A source is a causal route, not a vote

The Sanskrit term pramāṇa is often translated as a means or source of knowledge. That translation is useful only if we retain its sharper edge. In much classical Indian epistemology, a successful pramāṇa does more than raise confidence. It produces a veridical cognition, pramā, through an appropriate route. A lucky true guess lacks the right pedigree. A hallucination is not a defective instance of successful perception in the technical sense; it is a cognition that fails to qualify as pramāṇa-generated knowledge.[1]

This makes the traditions unusually helpful for a modern problem. We routinely use the word evidence for anything that moves belief: a sensor value, a witness statement, a feeling of certainty, a statistical association, an expert interpretation or a contemplative report. Once everything becomes evidence, it is tempting to combine it by weight. Pramāṇa theory asks a prior question: what is the process by which this cognition could be right rather than accidentally true?

Classical schools answer differently. Nyāya lists perception, inference, comparison and testimony in its opening taxonomy.[2] Yoga Sūtra I.7 gives perception, inference and āgama or authoritative transmission.[3] Buddhist epistemologists in the Dignāga-Dharmakīrti lineage recognise perception and inference as the two basic reliable cognitions, often analysing verbal knowledge through inferential structure.[4][17][18] Many later Mīmāṃsā and Advaita accounts defend further irreducible routes, including postulation and knowledge of absence.[16]

The differences are not a contest between narrow and open minds. They concern reducibility. Is testimony independent, or does it work because the hearer infers reliability? Is absence perceived, inferred, or known through a distinct operation? Does cognition disclose itself, or does a later awareness take the first cognition as its object? A source count is therefore a compressed causal theory.[15]

Why no universal ranking can work

Suppose someone asks whether perception is more reliable than testimony. The question lacks an object. For the colour of the indicator now in front of you, ordinary sight may be the better route. For the alloy inside a sealed component, a laboratory report may outrank unaided sight. For whether a colleague promised to complete a task, the colleague’s utterance is more than a noisy substitute for perception. It is partly constitutive of the promise. For the current sharpness of your own headache, a third-party instrument cannot simply replace first-person access.

This suggests the idea of epistemic jurisdiction. Jurisdiction is not social prestige. It is the bounded domain within which a route has the right kind of causal contact with its object. Jurisdictions can overlap. A person can feel breathlessness, a clinician can observe respiratory effort, and a sensor can measure oxygen saturation. Their agreement may support a broader diagnosis, but only through a model that relates felt sensation, visible behaviour and physiology. Their disagreement is not automatically evidence that one source is bad. It may reveal that the sources address different layers.

The strength of evidence is inseparable from the granularity of the claim. “There is a faint odour” is easier for smell to warrant than “a fire is developing”. “The participant reported stillness” is easier for an interview to warrant than “narrative self-processing ceased”. Precision in the claim is therefore not editorial tidiness. It determines whether the source and object are causally aligned.

Nyāya-style treatments of testimony commonly emphasise a competent, sincere speaker whose assertion is properly understood. Mīmāṃsā gives linguistic cognition, especially Vedic testimony, a different basis and resists making it depend in the same way on an ordinary author. The point here is not to settle that dispute. It is to see what the dispute exposes: the sentence alone does not tell us which causal theory of testimony is in play.

flowchart TD
  A["The silent alarm"] --> B["A source is a causal route,<br/>not a vote"]
  B --> C["Five channels, five native ranges"]
  C --> D["Where warrant leaks"]
  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 channels, five native ranges

Consider a wet courtyard at dawn. You see water on the stone. You infer that it rained. A neighbour says a pipe burst. You notice that you had already expected rain. During careful attention, you distinguish the glint of the surface from the thought “rain”. These are not five answers to the same question. They are routes with different native objects.

Routes are roles in a knowledge episode, not permanent labels attached to an artefact. A photograph can be the output of an instrument, the object of a viewer’s perception, part of an expert’s testimony and a premise in an inference. A contemplative journal can preserve a first-person report, become testimony for a researcher and later enter a statistical model. At every transfer, the relevant question changes from “what does this item contain?” to “what relation now connects this item, this knower and this claim?”

