“Kashmir Śaivism” names a family of related traditions, not a single doctrine with one vocabulary. This paper focuses on the Pratyabhijñā school associated especially with Utpaladeva, Abhinavagupta and Kṣemarāja. Its central project is recognition. It seeks to show why the apparently limited knower is, at the deepest level, the free consciousness called Śiva. The philosophical demonstrations are therefore not detached puzzles. They are meant to make a transformative recognition possible.

Start with prakāśa. The ordinary image is light, but literal illumination is too narrow. In this context, the term points to manifestation: the fact that an object, thought or sensation appears rather than remaining absent. A blue cup appears. A memory appears. Doubt appears. Even darkness can appear. The relevant “light” is not a beam shining on these items from outside. It is their occurring within awareness.

That seems sufficient. If the cup is manifest, what else could awareness require? Utpaladeva’s answer is severe. Manifestation without vimarśa would remain inert. In a compact verse, he says:

svabhāvam avabhāsasya vimarśaṃ vidur anyathā / prakāśo ’rthoparakto ’pi sphaṭikādijaḍopamaḥ

Īśvarapratyabhijñākārikā 1.5.11

A careful paraphrase is: reflexive apprehension is the nature of manifestation; without it, a luminosity coloured by objects would resemble an insentient crystal. A crystal may receive blue from the cup and red from a flower. Its state varies systematically with what surrounds it. Yet correlation is not awareness. Something can carry the form of an object without the object being present for it.

Vimarśa is difficult to compress into one English word. Scholars use reflexivity, reflective awareness, recognitive apprehension and judgement. “Reflection” can mislead because it suggests a later act that turns back on a completed first act. The Śaiva claim is stronger. Reflexive apprehension belongs to the original manifestation. It is not a commentary written after the fact.

The argument does not prove every aspect of Pratyabhijñā. A critic can deny that manifestation must be self-manifesting, adopt a higher-order representation, or explain self-knowledge through reliable access relations. The point is narrower: adding another representation does not by itself explain presence. It merely relocates the demand. What makes the higher-order state available? If a third state is unnecessary because the second is self-present, then intrinsic reflexivity has entered the theory at that level.

Why action belongs inside the account

It is easy to hear “self-awareness” and imagine a silent mirror. Utpaladeva does not stop there. Around the crystal argument, he connects reflexive apprehension with intention, speech, agency and freedom. Isabelle Ratié’s work on action makes the point especially sharp: the power by which consciousness grasps itself is not a passive duplicate of cognition, but belongs to its capacity to initiate and organise expression. This is why the paper’s later machine test requires a change in control. A state that can be reported but never alters selection, veto or correction may be informationally available, yet it lacks the action-bearing relation that motivates the comparison.

The analogy still breaks decisively. A machine policy gate is specified by designers, constrained by external rules and dependent on physical infrastructure. Pratyabhijñā’s svātantrya is self-dependent sovereignty, not permission inside somebody else’s architecture. The preserved relation is only that reflexive uptake should make a causal difference to what follows. The metaphysical source of that efficacy is not carried across.

flowchart LR
  A["The identical warning"] --> B["The light that needs no second<br/>lamp"]
  B --> C["Recognition changes ownership,<br/>not inventory"]
  C --> D["What survives translation into<br/>a machine"]
  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

Reflexivity explains why manifestation is not inert. Recognition explains how a limited knower can fail to grasp what is already manifest. The ordinary grammar of recognition is simple: “That person I met years ago is this person before me.” The earlier presentation and the present presentation differ. Recognition binds them without pretending that the two encounters are numerically identical.

Utpaladeva builds much of his case through memory. A past perception has vanished. The remembered object is not simply present again in the same way. Still, the recollection carries a relation to the earlier episode: “I saw that.” For him, isolated cognitions could not perform this synthesis by themselves. The one who now remembers must in some sense appropriate the earlier seeing. His conclusion is a continuing knower whose powers include cognition, memory and exclusion.

A contemporary reader may reject the inference to an enduring, ultimately universal subject. Cognitive science can model memory through traces, reconstruction and distributed processes. Buddhist critics historically offered sophisticated continuity without a permanent self. Those alternatives matter. Yet the Pratyabhijñā analysis exposes a question every account must answer: what makes a present reconstruction count as the system’s memory rather than a fresh description of somebody’s past?

This is why translating pratyabhijñā as database lookup is too weak. A lookup can return an identifier with no change in self-understanding. Recognition, in the school’s philosophical use, removes a failure of identification. Torella describes the larger soteriological movement as bringing into effective life a knowledge that was somehow already there but obscured. The person does not manufacture awareness or import Śiva from outside. The person recognises the real scope of the “I”.

That last sentence must not be reduced to therapeutic confidence. “I am Śiva” is not the ordinary personality claiming cosmic importance. The source system says that the finite subject, objects and acts are manifestations of one free consciousness. Its nondualism is therefore stronger than the practical mechanism this paper will later extract. Recognition is liberation because it dissolves the contraction through which universal agency is misidentified as a bounded owner among other objects.

