The map begins with a change in relation

“Kashmir Shaivism” is a convenient umbrella, not the name of one homogeneous doctrine. It covers overlapping Śaiva and Śākta traditions, including Pratyabhijñā, Spanda, Trika, Krama and Kaula materials, with different genres and historical purposes. The present argument focuses on the mature nondual Pratyabhijñā synthesis associated with Utpaladeva, Abhinavagupta and Kṣemarāja. That scope matters because a metaphysical claim in a philosophical digest cannot simply be read as if it were a neutral engineering taxonomy.[1]

The Sanskrit word tattva can mean a principle, reality or “thatness”. In this setting, the thirty-six tattvas provide a graded account of manifestation. They can be read downward, from the most inclusive level of consciousness to material embodiment, or upward, from ordinary experience towards the conditions that make it possible. They are not thirty-six detachable substances. They organise powers, standpoints, cognitive apparatus and worldly elements within a tradition-internal account.

The thirty-six-level terrain

The familiar arrangement has three broad regions. The five śuddha, or pure, tattvas describe the nondual field through changing emphases of “I” and “this”: Śiva, Śakti, Sadāśiva, Īśvara and Śuddhavidyā. Māyā then marks differentiation. The five kañcukas and Puruṣa articulate the bounded experiencer. Below them, Prakṛti and the twenty-three further principles largely reuse the Sāṅkhya account of intellect, ego-making, mind, capacities of knowing and acting, subtle qualities and gross elements.

Some texts and modern charts vary the display order of the kañcukas, and Śaiva traditions do not all interpret the scheme identically. This paper uses the analytic order given in Kṣemarāja’s discussion of contracted powers: kalā, vidyā, rāga, kāla, niyati. The point is not the graphic order. It is the relation between an uncontracted power and its bounded appearance.[2]

Māyā is not simply a claim that the world is unreal

A common shortcut imports a simplified Advaita vocabulary and translates māyā as “illusion”. That misses the mechanism at stake here. In Kṣemarāja’s compact account, consciousness manifests a diverse world according to differences between grasper and grasped. The bounded experiencer is described as a contraction of consciousness that contains a contracted world. The world is not dismissed as a blank error. Differentiation becomes the form in which manifestation is encountered.[2]

This is why the five kañcukas cannot be treated as independent add-ons. Torella’s historical study argues that they express a central tantric account of human finitude rather than a minor list of doctrinal attributes.[3] In the mature nondual synthesis, the same powers that are unrestricted at the inclusive level appear in contracted form at the bounded level. The causal shape is power, contraction, local standpoint, not resource, cap, poorer performance.

flowchart TD
  A["Five shutters in an observatory"] --> B["The map begins with a change<br/>in relation"]
  B --> C["Five contractions, not five<br/>shortages"]
  C --> D["Five tempting bridges and<br/>their break points"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#5B5CF6,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Five contractions, not five shortages

Kṣemarāja gives an unusually compact relation. Universal agency, omniscience, fullness, eternity and all-pervasiveness appear, through contraction, as kalā, vidyā, rāga, kāla and niyati respectively.[2] Read that sentence slowly. The left side names an inclusive power. The right side does more than name less of a commodity. It names the bounded mode in which that power is lived or expressed.

Kalā: from all-agency to determinate doing

Suppose you can type, walk, calculate and speak. At this moment you type one sentence. Your agency is not a fuel tank with ninety-nine per cent left. It is determinate: this body, this action, these available means, this consequence. Kalā names the contraction of universal doership into partial agency, “doing some” rather than doing all.

A machine comparison must therefore distinguish computation from action. More floating-point operations can generate more proposals without granting one additional permitted effect. Conversely, a small controller may possess decisive authority over a valve. Compute contributes to capability, but bounded agency is shaped by affordance and authority as well as processing capacity.

Vidyā: from all-knowing to situated disclosure

Change one variable. Give a librarian access to only shelf A. She can reason perfectly about every book she can inspect, yet shelf B remains outside her available evidence. Now give her a catalogue entry for shelf B but deny access to the book. Her factual situation changes again. Partial knowing depends on standpoint, access, representation and warrant, rather than only storage size.

Vidyā is the contracted form of omniscience in Kṣemarāja’s sequence. It does not mean that every finite belief is false. It means that knowing is local and partial. For machine systems, model weights, prompt tokens, retrieved documents, authorised evidence and current world state are different reservoirs. Collapsing all of them into “context” makes the analogy easier and the mechanism worse.

Rāga: from fullness to lack and directed pull

Imagine a system that already has everything its specification requires. It can still assign probabilities and choose outputs, but there is no unresolved deficit inside the control loop. Now introduce a gap between a target state and an observed state. Behaviour becomes organised around closing that gap. The change is not simply more preference. It is a relation of incompleteness.

In this source structure, rāga is associated with contracted fullness. The bounded subject appears as lacking and is drawn towards particular completion. This is why equating rāga with a training loss is especially risky. A loss shaped parameters in the past. It need not be represented, active or experienced at runtime. A live error signal can be a better functional candidate, but neither signal establishes felt desire.

