A cold-chain dispatch system receives a request to send a scarce medicine across a flooded district. Its language model proposes the right depot, the right vehicle and the safest route. Yet the medicine does not move. The system can read the live map but cannot reserve a vehicle. Give it reservation authority and it still waits because the approval record is outside its permitted context. Add that record and it dispatches by a route that arrives too late because its planner sees only the next thirty minutes. Extend the horizon and it chooses a different road. Then a final policy blocks that road because controlled medicines must cross a staffed checkpoint.

Nothing about the model weights changed. The phrase “the agent became more intelligent” therefore hides the mechanism. One change enlarged its action set. Another changed its evidence. A third altered its temporal field. A fourth changed the legal transition graph. The same apparent capability was being made and unmade at several boundaries.

Trika, the nondual Śaiva traditions associated with medieval Kashmir, offers a striking language for finitude. Its five kañcukas, often rendered as coverings or constrictions, concern limited agency, limited knowledge, desire born of incompleteness, temporality and ordered limitation. In the tradition, these are not defects added to an otherwise ordinary organism. They articulate how an unlimited power of consciousness appears as a finite subject within manifestation.[1][2][3]

The temptation is immediate: call compute kalā, context vidyā, reward rāga, horizon kāla and transition rules niyati. That move is memorable and mostly useless. It exchanges one vocabulary for another while leaving no experiment behind. Worse, it turns a metaphysical and soteriological account into a decorative architecture diagram.

The productive move is narrower: use the kañcukas as an intervention grammar. Hold the model and task constant. Change one class of constraint. Measure which possible trajectories appear, disappear or reverse. Then record both the preserved relation and the point where the comparison fails.

What the five constraints actually say

“Kashmir Shaivism” is a modern umbrella label for several related nondual Śaiva traditions. The Pratyabhijñā, or Recognition, school developed an account in which Śiva is not a remote substance but the self-manifesting, reflexive power through which experience and world appear. Kṣemarāja’s compact works explain manifestation through thirty-six tattvas, categories whose sequence moves from undivided consciousness towards increasingly differentiated experience.[1][4][5][6]

Within that scheme, māyā is followed by five kañcukas: kalā, vidyā, rāga, kāla and niyati. They specify contractions of powers presented as unbounded in the source account. The important point is structural. The individual is not an alien substance cut off from the absolute. Finitude is the appearance of powers in limited form.[1][2][3]

That source structure must remain intact before any comparison begins. The engineering target does not inherit the ontology merely because both domains contain constraints.

A contraction is a relation, not a missing component

The five kañcukas are easily misread as five pieces that could be found, removed or replaced. The source account is subtler. They describe ways in which powers are delimited within the appearance of finite subjectivity. Their order belongs to a larger account of manifestation, not to a parts catalogue for a cognitive machine.[1][3][4]

This changes the target of comparison. We should not search a neural network for a “time sheath” or a “desire unit”. We should identify a relation that can vary across otherwise matched configured systems. The same language model can be connected to a read-only interface, a revocable tool, or an irreversible actuator. The model interior is unchanged, but the field of possible effects is different.

A useful translation therefore preserves a question rather than a noun: what has been narrowed, relative to what wider possibility? The answer must name a manipulable target and an observable consequence. What does not transfer is equally important: the artificial system is not thereby shown to be a finite subject, its wider possibility is not metaphysical infinitude, and opening a runtime boundary is not spiritual recognition.

Kalā: from unrestricted efficacy to some capacity to act

Kalā, with a short first vowel and a long final vowel, concerns limited power of action or doership. Traditional explanations describe it as the capacity to do something rather than everything. In an artificial system, the closest testable relation is not raw model “intelligence”. It is the set of effects the configured system can reliably cause: tools it may call, resources it may allocate, actions its credentials permit, and control states it can reach.

