flowchart TD
  A["The beep that arrived late"] --> B["Bracketing the hidden claim"]
  B --> C["Variation finds structure, not<br/>certainty"]
  C --> D["Making a private event<br/>publicly criticisable"]
  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

Bracketing the hidden claim

Husserl begins from a simple difficulty. Ordinary perception is transparent to its object. We look through seeing towards the cup, through hearing towards the melody, and through remembering towards the meeting. In what he calls the natural attitude, the world is not usually experienced as a hypothesis. It is simply the field in which our concerns proceed.

Epoché changes the attitude. In Ideas I, Husserl describes bracketing as putting a thesis “out of action” while leaving it present. The observer does not convert belief into disbelief. She refrains from using the belief as a premise in the current description. [1] Contemporary scholarship likewise treats epoché and reduction as a redirection towards the correlation between an experience and the manner in which its object is given, not as a denial that anything exists. [2]

Return to Nadia. “The second tone was delayed” combines at least three claims. There was an auditory appearance. It was organised as a second item. The physical timing was delayed. Only the first two belong to the appearance as reported. The third is a claim about the stimulus. Epoché asks Nadia and the researcher to suspend the third while describing the first two.

Bracketing is a claim-boundary operation. It changes “what happened in the world” into “how the event was given in this occurrence”, and it changes “what caused my choice” into “what sequence, tension or transition I could discriminate”. This reduction in claim width is the first mechanism of reliability. A narrower claim has fewer hidden commitments and more visible failure points.

Three kinds of leakage

The practical reconstruction proposed here treats phenomenological discipline as control over three leakages.

World-claim leakage occurs when a manner of appearance becomes a fact about the external object. “The room became dangerous” may be a report of threat salience, not evidence that the room changed. Mechanism leakage occurs when an experience is described through a causal theory. “My prediction error increased” is not a first-person observation. Generality leakage occurs when one occurrence becomes a universal structure. “Anxiety always begins in the chest” exceeds the case.

Epoché directly controls the first two. It asks the observer to retain the external and causal theses as bracketed commitments. It cannot by itself control the third. For that, the method needs variation.

Variation finds structure, not certainty

A single careful description remains a single case. Husserl’s answer is eidetic variation. Hold an experience or object in view, vary aspects freely in imagination, and ask what can change without the phenomenon ceasing to be the kind of thing under investigation. In Husserl’s project, the aim is an eidos, an invariant essential structure rather than an empirical average. [1][3]

Take the cup. Change red to blue. It remains a cup. Change ceramic to steel. Still a cup. Rotate it so only a profile appears. Still the same object through a different presentation. Remove every possibility of further profiles and every practical unity across movement. At some point the intended object is no longer a perceived spatial thing but a flat sign or an imagined token.

The reader’s intuition is that whatever survives is the essence. The hidden assumption is that our variations have ranged over every relevant possibility and that imagination is a neutral generator. Neither follows. A participant may be unable to imagine an unfamiliar sensory organisation. A culture may stabilise one concept of agency. An interviewer may decide in advance which property counts as the same phenomenon.

The safe empirical move is a modal downshift: call the result a variation-relative structural constraint, not a necessary essence. This preserves the method’s power while lowering the claim to what the variation set actually tested.

Variation is not ordinary sampling. Sampling asks how often a feature occurs in a population. Variation asks what happens to a candidate relation when one condition is altered while the identity of the target is provisionally held fixed. The hard work lies in that proviso. If changing visual presentation into touch changes the object from “perceived cup” to “recognised utensil”, the comparison may have crossed phenomena rather than tested one. The worksheet must therefore record both the altered property and the reason the case still belongs to the comparison.

The most informative variation is often the one that breaks the claim. A sequence that survives colour, material and viewpoint but fails when exploratory movement is prevented suggests that sensorimotor availability, not visible shape alone, carries the unity. That is not yet a universal truth about objects. It is a design clue for the next case. A break condition converts imaginative agreement into an experimental question. It tells a researcher which manipulation should dissolve the reported structure, and it gives a rival account somewhere to win.

Here, V is the chosen set of variations. Each d_v is the description generated under one variation. The property p might be “the object is given through incomplete profiles”. If p survives every member of V, it is invariant relative to that set. The equation assumes that the variations are admissible, the descriptions are not prompted into agreement, and the property has been coded consistently. Remove those assumptions and the result becomes a record of the procedure, not of the phenomenon.

Why another person matters

Experience remains first-person in access. The method becomes public through the conditions under which a description is elicited, transformed and challenged. Another researcher can inspect the question sequence, repeat the variation, propose a counterexample, recode the transcript, or test whether the distinction predicts an independent perturbation.

This is not the same as voting on what Nadia felt. Agreement can be manufactured by shared training. Disagreement can arise from language rather than experience. Intersubjective criticism earns its value when it exposes dependencies: Did the same distinction arise under different interviewers? Did participants use their own terms before the codebook was revealed? Does the proposed structure survive a condition that changes attention but leaves the stimulus constant?

