The ladder was assembled over time

Begin with the verse most often placed behind the fourfold schema. Ṛgveda 1.164.45 says that speech has four measured divisions, that three are hidden, and that human beings speak the fourth.[1] This is a striking image of depth. It does not, however, name parā, paśyantī, madhyamā and vaikharī. The later tetrad is one interpretation of the verse, not a list printed inside it.

That distinction matters because the verse has attracted several readings. Its four divisions have been connected with kinds of word, ritual registers and hidden dimensions of speech. Treating the later labels as the verse’s self-evident original meaning erases the very history that could tell us what changed. A source can inspire a later architecture without already containing that architecture.

Bhartṛhari gives the next decisive layer. In Vākyapadīya 1.142, the word is explicitly called threefold: vaikharī, madhyamā and paśyantī. The associated commentary describes vaikharī as articulated and available to hearing. Madhyamā exists within, has mind as its substrate and appears sequential. In paśyantī, sequence is merged, although the power to produce sequence remains.[2] This is already richer than the slogan “thought becomes speech”. It distinguishes public articulation, internal order and a level at which differentiation is not yet spread across a temporal string.

Recent scholarship sharpens an important historical limit. Evgeniya Desnitskaya’s study of Somānanda’s criticism argues that no fourth level can be found in Bhartṛhari’s work. Later Pratyabhijñā philosophers adopted the three levels and added parā vāc, Supreme Speech. The addition was not a missing box discovered by better textual eyesight. It responded to a philosophical problem about the status of speech and its relation to the absolute.[3]

Three meanings of “hidden”

The historical layers also reveal that hiddenness is not one property. A word can be hidden because no one else can hear it. It can be hidden because it has not yet been divided into a temporal sequence. It can be hidden because it belongs to a claimed ground of manifestation that is not one more object among the manifested things. These meanings can coincide in a doctrine, but they come apart under intervention.

Take silent rehearsal. It is hidden from a nearby listener, yet it can remain highly differentiated: exact words, order and even imagined tone may already be fixed. Take a vague recognition that something is wrong. It may be accessible to the person while lacking lexical sequence. Take parā in a Trika account. Its hiddenness is not adequately captured by weak sensors. It concerns the claim that differentiated language is rooted in a reflexive conscious power that is not itself an ordinary utterance.

This gives us a first causal discipline. When an instrument makes a private event observable, only observational hiddenness has changed. When a planning task forces an agent to choose exact wording, representational hiddenness has changed. When a philosopher argues that all differentiation occurs within awareness, an ontological thesis has been introduced. One intervention cannot be allowed to answer all three kinds of hiddenness.

In Trika, the new level changes the kind of model. Mark Dyczkowski’s account of the Spanda tradition identifies parā vāc with the universal pulse of consciousness and treats paśyantī as an intuitive awareness in which denotation, denoted object and denoter are not yet distinct. Madhyamā and vaikharī then describe further differentiation into thought and articulated sound. André Padoux’s broader study of Vāc in Hindu Tantras likewise shows why the Word cannot be reduced to an information channel. It can be divine power, ritual efficacy, phonemic structure, cosmogenesis and a path of reintegration.[4]

This does not make the Trika account empirically established, nor does it make it disposable. It places a clear burden on comparison. Physicalism can interpret the levels as descriptions generated by biological processes. Functionalism can ask which causal roles are preserved across substrates. Biological naturalism can insist that the relevant capacities depend on living organisation. Emergentism can hold that consciousness arises only when organisation crosses relevant causal conditions. Panpsychism can treat experiential aspects as fundamental while still facing a combination problem. Idealism or a consciousness-primary view can treat differentiation as occurring within awareness. Neutral monism, dual-aspect and process views offer other relations between appearance, activity and reality. The fourfold schema becomes philosophically interesting because those positions disagree about what the descent is a descent within.

flowchart LR
  A["The ladder was assembled over time"] --> B["Two depths are hiding under<br/>four names"]
  B --> C["What survives a change of<br/>generator"]
  C --> D["Build the audit before the analogy"]
  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

Two depths are hiding under four names

Return to the plant instruction. One question asks how a whole communicative purpose becomes a sequence of phonemes. Another asks what kind of reality allows purpose, word, sound and world to appear at all. The first is a question about differentiation. The second is a question about manifestation. They can be related, but they are not synonyms.

