should never share one answer
  E-->>M: compare rival explanations
  E->>D: Separate function, authority, ownership and witness
  D-->>Q: revise scope or act

Separate function, authority, ownership and witness

The shuttle’s sentence, “I reduced speed”, initially looked like a transparent description of one conscious agent. We can now see at least four unresolved questions inside it. What process discriminated the hazard? What component had authority over steering? What generated the first-person ownership claim? What evidence, beyond those functions, bears on witnessing or experience?

Sāṅkhya’s answer is radical: the first three belong to prakṛti; puruṣa witnesses without acting. You need not accept that ontology to inherit its discipline. The practical lesson survives as a causal rule: never let a unitary surface supply a unitary agent before you have tested the roles underneath it.

For engineers, this changes system architecture. Keep model proposal, world state, permitted context, authority, typed action, evidence, readback and narration distinct. For researchers, it changes experimental design. Intervene on control and ownership separately; do not make first-person language the bridge to consciousness. For philosophy, it changes the burden of argument. A witness can no longer be smuggled in as the thing that discriminates, remembers and acts.

The deeper provocation remains. Perhaps conscious agency is not a primitive entity but an alignment among causal control, ownership and witnessing. Perhaps Sāṅkhya is right that their apparent unity is a misidentification. Perhaps a rival metaphysics explains the same separability better. The next decision is not to choose a metaphysics by intuition. It is to build cases in which the roles come apart, record what changes, and refuse every conclusion the intervention did not earn.