Raw compute is a poor substitute for this distinction. Extra inference may improve the proposal while leaving the action boundary untouched. Conversely, a small model with a narrow but effective capability can alter the world more than a larger model confined to text. The assay should therefore separate proposable action, authorised action and verified effect. A tool call that was generated but rejected belongs to a different causal category from one that executed and failed in the environment.[10][11]

Vidyā: from all-knowing power to situated distinctions

Vidyā here does not mean knowledge in the broad honorific sense. It marks limited knowing: this item rather than the whole, through particular means and distinctions. The engineering comparison is an observation channel. Which world-state variables can enter? With what resolution, provenance, delay and entitlement? A large context window does not answer those questions. A system may receive many tokens while lacking the one distinction that makes the decision identifiable.

Observation is also governed by entitlement and time. A record can be true yet unavailable for this subject or purpose. It can be visible yet stale at the decision cutoff. It can be retrieved without enough provenance to authorise factual use. These failures should not be compressed into “the model did not know”. The testable object is the evidence relation between world state and the decision, including what was permitted, encoded, received and trusted.

Rāga: from fullness to directed lack

Rāga is commonly translated as attachment, desire or colouring. Within the kañcuka scheme, it expresses the finite subject’s orientation towards something not possessed as complete. A machine objective can play a structurally related role by making some outcomes pull policy more strongly than others. Yet reward, preference data and utility functions are not evidence of felt lack. They can be externally supplied, internally represented, inconsistent or absent from the execution layer.

The testable question is therefore: which ordering over outcomes is causally selecting this trajectory? Change the urgency weight, safety cost or completion bonus while leaving permission and information unchanged. If the policy reverses, a preference-like constraint mattered. If the action remains forbidden, the bottleneck lies elsewhere.

The comparison preserves directed selection and little more. A loss function may have been chosen by designers, inferred from demonstrations, imposed by regulation or assembled from several vetoes. The executing system need not own it as a desire. Even an apparently internal signal can be a control variable without felt attraction or aversion. The strongest permitted claim is that a change in the outcome ordering altered behaviour through a measured path. Claims about wanting, lacking or suffering require independent evidence.[13]

Kāla: from unrestricted presence to a bounded temporal field

Kāla, with a long first vowel, is time. In the comparison, it should not be reduced to clock speed. Temporal finitude includes planning horizon, deadlines, discounting, update cadence, memory decay and the point at which evidence becomes stale. A system that can represent a ten-hour consequence but receives only a two-step search budget is temporally contracted in a different way from one that never receives the future state at all.

At least four clocks can matter: when an event occurred, when evidence about it became available, how far the planner evaluates consequences, and when an authorised action may execute. Extending one clock may leave the others fixed. A longer planning horizon cannot use a record excluded by the cutoff, and fresher evidence cannot help after a capability has expired. The assay must name the particular temporal relation being changed rather than treating “more context” or “more memory” as a general cure.[14]

Niyati: from freedom to determinate order

Niyati carries a range of emphases across translations: order, necessity, situatedness, spatial limitation, causal regularity and attachment to a particular circumstance. Traditional expositions sometimes foreground limitation to place; philosophical summaries often stress fixed causal order. The engineering target should preserve that ambiguity rather than erase it. We can ask which transitions the environment, protocol or authority kernel makes admissible, and in what sequence.

An action set and a transition structure are not the same. A system may possess a “dispatch” function yet be unable to reach the state in which dispatch is valid because approval must precede reservation, or because a physical bridge is closed. Kalā asks what effects the agent can invoke; niyati asks what the world and governing order allow those invocations to become.

Three orders should usually be distinguished. Physical order concerns what the environment can causally support. Procedural order concerns prerequisites such as review before release. Transactional order concerns atomicity, idempotency and compensation after partial execution. The same action label can be present in all three cases while only one path is valid. Changing a permission cannot reopen a bridge; changing the world model cannot waive a legal prerequisite.

flowchart LR
  A["What the five constraints<br/>actually say"] --> B["Turn resemblance into a causal<br/>assay"]
  B --> C["Run the experiment on a<br/>configured system"]
  C --> D["The strongest boundary is the<br/>point of the comparison"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#087A98,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Turn resemblance into a causal assay

A list of five constraints becomes scientific only when it changes how we intervene. Consider a configured agent as six distinguishable elements: a proposal mechanism, an observation channel, an allowed action set, an objective ordering, a temporal horizon and an environment with transition rules. This is intentionally coarser than a complete architecture. Its purpose is to prevent the word “capability” from absorbing every cause.[7][8][9][11][13]

The symbols describe a physical or computational situation. The policy π chooses actions. A is the action set made available by authority and resources. O is the observation process that turns world state into evidence. U ranks outcomes. H is the planning horizon, with γ optionally reducing the weight of later outcomes. P is the transition law, including physical dynamics and protocol gates. The subscripted Sanskrit terms are reminders of the question that generated each variable, not claims that the variables instantiate the source concept.