Perception: contact is selective

Perception seems closest to the world because it involves contact through a sense or instrument. Yet contact is always selective. Vision registers surfaces under lighting, not the whole object at once. A thermometer responds within a calibrated range, not to “fever” as a complete clinical condition. Indian debates sharpen this point further. Nyāya and Mīmāṃsā defend forms of realism about perceived objects; Buddhist accounts associated with Dignāga and Dharmakīrti place greater pressure on the relation between non-conceptual particulars and conceptual construction.[5][6]

The practical rule is modest. Perception directly supports what the channel could register under the conditions. From “water glints on the stone” to “it rained” lies an inference. That inference may be excellent, but naming it matters because a burst pipe can defeat it. Directness attaches to a relation, not to the confidence of the observer.

Inference: the bridge is the claim

Inference begins where direct access ends. Smoke indicates fire only because a relation between smoke and fire is assumed, learned or established, while relevant counterexamples are excluded. Modern models do the same work with different machinery. A risk score links observed variables to a hidden or future condition through a fitted relation. The output can be precise while the relation fails after a shift in population or environment.

In the courtyard, the smallest useful calculation is not a formula but a rival. Rain predicts a broad wet surface, cloud cover and perhaps water beyond the pipe’s reach. A leak predicts a local gradient around the pipe. Alter one observation and the preferred explanation changes. Inference is strengthened by interventions and discriminating consequences, rather than only by adding more observations that every rival predicts.

This is why the bridge itself deserves a record. “Smoke therefore fire” contains a relation learned from past cases. “A pattern of reports therefore a universal structure of consciousness” contains a much larger relation, spanning language, culture, training and ontology. Both are inferences, but their epistemic distances differ. The second demands more than a larger sample. It needs invariance tests across elicitation methods, traditions, populations and plausible rival constructions.

Testimony: knowledge can travel, but provenance travels with it

No individual can perceive every historical event, inspect every aircraft or reproduce every proof. Testimony is not an epistemic embarrassment. It is a condition of cumulative knowledge. Its native power is transmission: a proposition can reach someone who lacks direct contact with its object. But transmission introduces new relations. Did the speaker know? Was the statement meant literally? Did the hearer understand it? Is the claim still within scope?

The engineer’s statement about insulation may warrant action because her competence and inspection connect the words to the cable. A later summary, “engineering confirmed a fire risk”, adds an interpretation. A machine-generated digest that omits the inspection date may preserve the words while losing temporal entitlement.

Competence is also claim-specific. A renowned scholar may be an excellent source for what a text says and a poor source for whether a physiological mechanism exists. A meditation teacher may competently guide a practice while lacking the data needed to generalise its effects. An instrument vendor may accurately describe a device specification while overstating what a derived metric means. The title “expert” cannot travel farther than the relation that made the person expert.

Long testimony chains create a second problem. Each link can preserve the sentence while changing its force. A direct observation becomes a note; the note becomes a review; the review becomes a policy; the policy becomes a model’s retrieved context. Unless each transformation carries object, time, scope and source, the final reader receives language without entitlement. Testimony is not weakened by having provenance; provenance is part of how it works.

Introspection: content is not process

Introspection is not one agreed category in Indian thought. Nyāya analyses a later awareness of a prior cognition, often called anuvyavasāya. Mīmāṃsā schools develop rival accounts of how cognition presents or certifies itself. Buddhist philosophers dispute reflexive awareness, svasaṃvedana, and its implications.[1][4] Translating all of this as one faculty would erase the mechanism under debate.

A modern practical distinction helps. A person may have strong access to what is presently felt, seen in imagination or entertained as a thought. Access to the process that caused a choice is weaker. Classic work on verbal reports argued that people can be unaware of influential variables and then offer plausible causal explanations.[7][14] Later criticism rightly resists turning this into global scepticism about experience.[8] The better conclusion is typed: “I feel pressure in my chest” and “I chose this option because of the pressure” are different claims.

A useful first-person protocol therefore separates description from explanation. Description asks about qualities, temporal order, attention, bodily location, imagery and felt significance. Explanation asks why the event occurred, which process generated it and what it implies. The participant may contribute to both, but the second class should be marked as interpretation unless an additional route supports it. This separation does not demean the participant. It prevents a scientifically familiar vocabulary from being mistaken for direct inner inspection.

Memory adds another layer. A report produced during an event and a report reconstructed a week later are both first-person testimony, but they are not the same access episode. Repeated questioning can stabilise a narrative while moving it away from the original event. Conversely, careful elicitation may recover distinctions that an unstructured question misses. The relevant issue is not whether memory is trustworthy in general. It is which features of the event the procedure can still discriminate.