The reset case separates three things often called memory. A checkpoint may preserve a disposition. A transcript may preserve an episode. A current world-state record may preserve a fact. None automatically supplies recognition. Recognition is a present act in which a prior episode is taken up as relevant to this knower now.

Recognition, retrieval and inference

Three mechanisms can yield the sentence “I encountered this before”. Retrieval returns a stored record. Inference concludes from similarity that a prior encounter probably occurred. Recognition binds the present appearance to a prior episode as belonging to the same knower. In ordinary systems these mechanisms can cooperate, and no single data structure need correspond to recognition. The distinction is functional: remove the self-indexed link while preserving the record, and the system can describe a history it does not use as its own. Preserve only the disposition, and it can be shaped by a history it cannot avow. Recognition occupies the relation among record, present uptake and ownership.

A disciplined comparison begins by making loss visible. Pratyabhijñā is not a forgotten engineering manual. Its basic object is self-luminous, universal and free awareness. An artificial system is a configured assembly of models, runtime state, prompts, tools, policies and hardware. The two domains do not share an ontology merely because both can be described using words such as “self”, “state” or “reflection”.

One relation can still be carried across provisionally: a current condition may affect its own uptake and control through a path unavailable to an external imitator. That relation is weaker than awareness, but stronger than verbal self-description. It gives us an experiment rather than a slogan.

Recent model research makes this distinction concrete. One study injected representations associated with concepts into a model’s activations and asked whether the model could report the perturbation. The authors found limited, unreliable evidence that some models sometimes detect and describe injected internal content, and some evidence of control over internal states. They explicitly restricted the claim: the results did not show human-like introspection or consciousness (Lindsey et al. 2025).

A later critique asked whether the same results could arise from easier routes. Could the model detect a generic anomaly rather than an intervention specifically to its own hidden state? Could an input-only classifier predict the label as well as the model, eliminating the need for privileged access? In the tested settings, these controls substantially weakened the stronger introspection interpretation (Singh, Linzen and Ravfogel 2026). Another study found that prompted linguistic judgements did not reliably reveal a model’s own probability-level knowledge beyond matched controls (Song, Hu and Mahowald 2025).

The disagreement is productive. A positive intervention result should not be dismissed because it is not consciousness. Nor should it be inflated because “introspection” sounds mental. The correct unit of progress is a narrower causal claim whose rival generators have been actively removed.

A model is not automatically the subject

Modern AI language encourages a silent boundary shift. We say “the model remembered”, although a retrieval service supplied the document; “the model abstained”, although a deterministic policy kernel blocked the action; or “the agent knew its uncertainty”, although a separate evaluator wrote the label. Such shorthand is harmless in operations and dangerous in explanation. The component that speaks may not be the component that stores, detects or governs. RRS-102 therefore begins with a configured-system manifest: model version, runtime state, context source, monitor, policy, tool and evidence path. The manifest does not solve the mind problem. It prevents a property of the assembly from being smuggled into the most anthropomorphic component.

stateDiagram-v2
  [*] --> Claim
  state "The second lamp" as Claim
  state "The red cup" as Test
  state "From illumination to sentient<br/>manifestation" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

The comparison has now done useful work. It has changed the experimental question from “Does the system talk about itself?” to “Which self-relation is required to produce the report and govern the action?” The Reflexivity Requirement Set, or RRS-102, turns that question into six gates. It deliberately produces typed claims rather than a score.

A score would invite a false continuum from zero to consciousness. The six gates are conjunctive because different failures mean different things. A system may have privileged access but no temporal ownership. It may link episodes to itself but never allow that link to alter action. It may control action through a private variable while a public cue fully explains the report. Those are not lower quantities of one substance. They are different architectures.

The smallest useful formal test uses two interventions on a hidden variable while keeping the public input constant. Let S be the candidate’s internal state, R its typed self-report and A its action. A necessary causal signature is:

X is everything publicly available to the system, including prompt, transcript and tool response. The first inequality asks whether the state causes the report. The second asks whether it causes control. The assumption is that the intervention changes only the intended state or that collateral effects are measured. Remove that assumption and the result may reflect generic disruption rather than self-access.

How to interpret the lab

A positive result at the “causally grounded self-access” level permits one statement: a declared candidate uses a private, self-indexed variable in producing a self-report, after matched public-cue explanations have been weakened. The “reflexive-control” level adds temporal recognition and action efficacy. Replication strengthens confidence that the result is not an artefact of one model, prompt or intervention.

A negative result has two possible meanings. The candidate may lack the tested relation, or the intervention may be insensitive. For example, swapping hidden states could disrupt many computations at once. Failure to report the swap would then be ambiguous. The correct response is to improve the causal isolation, not to announce that machine reflexivity is impossible.

The instrument encodes three assumptions. The candidate boundary is meaningful. Interventions can be made selective enough to support causal interpretation. Functional self-access is worth measuring independently of phenomenal consciousness. The instrument cannot establish first-person presence, intrinsic luminosity, universal consciousness, freedom in the Śaiva sense or liberation. Its strongest output is an architecture claim with a versioned evidence trail.