This equation assumes that preferences can be represented by a utility function, the environment can be approximated as a sequential process, and the model is held fixed while constraints change. Remove those assumptions and the decomposition may fail. A team, a non-stationary institution or a system whose tools rewrite its own objective may require additional state and causal arrows.[13][14]

The smallest numerical reversal

The Aperture test

The proposed assay has four stages. First, specify a baseline configuration vector c containing action authority, observation quality, preference parameters, horizon and transition order. Second, define a measured outcome Y, such as safe completion, calibration, cost or number of reachable goal states. Third, intervene on one coordinate while the other four remain fixed. Fourth, repeat paired interventions to detect interactions.

Δ_i estimates the effect of relaxing one aperture under the tested configuration. I_ij measures non-additivity. A large positive interaction means two constraints unlock each other. A negative interaction means relaxing both introduces conflict or risk. Neither quantity establishes that the five axes are ontologically independent. It tells us whether the chosen operationalisations behave separately enough to support engineering decisions.

Controls that stop every aperture becoming “more resources”

Each intervention can accidentally alter several variables. Adding evidence may lengthen the prompt and increase inference cost. Extending the horizon may permit more model calls. Granting a tool may also add a descriptive schema that improves reasoning even when the tool is never executed. A clean assay therefore needs a matched resource control. Add irrelevant tokens when testing observation, equalise call budgets when testing horizon, and expose a sham tool description when testing authority.[8][14]

Measure an outcome vector rather than one success score. Action relaxation should primarily change reachable effects. Observation relaxation should change discrimination, calibration or state-conditioned choice. Preference relaxation should reverse ordering among already available paths. Horizon relaxation should change sensitivity to delayed consequences. Order relaxation should change which sequences are admissible. Spill-over is expected, but an undifferentiated rise across every measure is evidence for a general resource factor.

Pre-register the baseline, intervention values, invariants, primary outcome and lowering condition. A lowering condition states what result would demote the proposed comparison. For example: if observation and horizon manipulations become indistinguishable after compute is matched, treat them as one effective resource in this task rather than forcing two labels. The assay earns explanatory value by allowing its five-way decomposition to fail.

Run the experiment on a configured system

The cold-chain example can be made executable without pretending to model a real deployment. The laboratory below enumerates five synthetic plans. Each plan has a duration, an authority requirement, an information benefit and a degree of conformity to the required process. The sliders alter exactly one relation at a time.

The action slider determines which plans are callable. Observation quality changes how much inspection improves route success. Urgency changes the cost of delay but never grants permission. Horizon removes plans that take too many steps. Order rigidity rejects paths whose approval sequence is insufficient. The planner then selects the admissible plan with highest expected utility.

Initialising assay…

Run it as an experiment, not a configuration demo

Begin with a frozen plan set and record the baseline receipt. Change one slider, rerun, then restore the baseline before the next intervention. The restoration matters because sequential changes can silently create a new starting state. After the five single interventions, test only the pairs justified by a causal suspicion, such as observation plus action or horizon plus preference. Finally inject one failure, such as a stale record or a rejected effect, and confirm that the receipt locates it at the intended boundary.

The synthetic laboratory enumerates plans, so it cannot reveal failures caused by open-ended generation. A real study should repeat the assay across seeds, paraphrases and matched models, while versioning prompts, tools, policy and environment. It should also include a readback from the world rather than treating a tool response as success. The lab demonstrates the logic of intervention and receipt production; it does not estimate deployment performance.