Contemplative inquiry: training changes resolution, not jurisdiction

Contemplative practice can change what a participant notices, how steadily attention is held and how finely experience is described. Neurophenomenology and micro-phenomenology treat disciplined first-person methods as capable of constraining research rather than as private decoration.[10][11][12] Yet “contemplative inquiry” is a modern umbrella, not a universally independent classical pramāṇa. Yoga’s own opening list names perception, inference and transmission. Buddhist accounts may place cultivated or yogic cognition under perception after stages of hearing, reflection and practice.

Tradition-internal discipline matters as well. A Yoga practitioner, an Advaita inquirer and a Buddhist meditator may use similar English words while training different attentional operations, accepting different accounts of error and pursuing different ends. “Witness”, “non-self” and “pure awareness” cannot be merged into a neutral latent variable merely because reports share terms such as silence or spaciousness. A serious comparison must preserve the practice instruction, the expected transformation, the tradition’s own defeaters and the point at which translation fails.

Training can improve a channel without making it unlimited. A randomised two-week study found improved metacognitive accuracy for memory but not perception, a useful warning against a general expert-introspection bonus.[9] Cross-sectional findings linking meditation experience with introspective accuracy are suggestive but cannot by themselves separate training from selection.[13] Greater experiential resolution does not grant causal or metaphysical jurisdiction.

The strongest use of contemplative inquiry is therefore comparative and generative. It can refine the vocabulary of experience, identify transitions that coarse questionnaires miss and produce predictions for behavioural or physiological study. It is weakest when the conclusion is already supplied by doctrine and the report is treated as confirmation. A practice can be transformative for a person while remaining underdetermining as evidence for a public ontology.

Where warrant leaks

Most mixed-method errors are not false observations. They are valid observations moved into a claim they cannot discriminate. Four leaks recur.

Claim drift turns “I experienced boundarylessness” into “the self-model switched off”. Dependency laundering treats several outputs derived from one source as independent corroboration. Authority stacking combines expert status, technical vocabulary and institutional repetition until the original bridge disappears. Category conversion turns a report about experience into a claim about what reality ultimately is.

These leaks often hide inside agreement. Imagine five papers that all classify an experience as “non-dual”. The first develops the coding scheme from interviews with one teacher’s students. The second trains raters on examples from the first. The third uses those ratings as labels for a classifier. The fourth applies the classifier to a new recording. The fifth asks a language model to synthesise the literature. Five documents now agree, but the agreement may trace back to one vocabulary and one initial judgement. This is a lineage, not five independent confirmations.

Conflict can be more informative than agreement. If a participant reports calm while physiology shows arousal, at least four possibilities remain: felt calm coexists with arousal; the physiological metric is non-specific; the report is reconstructed; or the measures refer to different intervals. Declaring one source authoritative too early destroys the chance to distinguish them. A typed conflict record keeps the discrepancy alive long enough to design the next observation.

This experiment does not imply that the practitioner is right about metaphysics. It asks for a mechanism-sensitive difference within the domain of report. Can the person distinguish onset from recognition, sensory attenuation from reduced identification, or attentional stability from retrospective narrative? Are those distinctions reproducible under varied prompts and linked to independent behavioural or physiological changes? A positive result strengthens a claim about trained experiential discrimination. It still does not prove that awareness is fundamental, emergent or reducible.

stateDiagram-v2
  [*] --> Claim
  state "Thought experiment 1: change<br/>one channel, hold the<br/>conclusion fixed" as Claim
  state "Why no universal ranking can work" as Test
  state "Thought experiment 2: the<br/>perfect witness transplant" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build the bridge before the conclusion

The Local Warrant Principle becomes useful only when it changes a workflow. The Warrant Bridge Protocol does this with seven fields.

1. Target claim. Write one proposition precise enough to be contradicted. “Meditation changes the self” is too broad. “During this task, trained participants report less narrative elaboration after the cue” is testable.

2. Native route. Name the route that directly accesses the claim. A report may be native to current lived content. An intervention is native to a causal contrast. A text is native to what it states within its transmission conditions.

3. Access conditions. Record calibration, competence, attention, elicitation, timing and scope. A route name without operating conditions is not a pedigree.

4. Bridge. State the assumption that carries the evidence beyond its native range. For example: “Reduced first-person narrative density indexes reduced activity in a specified self-referential process.”