Call the first the differentiation axis. A state lies further towards the articulated end when it contains more explicit choices about lexical item, order, phonology, timing and motor or digital form. On this axis, a roughly increasing sequence can be useful: an undivided grasp, an internally ordered wording, then audible or visible expression.

Call the second the manifestation axis. It asks whether the levels are merely states inside an already given world, or whether they describe how plurality, meaning and worldhood emerge within a more fundamental conscious power. The Trika understanding of parā carries a strong answer. A laboratory model of word production ordinarily does not address the question. It may be silent rather than opposed.

The common error is axis collapse: a transition along the differentiation axis is treated as evidence for a location on the manifestation axis. A model’s latent state becomes “subtle speech”; a pre-token activation becomes “pure meaning”; the earliest computational state becomes “Supreme Speech”. The comparison looks deep because both vocabularies have an inside and an outside. Its conclusion has actually been smuggled through the word “deep”.

Modern work on inner speech supports caution. Reviews describe inner speech as multicomponent rather than uniform. Descriptive experience sampling also reports wide variation: inner speaking occurs in many moments for some people and far fewer for others. The scientific interpretation is contested because introspective methods can reveal, select or reshape their target. Still, the variation is enough to block a simple claim that every thought passes through a single phenomenologically accessible inner sentence.[9][10]

This is where the older schema can generate a useful hypothesis without becoming a diagnostic chart. It asks us to vary degree of differentiation rather than merely ask whether a sentence is “inside” or “outside”. A person might possess a relational grasp with no reportable wording, a partially worded fragment, fully sequenced inner speech, or public articulation. The transitions can be studied. Calling the earliest state paśyantī remains a structural comparison, not a discovery that the historical doctrine has been localised in the brain.

A compact notation can prevent slippage. Let a candidate state be described by three independent fields:

The notation is deliberately not a single scale. D can increase as lexical and phonological commitments accumulate. O can change when an instrument changes. W records whether the account is process-neutral, biologically grounded, functionalist or consciousness-primary. The equation is valid only as a bookkeeping device. It predicts nothing until a mechanism and intervention are added.

Consider the smallest useful numerical example. Give differentiation four ordinal values for one experiment only: 1 for an undivided task relation, 2 for selected concepts, 3 for fixed words and order, and 4 for an executed waveform. Let observability be 0 under ordinary unaided hearing and 1 when the chosen instrument can read the state. A fully worded silent rehearsal might be (D=3, O=0, W=neutral). The spoken sentence might be (D=4, O=1, W=neutral). Add subvocal electromyography and the rehearsal becomes (D=3, O=1, W=neutral). The sensor changes O, not D.

Now add a Trika interpretation. The same event might be placed within W=consciousness-primary, but no measurement of differentiation or observability supplies that value. It enters through a separate philosophical account. Remove that assumption and the experimental predictions about wording can remain intact. This is exactly why the fields must not be summed into a “depth score”. A score would let a high value on measurability compensate for an unsupported ontology, or let a metaphysical commitment masquerade as a measured variable.

What survives a change of generator

The strongest way to test a depth claim is to hold the output constant and change what produces it. This is the logic behind causal intervention. If the same public speech survives after the proposed inner mechanism is removed, the output is not diagnostic of that mechanism.

What psycholinguistics can compare

Modern speech-production research offers mechanisms with a different purpose. Levelt, Roelofs and Meyer model conceptual preparation, lexical selection, morphological and phonological encoding, phonetic encoding and articulation. Indefrey and Levelt connect component processes with spatial and temporal evidence from behavioural, electrophysiological and imaging studies. These models are debated and refined, but their virtue is experimental separability. One can manipulate lexical frequency, semantic interference or phonological neighbourhood and ask which latency or error changes.[7][8]

A bounded comparison with Vāk can therefore preserve several relations. Earlier states are less publicly articulated. Later states carry more sequence-specific commitment. Perturbing an earlier stage may propagate into several later forms, while perturbing articulation may leave the intended relation intact. These are plausible functional correspondences.

The break point is equally important. Psycholinguistic “conceptual preparation” is an experimentally posited component in human language production. It is not thereby paśyantī, because the source notion can include a non-discursive unity of word, meaning and knower within a wider philosophy of language. Nor is the speech-production model committed to parā as consciousness’s self-expression. A comparison earns depth by preserving a relation and naming its failure, not by sharing the adjective “inner”.