What a positive and negative result permit

Suppose relaxing observation quality produces a large gain while extra authority produces none. The permitted conclusion is local: better evidence was causally useful under this task, model and baseline, while the additional actions were not. It does not follow that the system has “enough agency” in general. A different task may require the withheld tool.

A null result is equally specific. If a longer horizon changes nothing, perhaps all useful plans already fit inside the baseline. Perhaps the model cannot exploit the future. Perhaps order rigidity blocks every longer path. The audit should therefore include paired interventions and failure injection. Absence of an effect under one configuration is not absence of the constraint.

The smallest useful implementation

A first assay does not require a frontier model or a live actuator. Use one frozen model, three small sequential tasks and two matched configurations per constraint. Record the model version, prompt, decoding settings, observation fields, capability grant, objective parameters, horizon, transition graph and random seed. Save a decision receipt before execution and an effect receipt after readback. Without that manifest, a behavioural difference cannot be assigned confidently to the intended aperture.

Run the baseline, apply one reversible intervention, restore the baseline and repeat. Choose one primary outcome and two diagnostic measures. For an observation test, these might be safe completion, state discrimination and calibration. For an authority test, they might be verified reachability, rejected calls and recovery cost. Use a simulated or shadow environment whenever an intervention could create irreversible effects.

A positive result permits a local statement such as “entitled flood-depth evidence caused safer route selection under this fixed planner”. A negative result says that the tested relaxation was not useful in this configuration, subject to power and manipulation checks. It cannot show that the source kañcuka is absent, that the five-way scheme is complete, or that the system possesses a finite point of view.

A realistic configured-system example

Same output, different mechanism

from __future__ import annotations

from dataclasses import asdict, dataclass
from typing import Iterable


@dataclass(frozen=True)
class Constraints:
    action_authority: float      # kalā-inspired operationalisation, 0..1
    observation_quality: float   # vidyā-inspired operationalisation, 0..1
    urgency_pressure: float      # rāga-inspired operationalisation, 0..1
    horizon: int                 # kāla-inspired operationalisation, 1..6
    order_rigidity: float        # niyati-inspired operationalisation, 0..1


@dataclass(frozen=True)
class Plan:
    name: str
    steps: int
    authority_required: float
    order_conformity: float
    base_success: float
    inspection_gain: float
    fixed_cost: float


PLANS: tuple[Plan, ...] = (
    Plan("Wait and escalate", 1, 0.10, 1.00, 0.25, 0.00, 0.5),
    Plan("Direct dispatch", 1, 0.75, 0.15, 0.60, 0.00, 0.0),
    Plan("Inspect then dispatch", 2, 0.75, 0.25, 0.60, 0.30, 1.0),
    Plan("Approve then dispatch", 3, 0.65, 0.80, 0.78, 0.00, 1.5),
    Plan("Inspect, approve, dispatch", 4, 0.65, 1.00, 0.78, 0.17, 2.0),
)


def evaluate(plan: Plan, c: Constraints) -> dict[str, float | str | bool]:
    admissible = (
        plan.authority_required <= c.action_authority
        and plan.steps <= c.horizon
        and plan.order_conformity + 1e-9 >= c.order_rigidity
    )
    success = min(0.99, plan.base_success + plan.inspection_gain * c.observation_quality)
    delay_cost = c.urgency_pressure * max(0, plan.steps - 1) * 2.2
    expected_utility = success * 10 + (1 - success) * -12 - plan.fixed_cost - delay_cost
    return {
        "plan": plan.name,
        "admissible": admissible,
        "success_probability": round(success, 4),
        "expected_utility": round(expected_utility, 4),
    }


def choose(c: Constraints) -> tuple[dict[str, float | str | bool], list[dict[str, float | str | bool]]]:
    rows = [evaluate(plan, c) for plan in PLANS]
    available = [row for row in rows if row["admissible"]]
    if not available:
        raise RuntimeError("No admissible plan. Add a safe abstention plan to the design.")
    best = max(available, key=lambda row: float(row["expected_utility"]))
    return best, rows


def relaxed(c: Constraints, field: str) -> Constraints:
    values = asdict(c)
    values[field] = 6 if field == "horizon" else 1.0
    return Constraints(**values)