5. Independent bridge test. Design an observation that could fail even when the original evidence remains. If the bridge predicts a behavioural effect, manipulate the proposed process or compare a rival mechanism.

6. Dependency graph. Mark shared participants, labels, prompts, models and source documents. Agreement among dependent outputs is consistency, not replication.

7. Decision limit. State the action the evidence permits. A study may justify refining a protocol, not announcing a universal theory of mind.

Run the protocol on the aircraft

The target claim is not “something is wrong”. It is “an electrical fault is sufficiently plausible that diversion is the safer action”. Smell is native to the presence and character of odour. The panel is native to sensor readings within its detection envelope. The maintenance message is testimony about prior work. Radio crackle supports an inference only through a model connecting interference to electrical failure. Dread is native to the captain’s current experience. The contemplative check supports a finer claim that imagined smoke modulates the dread.

The crucial bridge links these source-level facts to hidden fault and operational risk. An independent bridge test might isolate power to the suspected galley circuit, compare current draw, inspect for additional heat or ask another crew member who has not heard the first report to check the area. Shared dependence must be marked. If the panel and maintenance message both rely on the same diagnostic system, their agreement is weaker than it appears. The decision limit can remain precautionary: diversion may be justified under asymmetric harm even before fault is established. The protocol does not force certainty before action. It forces clarity about what is known, inferred and risk-managed.

Five designs for an independent bridge test

An intervention test changes the proposed cause while holding the reporting method stable. A replacement test produces the same report through a rival mechanism and asks whether downstream predictions differ. A blind prediction test commits the bridge to an outcome before the independent route is observed. A negative-control test uses a condition that shares nuisance factors but lacks the proposed mechanism. A transport test moves the bridge to a new population, practice, instrument or environment. None is universally best. The right test is the one that could preserve the original evidence while making the bridge fail.

This last condition is essential. Repeating the same questionnaire, retraining a classifier on the same labels or asking another expert steeped in the same doctrine may increase consistency without testing the bridge. Independence is not physical separation. It is the possibility of a meaningful disagreement generated through a different causal path.

function evaluateWarrant(input) {
  if (!input.conditionsPassed || input.activeDefeater) return "DEFEATED";
  if (input.typedConflict) return "CONFLICT_REQUIRES_PARTITION";

  const native = nativeClaims[input.channel].includes(input.claim);
  if (native) return input.sharedDependency
    ? "NATIVE_WARRANT_DEPENDENCY_FLAG"
    : "NATIVE_WARRANT";

  if (input.bridgeExplicit && input.bridgeIndependentlyTested) {
    return input.sharedDependency
      ? "BRIDGED_PROVISIONAL_DEPENDENCY_FLAG"
      : "BRIDGED_PROVISIONAL";
  }
  return "REPORT_ONLY";
}

The artefact encodes three assumptions. First, native ranges are declared before the result. Second, a bridge test is independent enough to fail while the original evidence remains. Third, dependency flags reduce the force of corroboration but do not erase the source records. A positive result permits the bounded statement shown. A negative result means the claim should be narrowed, the bridge tested or the decision deferred. The tool cannot determine historical truth, contemplative attainment, neural causation or metaphysical reality.

Let the best-tested bridge set the claim width

When several ways of knowing converge, do not ask first whether the evidence is “strong”. Ask what each route directly knows, what claim you are trying to make and which bridge joins the two. This changes how a pilot handles an alarm, how a scientist combines reports and physiology, how an organisation reuses expert testimony and how a contemplative tradition enters public inquiry.

The practical decision is simple: do not release a claim at a wider level than its best-tested bridge permits. Preserve perception as perception, inference as inference, testimony with provenance, introspection as typed first-person evidence and contemplative inquiry as disciplined but bounded access to experiential structure. Then let the routes constrain one another through interventions, not through prestige or arithmetic alone.

This discipline is neither scepticism nor relativism. Scepticism would flatten every route into doubt. Relativism would allow each route its own untouchable truth. Local warrant does the opposite. It grants each method real authority where its causal access is discriminating, and it makes every extension answerable to a test.

Glossary

Pramāṇa

A source, means or episode of veridical cognition, analysed differently across Indian traditions.

Pramā

A successful cognition or knowledge result, in the relevant technical usage.

Prameya

The object or matter to be known.

Āgama / śabda

Authoritative transmission or verbal testimony, with different accounts of authority across schools.