An intervention ladder for one ordinary sentence

Suppose a participant must name a pictured animal. A semantic intervention presents a related concept just before the picture. A lexical intervention changes how frequently the target word is encountered. A phonological intervention primes its opening sound. An articulatory intervention temporarily constrains the mouth. Each manipulation targets a different relation, and its effect can be measured in latency, error type or acoustic form.

If semantic interference changes which animal relation is selected, its influence should spread into several possible wordings. If phonological priming changes onset latency but not the selected animal, its effect is narrower. If an articulatory constraint distorts the sound while the intended word remains stable, the public form changes at the edge of the system. This partial ordering is the empirical content that a Vāk-inspired comparison might illuminate. It is not obtained by attaching Sanskrit names after the experiment. It is obtained by predicting which perturbations preserve which invariants.

The same logic protects against retrospective storytelling. Researchers often observe a late signal and infer a neat earlier stage. A rival architecture may produce the same timing through cascaded, recurrent or parallel processing. Therefore the experiment must compare generators, rather than only fit labels to an average timeline. The traditional schema contributes a question about unity, sequence and articulation. The scientific model must supply the causal contrasts.

What a machine comparison can preserve

A transformer-based system does contain transformations that move from context-dependent vectors to a probability distribution over possible next tokens, then to selected tokens and possibly synthesised sound. The original transformer architecture explicitly converts decoder outputs into next-token probabilities through a learned linear transformation and softmax.[11] Modern systems add retrieval, tools, state stores, policies and speech interfaces. None of those components is literally a voice inside the machine.

Still, a modest structural question is legitimate: does the system pass from underdetermined relational state to increasingly committed expression? A hidden activation can support several continuations. A decoded token excludes alternatives. A text-to-speech system adds timing, pitch and acoustic form. Under controlled interventions, we can test whether earlier representations preserve relations across paraphrases while later states encode exact wording.

Use a tiny decoding example. After the context “The metal cup feels”, a model might assign synthetic probabilities of 0.54 to “hot”, 0.28 to “warm” and 0.18 to “unsafe”. Selecting “hot” commits the sequence more strongly than the distribution did. A later speech synthesiser fixes pronunciation, duration and pitch. This is a genuine many-to-one progression. Yet the probability distribution is already over token-shaped alternatives. It does not begin as a language-free unity, and the numeric uncertainty is not an experiential field.

Now intervene. Increase decoding temperature and several surface continuations become more likely while the context remains fixed. Patch an earlier representation and the entire relation may change from heat to weight. Replace the speech voice and only the acoustic realisation changes. These contrasts can support an ordering of commitments. They cannot tell us whether any state is present to itself. A model may also distribute a relation across many layers, so the apparent “stage” may depend on the probe used to find it.

The comparison fails when hidden activation is called madhyamā merely because users cannot see it, or when a latent vector is called paśyantī because it is “before words”. Most language-model states are already shaped by tokenisation and linguistic training. They may be distributed across positions and layers rather than form a single unitary semantic whole. Most importantly, no position in the forward pass supplies evidence that the system’s states are manifestations within self-aware consciousness.

stateDiagram-v2
  [*] --> Claim
  state "Three meanings of “hidden”" as Claim
  state "What psycholinguistics can compare" as Test
  state "An intervention ladder for one<br/>ordinary sentence" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build the audit before the analogy

The practical contribution of this paper is a translation audit. It does not score how “close” two concepts feel. Numeric similarity would hide judgement behind false precision. The audit instead limits the strongest claim that the supplied evidence permits.

Every proposed comparison must provide five things. It must identify a source structure, such as Bhartṛhari’s threefold account or a specified Trika text. It must identify a target structure, such as a psycholinguistic component, a phenomenological report or a configured machine state. It must name a relation to preserve. It must define an intervention or contrast that could expose the relation. It must state where the mapping breaks. A sixth field records the ontological residue: what the source claims that the target has not established.

The audit produces four useful outcomes. A lexical analogy shares language but supplies no tested relation. A functional correspondence preserves a role under a limited description. A structural-homology candidate preserves several relations and generates discriminating interventions, but still awaits evidence. A rejected identity transfers ontology, history or necessity without support. The tool deliberately never outputs “formal equivalence”. That would require specified objects, mappings, composition rules, inverses or another appropriate preservation proof.