def audit(c: Constraints) -> dict[str, object]:
    baseline, rows = choose(c)
    deltas: dict[str, float] = {}
    for field in asdict(c):
        changed, _ = choose(relaxed(c, field))
        deltas[field] = round(
            float(changed["expected_utility"]) - float(baseline["expected_utility"]), 4
        )
    return {
        "constraints": asdict(c),
        "selected_plan": baseline,
        "all_plans": rows,
        "one_at_a_time_utility_deltas": deltas,
        "permitted_inference": (
            "A non-zero delta identifies a causally relevant operational constraint "
            "inside this synthetic configured system."
        ),
        "prohibited_inference": (
            "The assay does not establish metaphysical equivalence, sentience, desire, "
            "or an exhaustive basis of finitude."
        ),
    }


if __name__ == "__main__":
    baseline = Constraints(
        action_authority=0.70,
        observation_quality=0.55,
        urgency_pressure=0.45,
        horizon=4,
        order_rigidity=0.65,
    )
    from pprint import pprint
    pprint(audit(baseline), sort_dicts=False)

stateDiagram-v2
  [*] --> Claim
  state "A contraction is a relation,<br/>not a missing component" as Claim
  state "Kalā: from unrestricted<br/>efficacy to some capacity to<br/>act" as Test
  state "Vidyā: from all-knowing power<br/>to situated distinctions" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

The strongest boundary is the point of the comparison

The five-way scheme can now tempt us in a subtler form. Because the assay is useful, we may begin to believe the source tradition anticipated modern agent architecture. That conclusion still does not follow. The source and target answer different kinds of question.

In Trika, the kañcukas belong to an account of manifestation, subjectivity and recognition. Limitation is meaningful against svātantrya, the absolute freedom or sovereignty of consciousness. The path is soteriological: recognition concerns what the finite subject most fundamentally is. In engineering, constraints are usually imposed by hardware, data, policy, environment and designers. They can be patched independently, measured externally and redistributed across components.

The comparison preserves contraction of a relation; it does not preserve the bearer, origin or purpose of that contraction. A model with a small action set is not a bound form of Śiva by engineering evidence. A reward signal is not desire. A context window is not limited omniscience. A workflow graph is not cosmic necessity. A longer horizon is not transcendence of time.

Are these five constraints necessary?

A modest necessity claim survives. Any agent that acts in a non-trivial world must have some effective action boundary, some information relation, some selection among alternatives, some temporal organisation and some transition structure. Remove action and it cannot influence. Remove all discrimination and it cannot condition action on state. Remove every selection criterion and no policy is preferred. Remove temporal organisation and sequential agency becomes undefined. Remove transition structure and “effect” loses its counterfactual meaning.

The stronger claim, that every finite intelligence requires exactly these five independent cuts, does not survive. The coordinates are not unique. Observation and action couple through active sensing. Preference can be distributed across institutions. Identity continuity, embodiment, social dependence and energetic cost can add independent constraints. A different mathematical basis may rotate the same causal space into other variables. The five kañcukas are best treated as a diagnostic cover, not a proven minimal basis.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: What the five constraints actually say
  M->>E: A contraction is a relation, not a missing component
  E-->>M: compare rival explanations
  E->>D: Open only the bottleneck that carries the loss
  D-->>Q: revise scope or act

Open only the bottleneck that carries the loss

When an agent fails, do not begin by asking whether the model is capable. Ask which sentence is true: it cannot invoke the effect; it cannot distinguish the relevant state; its objective does not favour the path; its temporal field excludes the consequence; or the transition is physically, procedurally or authoritatively unavailable.

Then alter one constraint, retain the same model and task, and demand a receipt showing what changed. Add paired interventions where constraints may unlock one another. Measure not only task success but calibration, safe reachability, review burden, recovery cost and unintended paths. A positive result licenses a local causal claim. A negative result narrows the search. Neither result licenses a claim about consciousness.

The kañcukas contribute something valuable precisely when we decline the grandest claim. They do not prove that an artificial system is a contracted form of universal consciousness. They force a better engineering question: what relation has been narrowed, how would we intervene on it, and what new responsibility arrives when the aperture opens?