Kāla: from non-sequential availability to before and after

Listen to a melody. A vanished note remains relevant through retention; the present note occupies the point of action; the next note is anticipated but unavailable. Finite temporality is more than reading a clock. It organises what can still be changed, what has become fixed and what is only possible.

Kāla contracts eternity into sequence. An autoregressive decoder also emits tokens in sequence, but the resemblance is only partial. Transformer training can process positions in parallel, and a non-autoregressive architecture still operates under deadlines, state changes and irreversible effects. The relevant machine target would need a rule that makes past, present and future differently available to the configured system.

Niyati: from pervasiveness to this order and placement

Move the same agent from a sandbox to a regulated production environment. Its model and prompts remain identical. Yet it now acts as this identity, on this record, under this policy, from this jurisdiction, through this permitted dependency path. The agent is more than smaller. Its place in an order constrains what any action means.

Niyati is often rendered through order, necessity, causal determination, spatial placement or restricted pervasiveness. No single English word captures the field. The source relation is clearer: all-pervasiveness appears as localisation within a determinate order. A causal attention mask captures one information-flow rule. It does not by itself capture identity, location, law, dependency or necessity.

Worked example 2: a five-coordinate operatorA minimal formal model, and why it is insufficient

Let a five-coordinate vector represent candidate capacities: agency a, knowing k, fullness f, temporal availability t and pervasiveness p. Normalise an inclusive profile to one on each coordinate. A diagonal operator can then produce a bounded profile.

The result is b = [0.25, 0.15, 0.40, 0.20, 0.30]ᵀ. That is useful for an engineered resource profile. It is not yet a model of the kañcukas. The same vector could be produced by a permission policy, hardware scarcity, a damaged sensor, strategic abstention or a subject whose world is constituted differently. The output alone does not identify the operator.

Five tempting bridges and their break points

The obvious machine mapping is memorable: kalā as bounded compute, vidyā as finite context, rāga as loss objective, kāla as autoregression and niyati as causal masking. Its value is not that it is already true. Its value is that each row can be converted into a break experiment.

A disciplined comparison has five pieces. Name the source relation. Name the target mechanism. State the relation claimed to survive. Intervene on the target while preserving plausible rivals. Record the point at which the analogy fails. Without those pieces, a Sanskrit label merely redescribes a familiar component.

Kalā and bounded compute: retain the action relation

A compute budget limits search depth, model size or the number of proposals. That is a real mechanism. Yet a proposal is not an action. An external authority kernel may reject every proposal; a tiny embedded controller may directly change the world. The mapping survives only if “bounded compute” is widened into a configured action repertoire that includes tools, permissions, effect channels and reversal.

Hold the model and compute constant. Change only the typed actions and authority grants. If effective agency changes sharply, compute was not the load-bearing analogue. The result does not refute every kalā comparison. It locates the target relation in controlled action rather than raw processing.

Vidyā and finite context: distinguish access from knowledge

Language models operate under finite token windows, and long-context methods explicitly change how information persists or is retrieved.[11] This makes context an intuitively strong candidate. But a context window is an interface limit. The configured system may also possess model parameters, retrieval, a state store, tools and a human collaborator. A document outside the prompt can remain available through search. A document inside the prompt can remain unusable because policy forbids it.

The better target is permitted evidence at the decision point. Freeze the candidate system boundary, then move the same fact among weights, prompt, retrieval and authoritative world state. If the claim “the system knows” changes merely because the evaluator redraws the boundary, the comparison has not stabilised.

Rāga and loss objective: the historical-cause error

A training objective explains why parameters changed. It may not be a state the deployed system currently represents. A model can continue producing goal-directed behaviour after the training loss has vanished from runtime. Conversely, a controller can represent a live deficit without learning from gradient descent. Equating rāga with loss therefore confuses a historical cause of policy formation with an active relation of lack.

A stronger experiment compares three systems with matched output: one shaped only by past optimisation, one receiving a live error signal, and one following a fixed rule. Remove the signal. If only one system changes, the active discrepancy is causal. Even then, the conclusion is functional: an error variable organises action. It does not show felt attraction or dissatisfaction.

Kāla and autoregression: preserve temporal asymmetry, not token order

The original Transformer removed recurrence from its core architecture and gained parallelism during training, while autoregressive language models still generate one token after another.[9][10] That mixture already warns against equating a whole machine temporality with one decoding rule. Autoregression orders expression. It does not by itself determine what the system can revise, remember, anticipate or commit.

A more faithful target is transaction time. Before a commit, alternatives remain open. During execution, effects may become partially visible. After an irreversible effect, the past cannot be regenerated by sampling a different token. The machine analogue becomes interesting when state and authority make temporal positions causally unequal.