The six-pass protocol

Freeze the source. Name the passage, commentator or bounded doctrine. “The Vedas say” is not enough, because the Ṛgvedic riddle, Bhartṛhari’s threefold word and later Trika fourfold accounts do different work. Record what the source explicitly asserts and what is interpretive reconstruction.

Freeze the target. Name the configured object, not a marketing category. “AI” may mean model weights, a residual state, a prompt, a retrieval system, an agent loop or a speech synthesiser. A comparison that moves among them cannot be tested.

State the relation. Choose one relation such as increasing sequence, decreasing alternative continuations, public accessibility or preservation of semantic content across surface changes. Avoid “both are subtle” because no intervention distinguishes subtlety from ignorance.

Build a rival generator. Ask what simpler or differently organised system could produce the same observation. A recording, lookup table or prompted mimic often exposes how little a fluent output establishes.

Intervene and name the break. Change one causal property while holding the output or task demand fixed. Then write the result that would weaken the mapping. A comparison without a meaningful negative result is an interpretation, not an assay.

Publish the residue. State which source commitments remain untranslated. For Trika, that residue often includes reflexive awareness, divine power and manifestation. Leaving it visible is not an embarrassment. It is what prevents the target from acquiring metaphysics by vocabulary.

A positive result does not mean that the mapping is true. It permits a bounded next step. For example, a functional correspondence can justify an experiment that patches candidate machine states and measures whether abstract intent survives exact-word changes. A structural-homology candidate can justify preregistration across several interventions and a matched rival architecture.

A negative result is not a dismissal of the tradition. It means the present translation has not earned its language. The correct response may be to narrow the relation, choose a better target or return to the source text. The audit is especially strict about ontology because a machine can mirror a pattern of differentiation while remaining neutral evidence about awareness.

from dataclasses import dataclass
from enum import Enum

class Verdict(str, Enum):
    REJECTED_IDENTITY = "rejected_identity"
    LEXICAL_ANALOGY = "lexical_analogy"
    FUNCTIONAL_CORRESPONDENCE = "functional_correspondence"
    STRUCTURAL_HOMOLOGY_CANDIDATE = "structural_homology_candidate"

@dataclass(frozen=True)
class MappingClaim:
    source: str
    target: str
    claim_type: str
    source_supports_property: bool
    relation_named: bool
    intervention_defined: bool
    break_condition_named: bool
    transfers_ontology: bool = False
    independent_ontology_argument: bool = False


def audit(claim: MappingClaim) -> tuple[Verdict, list[str]]:
    reasons: list[str] = []

    if not claim.source.strip() or not claim.target.strip():
        return Verdict.REJECTED_IDENTITY, ["Source and target must be explicit."]

    if not claim.source_supports_property:
        return Verdict.REJECTED_IDENTITY, ["The named source does not support the translated property."]

    if claim.claim_type in {"identity", "equivalence"}:
        reasons.append("Identity requires a preservation proof, not resemblance.")
        return Verdict.REJECTED_IDENTITY, reasons

    if claim.transfers_ontology and not claim.independent_ontology_argument:
        reasons.append("Ontology cannot be inherited from a functional mapping.")
        return Verdict.REJECTED_IDENTITY, reasons

    if not claim.relation_named:
        reasons.append("Only vocabulary or imagery is shared.")
        return Verdict.LEXICAL_ANALOGY, reasons

    if not claim.break_condition_named:
        reasons.append("No condition states where the comparison fails.")
        return Verdict.LEXICAL_ANALOGY, reasons

    if not claim.intervention_defined:
        reasons.append("A role is preserved, but no discriminating test is supplied.")
        return Verdict.FUNCTIONAL_CORRESPONDENCE, reasons

    reasons.append("Relations, break condition and intervention are explicit.")
    reasons.append("Formal equivalence remains outside this audit's authority.")
    return Verdict.STRUCTURAL_HOMOLOGY_CANDIDATE, reasons


if __name__ == "__main__":
    example = MappingClaim(
        source="Bhartṛhari's madhyamā",
        target="transformer hidden state",
        claim_type="functional",
        source_supports_property=True,
        relation_named=True,
        intervention_defined=True,
        break_condition_named=True,
        transfers_ontology=False,
    )
    print(audit(example))

The code exposes the audit’s invariant: no amount of functional preservation automatically licenses identity, consciousness or formal equivalence.