Methods such as micro-phenomenological interviewing and Descriptive Experience Sampling try to improve specificity through concrete episodes, non-leading probes, temporal anchoring and iterative training. [6][7] Their existence does not settle reliability. A study of untrained micro-phenomenological self-inquiry failed to replicate an earlier reliability result and argued that training may be necessary. [12] The lesson is methodological, not tribal: first-person methods need their own error models, controls and negative results.

stateDiagram-v2
  [*] --> Claim
  state "Three kinds of leakage" as Claim
  state "Why another person matters" as Test
  state "What reciprocal constraint<br/>really means" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Making a private event publicly criticisable

The practical value of phenomenology appears before theory testing. It helps determine what the task actually asks a participant to do. A button labelled “saw it” may combine detection, confidence, object recognition and willingness to report. A meditation instruction labelled “focus on breath” may be enacted as narrow fixation, rhythmic tracking, bodily monitoring or suppression of distraction. A model trained on the resulting labels inherits the ambiguity.

A current example makes this concrete. Cavaletti and Heimann used a small set of micro-phenomenological interviews to examine whether participants understood and enacted a time-perception task as intended. Their claim is not that interviews revealed the true mechanism of time. The interviews informed face validity and experimental design. [14] This is the right level of ambition.

Phenomenology is strongest as construct engineering. It can separate dimensions that a task has collapsed, identify unexpected strategies, locate temporal transitions, and generate perturbations. It becomes weaker when it jumps from reported structure to neural cause, metaphysical ontology or population-wide necessity.

This placement changes the research sequence. Before a large study, a small descriptive pilot asks whether participants enact the nominal task through one strategy or several. During analysis, the same distinctions can reveal why identical button presses hide different routes. After an unexpected result, re-elicitation can generate alternatives, but it cannot retroactively purify a contaminated trial. The method is most valuable where a construct is still plastic enough to redesign. Once a vague label has become a benchmark, a dataset and a model target, its ambiguity becomes expensive infrastructure.

What reciprocal constraint really means

Varela’s neurophenomenology proposed that disciplined first-person accounts and cognitive science should constrain one another. Later work developed this as a research programme relating experiential categories to moment-to-moment brain dynamics. [10][11][13] “Mutual constraint” is often repeated too loosely. It does not mean that a neural signal validates a report merely by correlating with it. Nor does a report explain the signal. Whether subjective explanation must be wholly mechanistic remains contested; one recent argument explicitly rejects that assumption. [16] The methodological point here is narrower: neither evidence stream inherits the warrant of the other.

A real constraint removes live alternatives. If participants distinguish continuous swelling from discrete comparison, and those two report structures predict different responses to working-memory interference, the first-person distinction has changed the experiment. If a neural transition consistently precedes the reported onset under blinded coding, the third-person result pressures the temporal description. Each side can force revision without being translated into the other.

Independence here is a matter of dependency structure, not headcount. Two coders using the same theory-laden codebook are not two independent paths. Two interviewers trained from the same leading examples may reproduce one bias. A behavioural perturbation chosen after reading the transcripts can also overfit their categories. The evidence chain becomes stronger when at least one challenge is specified before coding, when prompt language and hypothesis language are separated, and when a negative case is allowed to remain negative. Public criticisability requires a route by which the favoured description can genuinely lose.

Executable lab: the phenomenological warrant gate

The following instrument implements the bracket, vary, contest protocol without a numerical score. It returns a claim class and the missing gates. A positive result permits a description to enter the next research stage as a provisional constraint. A negative result means the claim needs narrower wording, broader variation or a real rival. It does not mean the experience did not occur.

{}

How to interpret the output

The gate encodes four assumptions. A concrete episode is more testable than a trait summary. Bracketed claims remain visible instead of being silently deleted. Variation must include a failure condition. Public criticism requires a path that is not generated by the same prompt and codebook. A positive result permits only the next stage. A negative result identifies missing discipline. The instrument cannot establish consciousness, neural cause, clinical truth or universal necessity.

from __future__ import annotations

from dataclasses import dataclass, asdict
from enum import Enum
from typing import List


class ClaimClass(str, Enum):
    WITHHOLD = "WITHHOLD"
    DESCRIPTIVE_CANDIDATE = "DESCRIPTIVE_CANDIDATE"
    STRUCTURAL_HYPOTHESIS = "STRUCTURAL_HYPOTHESIS"
    CROSS_CASE_CONSTRAINT = "CROSS_CASE_CONSTRAINT"