Anuvyavasāya

In Nyāya, a higher-order or subsequent cognition that takes a prior cognition as its object.

Svasaṃvedana

Reflexive awareness or self-cognition in contested Buddhist debates.

Native claim range

The class of claims a route can directly discriminate under stated conditions.

Bridge

An explicit assumption linking source-level evidence to a claim outside its native range.

Defeater

Information that undercuts the route, the evidence, or the link from evidence to claim.

Dependency laundering

Treating several outputs derived from one source or model as independent corroboration.

References

1. Phillips, Stephen. “Epistemology in Classical Indian Philosophy.” Stanford Encyclopedia of Philosophy. plato.stanford.edu/entries/epistemology-india/.
2. Gautama. The Nyāya Sūtras, especially 1.1.3. Public-domain edition and translation. archive.org/details/TheNyayaSutrasOfGotama.
3. Patañjali. Yoga Sūtra I.7, “pratyakṣānumānāgamāḥ pramāṇāni.” GRETIL electronic text. gretil.sub.uni-goettingen.de/.../sa_pataJjali-yogasUtra.htm.
4. Tillemans, Tom. “Dharmakīrti.” Stanford Encyclopedia of Philosophy. plato.stanford.edu/entries/dharmakiirti/.
5. Chadha, Monima. “Perceptual Experience and Concepts in Classical Indian Philosophy.” Stanford Encyclopedia of Philosophy. plato.stanford.edu/entries/perception-india/.
6. Matilal, Bimal Krishna. Perception: An Essay on Classical Indian Theories of Knowledge. Oxford University Press, 1986. global.oup.com/academic/product/perception-9780198239765.
7. Nisbett, Richard E., and Timothy D. Wilson. “Telling More Than We Can Know: Verbal Reports on Mental Processes.” Psychological Review, 84(3), 1977, 231–259. doi.org/10.1037/0033-295X.84.3.231.
8. Schwitzgebel, Eric. “The Unreliability of Naive Introspection.” The Philosophical Review, 117(2), 2008, 245–273. faculty.ucr.edu/~eschwitz/SchwitzAbs/Naive.htm.
9. Baird, Benjamin, Michael D. Mrazek, Dawa T. Phillips, and Jonathan W. Schooler. “Domain-Specific Enhancement of Metacognitive Ability Following Meditation Training.” Journal of Experimental Psychology: General, 143(5), 2014, 1972–1979. pubmed.ncbi.nlm.nih.gov/24820248/.
10. Varela, Francisco J. “Neurophenomenology: A Methodological Remedy for the Hard Problem.” Journal of Consciousness Studies, 3(4), 1996, 330–349. philpapers.org/rec/VARNAM.
11. Lutz, Antoine, and Evan Thompson. “Neurophenomenology: Integrating Subjective Experience and Brain Dynamics in the Neuroscience of Consciousness.” Journal of Consciousness Studies, 10, 2003, 31–52. .
12. Petitmengin, Claire, Martijn van Beek, Michel Bitbol, Jean-Michel Nissou, and Andreas Roepstorff. “Studying the Experience of Meditation through Micro-Phenomenology.” Current Opinion in Psychology, 28, 2019, 54–59. doi.org/10.1016/j.copsyc.2018.10.009.
13. Fox, Kieran C. R., et al. “Meditation Experience Predicts Introspective Accuracy.” PLoS ONE, 7(9), 2012, e45370. doi.org/10.1371/journal.pone.0045370.
14. Johansson, Petter, Lars Hall, Sverker Sikström, and Andreas Olsson. “Failure to Detect Mismatches between Intention and Outcome in a Simple Decision Task.” Science, 310(5745), 2005, 116–119. doi.org/10.1126/science.1111709.
15. Ganeri, Jonardon. Philosophy in Classical India: The Proper Work of Reason. Routledge, 2001. routledge.com/.../9780415240352.
16. Kumārila. Scholarly overview in “Kumārila.” Stanford Encyclopedia of Philosophy. plato.stanford.edu/entries/kumaarila/.
17. Dignāga. Pramāṇasamuccaya, perception chapter materials and translations. archive.org/details/dignagaonperception....
18. Hugon, Pascale. “Tibetan Epistemology and Philosophy of Language.” Stanford Encyclopedia of Philosophy. plato.stanford.edu/entries/epistemology-language-tibetan/.