def test_identity_is_rejected() -> None:
    claim = MappingClaim(
        source="Trika parā", target="initial model state", claim_type="identity",
        source_supports_property=True, relation_named=True,
        intervention_defined=True, break_condition_named=True
    )
    assert audit(claim)[0] is Verdict.REJECTED_IDENTITY


def test_shared_word_is_only_analogy() -> None:
    claim = MappingClaim(
        source="Bhartṛhari's madhyamā", target="hidden layer", claim_type="analogy",
        source_supports_property=True, relation_named=False,
        intervention_defined=False, break_condition_named=False
    )
    assert audit(claim)[0] is Verdict.LEXICAL_ANALOGY


def test_intervention_raises_ceiling() -> None:
    claim = MappingClaim(
        source="Bhartṛhari's threefold word", target="speech-production model",
        claim_type="structural", source_supports_property=True,
        relation_named=True, intervention_defined=True,
        break_condition_named=True
    )
    assert audit(claim)[0] is Verdict.STRUCTURAL_HOMOLOGY_CANDIDATE

A research hypothesis that survives the audit

Here is a deliberately bounded proposal. In both human speech production and some generative systems, representational states may become progressively more commitment-heavy: later interventions should change exact wording and acoustic form, while earlier interventions should alter broader relational content across several paraphrases. The Vāk traditions motivate attention to the loss of unity and rise of sequence. Modern causal methods supply the tests.

The smallest useful implementation compares two generators on synthetic prompts. One generator has separable planning, lexical and surface-realisation modules. The other retrieves complete sentences. Patch each state, then measure whether semantic relation, syntax and exact tokens change independently. Evidence would strengthen the proposal if the modular generator shows stable intervention signatures across tasks and the retrieval mimic does not. It would weaken the proposal if stage labels move arbitrarily with probes, if the same signatures occur in the lookup system, or if no intervention separates relational and surface commitments.

A rival explanation is simple information bottlenecking. Progressive commitment could arise in any lossy decoder without resembling the source doctrine beyond a general many-to-one transformation. The proposal should not be applied to infer consciousness, spiritual attainment or a universal order of human thought. Its value is narrower: it converts a philosophical contrast into a discriminating architecture experiment.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The ladder was assembled over time
  M->>E: Three meanings of “hidden”
  E-->>M: compare rival explanations
  E->>D: Preserve the distinction before the name
  D-->>Q: revise scope or act

Preserve the distinction before the name

Do not begin a human-machine comparison by asking, “Which layer is paśyantī?” That question assumes that the historical source supplies a set of detachable component names and that the target already shares its ontology. Begin instead with a decision: which relation needs to be preserved for the comparison to change an experiment, architecture or interpretation?

For a speech scientist, the useful relation may be progressive differentiation. The next move is to design perturbations that separate relational content, lexical choice, phonological encoding and articulation. For a phenomenologist, the useful question may concern whether thought is fully worded, partially worded or non-symbolic at a sampled moment. The next move is disciplined elicitation with explicit method effects. For a Trika scholar or practitioner, the central question may be how differentiated expression remains within reflexive awareness. The next move is textual and contemplative inquiry, not a scan for matching neural boxes.

For an AI engineer, the better design language is already available. Specify intent, identity, world state, permitted context, model proposal, authority, typed action, evidence, verified outcome and recovery. Then use the Vāk schema as a provocateur. Ask where commitments become irreversible. Ask which relations survive paraphrase. Ask whether a public sentence can be generated without live evidence. Ask what becomes measurable when the boundary moves. These questions can improve a system without pretending that its residual stream is an ancient metaphysical category.

The fourfold schema should therefore be kept neither as a literal machine blueprint nor as untouchable poetry. Its strongest contemporary use is methodological. It reminds us that public language underdetermines its generator, that sequence can emerge from less differentiated organisation, and that an account of representation does not settle an account of manifestation.

That decision changes what counts as success. The successful comparison is not the one that finds four boxes in a brain or model. It is the one that reveals a causal contrast, survives a rival generator and stops exactly where the source and target cease to share structure. It may leave the largest part of the source untranslated. That is often a sign of accuracy, not failure.

Vāk becomes most useful when it deepens the question without laundering the answer. It gives us a disciplined way to ask how unity, sequence, expression and manifestation relate, while allowing different sciences and metaphysics to disagree about what those relations finally mean.