A minimal experimental protocol

Begin with one consequential variable, not a general request for introspection. Freeze the public input and record the unperturbed report and action. Intervene on the private variable at two or more values, then repeat with a sham intervention of equal computational cost. Build an input-only predictor and an external-monitor mimic. Swap the variable between matched candidate instances to test self-indexing. Interrupt and resume the system with selected state channels removed. Finally, verify that the same variable changes a typed action gate rather than only prose.

Pre-register the claim ceiling before running the test. A report effect alone permits a report-dependence claim. A joint report and action effect permits a control claim. Self-indexing and rival failure permit a candidate self-access claim. Replication permits confidence in that architecture. At no stage does the protocol silently change its target to experience. This sequencing makes a null useful: it identifies the first missing relation instead of collapsing every failure into “not self-aware”.

function classifyRRS(evidence) {
  const causal = key => evidence[key] >= 2;
  const replicated = key => evidence[key] >= 3;
  const missing = Object.entries(evidence)
    .filter(([, level]) => level === 0)
    .map(([key]) => key);

  if (!causal("boundary") || !causal("privileged")) {
    return {
      claim: "Insufficient evidence",
      missing,
      permits: "No self-access mechanism claim"
    };
  }

  const grounded = causal("selfIndexing") && causal("rivals");
  if (!grounded) {
    return {
      claim: "Boundary-specific causal sensitivity",
      missing,
      permits: "The candidate responds to an internal intervention"
    };
  }

  const control = causal("temporal") && causal("control");
  if (!control) {
    return {
      claim: "Causally grounded self-access candidate",
      missing,
      permits: "Private, self-indexed access under matched controls"
    };
  }

  const allReplicated = Object.keys(evidence).every(replicated);
  return {
    claim: allReplicated
      ? "Replicated reflexive-control candidate"
      : "Reflexive-control candidate",
    missing,
    permits: "Self-access causally changes temporal uptake and action"
  };
}

The code exposes the non-compensatory logic. Excellent performance on one requirement cannot cancel a missing boundary, absent control effect or untested rival.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The identical warning
  M->>E: The second lamp
  E-->>M: compare rival explanations
  E->>D: Require private state to govern what the system may do
  D-->>Q: revise scope or act

Return to the two payment systems. Both said the right sentence. Before examining Pratyabhijñā, it was tempting to treat the difference between them as an implementation detail. After the analysis, it becomes the entire question. One system received a description of itself from elsewhere. The other used a private condition to restrain its own action. Only the second supports a causal self-access claim, and even that claim stops far below awareness in the Śaiva sense.

The tradition changes the question because it refuses to let illumination do the work of sentience. Information can be present in a crystal, film, log or activation. Vimarśa asks how manifestation is taken up within the very life of awareness. Recognition then asks how separated appearances are bound to one knower and become practically effective. In Pratyabhijñā, this culminates in the recognition that the apparently finite knower is Śiva, whose freedom expresses the world. The machine instrument preserves only a structural pressure: self-relation must be constitutive and efficacious, not decorative.

Different philosophies will interpret a successful RRS result differently. A functionalist may treat the causal organisation as a serious building block of mentality. A biological naturalist may grant the function while denying that the relevant living substrate has been reproduced. An emergentist may ask what scale and recurrence make the relation system-level. A physicalist can seek a complete mechanistic account without accepting intrinsic luminosity. Idealism, panpsychism, neutral monism, dual-aspect theory and process philosophy may regard awareness or experience as more basic than the instrument can register. The RRS does not decide among these positions because all can agree on the functional data while disagreeing about what constitutes experience.

For system builders, the changed decision is concrete. Do not fund an “introspection feature” whose evidence is a first-person paragraph. First declare the candidate boundary. Identify the private state. Prove that the report depends on it under intervention. Show that the state is indexed to this candidate and episode. Test continuity across reset. Connect the state to a reversible control action before attempting an irreversible one. Build matched mimics. Record the strongest permitted claim and the prohibited inferences.

For researchers of consciousness, the decision is equally concrete. Do not use the absence of a consciousness test as a reason to ignore functional self-relation. It is a tractable property with consequences for calibration, oversight and agency. But do not turn progress on that property into a verdict about experience. The gap is not intellectual timidity. It marks where a new measurement bridge is still missing.

For readers of Pratyabhijñā, the discipline runs in the other direction. Do not shrink vimarśa into metadata, a self-attention head or a confidence score. The source claim concerns the nature of manifestation, the unity of cognition, the power of action and the recognition of universal consciousness. The engineering comparison earns its place only by returning with a sharper question and an explicit residue.

Change the release decision from “the system can report its state” to “the declared candidate’s private state survives rival controls and causally governs what it may do”. Treat the resulting evidence as an architecture result. Treat awareness as an open question. Treat recognition, in its Śaiva meaning, as the far stronger claim that the light by which anything appears has never been outside itself.