@dataclass(frozen=True)
class Ledger:
    occurrence: str
    bracketed_claim: str
    rival_generator: str
    variations: List[str]
    break_condition: str
    delay_recorded: bool
    non_leading: bool
    own_words_first: bool
    independent_challenge: bool
    negative_case_allowed: bool


def evaluate(ledger: Ledger) -> dict:
    missing: list[str] = []
    concrete = len(ledger.occurrence.strip()) >= 40
    bracketed = len(ledger.bracketed_claim.strip()) >= 15
    varied = len([v for v in ledger.variations if v.strip()]) >= 3
    breakable = len(ledger.break_condition.strip()) >= 15
    rivalled = len(ledger.rival_generator.strip()) >= 20

    if not concrete:
        missing.append("one concrete, appearance-bound occurrence")
    if not bracketed:
        missing.append("an explicit world or causal claim placed in brackets")

    claim = ClaimClass.WITHHOLD
    if concrete and bracketed and ledger.delay_recorded:
        claim = ClaimClass.DESCRIPTIVE_CANDIDATE

    if claim == ClaimClass.DESCRIPTIVE_CANDIDATE:
        if varied and breakable and ledger.non_leading and ledger.own_words_first:
            claim = ClaimClass.STRUCTURAL_HYPOTHESIS
        else:
            if not varied:
                missing.append("at least three declared variations")
            if not breakable:
                missing.append("a result that would break the proposed structure")
            if not ledger.non_leading:
                missing.append("a non-leading elicitation check")
            if not ledger.own_words_first:
                missing.append("participant vocabulary recorded before the codebook")

    if claim == ClaimClass.STRUCTURAL_HYPOTHESIS:
        if rivalled and ledger.independent_challenge and ledger.negative_case_allowed:
            claim = ClaimClass.CROSS_CASE_CONSTRAINT
        else:
            if not rivalled:
                missing.append("a plausible rival generator")
            if not ledger.independent_challenge:
                missing.append("an independent challenge path")
            if not ledger.negative_case_allowed:
                missing.append("permission to retain negative or divergent cases")

    return {
        "claim_class": claim.value,
        "missing_gates": missing,
        "ledger": asdict(ledger),
        "permits": {
            ClaimClass.WITHHOLD: "re-elicitation only",
            ClaimClass.DESCRIPTIVE_CANDIDATE: "variation design",
            ClaimClass.STRUCTURAL_HYPOTHESIS: "independent perturbation",
            ClaimClass.CROSS_CASE_CONSTRAINT: "construct refinement or preregistration",
        }[claim],
        "cannot_establish": [
            "neural mechanism",
            "metaphysical truth",
            "universal essence",
            "clinical diagnosis",
        ],
    }

The strongest rival explanation

The strongest rival is that disciplined interviewing does not improve access to experience. It merely teaches participants a shared descriptive game. Concrete language, temporal anchoring and variation could produce coherent narratives without bringing the account closer to what occurred.

That rival can produce the same visible evidence: detailed transcripts, coder agreement and repeatable categories. It weakens only when the categories predict something they did not generate. A phenomenological distinction becomes more credible if it anticipates a perturbation response, survives blinded re-elicitation, appears in participant language before instruction, and fails where its stated break condition predicts failure.

Automation sharpens the boundary. A July 2026 preprint describes an LLM-based interviewer grounded in Descriptive Experience Sampling and states that validation is still pending. [15] Such a system could make prompts inspectable and repeatable. It could also scale one hidden bias across thousands of interviews. A stochastic elicitor must therefore log every question, prohibit supplied vocabulary, preserve uncertainty, and route unusual or sensitive material to trained human review. Scale is not intersubjectivity. It is repeated dependence unless independence is designed.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The beep that arrived late
  M->>E: Three kinds of leakage
  E-->>M: compare rival explanations
  E->>D: Make the route from experience to claim inspectable
  D-->>Q: revise scope or act

Make the route from experience to claim inspectable

When a study concerns perception, attention, emotion, agency or contemplative experience, do not ask whether first-person reports are trustworthy in general. That question is too coarse to guide a design. Ask which transformations stand between occurrence and claim, which commitments have been bracketed, which variations have been attempted, and which independent path could expose a manufactured description.

The changed decision is to use phenomenology before measurement as a construct-control layer, then again after measurement as a source of discriminating alternatives. Do not use it as a private veto over external evidence. Do not discard it as mere anecdote. Build a typed evidence chain.

Operationally, this means funding a small elicitation and variation phase before fixing labels, measures or model targets. Preserve the exact prompts and delays. Separate appearance, world claim, mechanism claim and generality in the data schema. Predeclare at least one break condition and one independent challenge. Then decide whether the resulting distinction deserves a larger study. This adds friction early, where revision is cheap, and removes false precision later, where a confident but underspecified construct can misdirect an entire programme.

For Nadia’s time-perception study, the practical consequence is immediate. The first report, “I compared two gaps”, should not enter the model as a mechanism label. It becomes a bracketed causal story. The second report, “a swelling narrowed into two events before the keypress”, becomes a descriptive candidate. Variation tests whether the structure survives changed wording, memory load and response timing. An independent perturbation decides whether the distinction improves the experiment.

This is a humbler result than private certainty and a stronger result than silence. Phenomenological rigour does not make experience incorrigible. It makes the route from experience to claim inspectable enough to disagree with.