Niyati and the causal mask: broaden the candidate or reject the bridge

A causal mask prevents a token position from attending to future positions. It specifies an information dependency. Niyati concerns a bounded being’s placement in a determinate order. A more adequate machine target would combine identity, topology, jurisdiction, policy, state ownership and causal dependency. That target is no longer one tensor mask. It is a configured system relation.

Produce the same refusal through three routes: mask the information, omit it from context, or deny the action through policy. If the observed output is identical, output-only evidence cannot tell which mechanism mattered. Patch each route separately. The route whose intervention restores the action is the causal mechanism for that case. The Śaiva term should not be attached until that mechanism and its residue are explicit.

When a correspondence becomes structurally interesting

A functional correspondence says that one target mechanism performs a role sufficiently similar to the named source relation for a controlled comparison. A structural claim asks more. The relation should survive changes of implementation. A context-limited model, a recurrent agent and a tool-using system might realise partial evidence differently. If the proposed vidyā relation disappears whenever the storage mechanism changes, the comparison belongs to one implementation, not to the higher-order organisation.

A useful test is a commutation challenge. Start with a wider configured capacity, apply the proposed contraction and measure the bounded relation. Then translate the system into a second implementation and apply its corresponding operator. If both routes preserve the same declared dependencies, interventions and break conditions, the comparison gains structural support. If the routes agree only on output while internal authority, evidence or temporal relations diverge, the square does not commute in the sense required by the claim.

Even a successful result should be called a structural homology candidate, not formal equivalence. Formal equivalence would require specified objects, allowable transformations and a proof that the relevant relations are preserved. The comparison ladder rises through harder preservation tests, not through stronger adjectives.

stateDiagram-v2
  [*] --> Claim
  state "An open field narrows through<br/>five unlike apertures" as Claim
  state "Would every closure feel like<br/>the same limitation?" as Test
  state "The thirty-six-level terrain" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

A lab for making comparisons fail

A useful cross-domain comparison should be easier to reject after it is stated clearly. The comparison-break ledger below turns the source relation into an experimental obligation. It does not score consciousness. It classifies the warrant of one proposed mapping.

status: functional_correspondence
source_relation: named
target_mechanism: named
intervention: specified
candidate_boundary: declared
break_condition: explicit
ontology_leak: false

How to use the instrument

Begin with the source relation, not the preferred machine component. For vidyā, write “omniscience appears as partial knowing from a bounded standpoint”, then identify the target mechanism that might preserve partial availability. Declare the candidate boundary. Specify an intervention, such as moving one fact from prompt to retriever while holding the task constant. Name the break condition: if the conclusion depends only on the evaluator’s boundary choice, lower the mapping.

A “functional correspondence” result permits a matched experiment. It means enough terms are specified for the proposed relation to be wrong in an informative way. An “under-specified” result means the missing field must be repaired before interpretation. A “rejected” result means the claim has crossed from a structural or functional comparison into an unsupported inference of subjectivity or metaphysical identity.

function classifyComparison(c) {
  if (c.ontologyLeak || !c.sourceNamed || !c.targetNamed) {
    return "rejected";
  }
  if (c.interventionNamed && c.boundaryNamed && c.breakNamed) {
    return "functional_correspondence";
  }
  if (c.interventionNamed || c.breakNamed) {
    return "testable_analogy";
  }
  return "under_specified";
}

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: Five shutters in an observatory
  M->>E: An open field narrows through five unlike apertures
  E-->>M: compare rival explanations
  E->>D: Map the contraction relation, not the component label
  D-->>Q: revise scope or act

Map the contraction relation, not the component label

When a machine component resembles a kañcuka, do not ask, “Which Sanskrit term names this limit?” Ask a harder sequence. What inclusive power stands on the source side? What operator produces its contracted appearance? What bounded subject–world relation results? Which configured machine mechanism might preserve that relation? What intervention separates it from a cheaper explanation? Where does the metaphysical residue remain?

This changes comparative philosophy. The thirty-six tattvas no longer serve as a decorative ontology pasted over an architecture diagram. They become a source model whose internal causal commitments constrain the comparison. Māyā matters because it changes the grasper–grasped relation. The kañcukas matter because they are distinct forms of that contraction. Puruṣa matters because finitude is attributed to a bounded experiencer, not to a resource dashboard.

It also changes machine design. Compute is not authority. Context is not knowledge. A training loss is not a live desire. Token order is not the whole of time. A causal mask is not a complete account of placement and necessity. Keeping these differences visible improves system boundaries, evaluation plans and failure analysis even when every metaphysical comparison is rejected.

A passing result permits a bounded functional experiment. It does not permit an equivalence claim. A failed result is not wasted. It tells us whether the error lay in the source reconstruction, the target boundary, the intervention or the inference. That is more valuable than an analogy protected from failure by vagueness.

The observatory’s shutters remain useful, but only after we stop mistaking them for the sky. Some shutters merely reduce a resource. Some alter what the instrument can lawfully do. None, by itself, explains how there comes to be a standpoint for which the sky is partial. The comparison earns its place when it sharpens that missing question rather than pretending to have answered it.