The camera that invents beads

Imagine a laboratory that can interrupt you at arbitrary moments and ask, “What was in consciousness just before the signal?” You answer with a noun phrase: the hum of the air conditioner, tomorrow’s meeting, a line from a song, pressure in the jaw. After a hundred probes, the researcher has a distribution of contents. The method appears to have sampled the stream.

The hidden assumption is that whatever can be named at the probe is the whole state that mattered at that moment. Once that assumption enters the design, transitions become empty intervals. A thought is treated as a bounded item; the next item replaces it; continuity is reconstructed afterwards from similarity between labels.

James objected to exactly this picture. In The Principles of Psychology, he distinguished the discrete things thought may be about from the subjective flow in which they are apprehended. The objects can change abruptly while the thought that takes them up still bears ownership, context and felt relation to what came before. He described slower passages as resting places and faster passages as “flights and perchings”, but both belonged to one stream.[1]

A sample is not a slice of the stream

Sampling contains two transformations that are easy to confuse. First, a living episode is interrupted or retrospectively marked. Second, the participant and researcher translate that episode into a record. The first transformation may disturb the process. The second can erase features even when the interruption is harmless. A person may accurately report “thinking about lunch” while the record still loses whether the thought was intruding, deliberately summoned, fading, or carrying a sense that something had been forgotten.

This matters because a database row is not a miniature consciousness. It is a projection chosen for a purpose. A colour camera can map wavelength differences while missing temperature; a microphone can preserve pressure changes while missing spatial attention. Neither failure is mysterious. The instrument was not designed to retain the omitted relation. Mental sampling deserves the same discipline.

Three claims must therefore remain separate. Phenomenological continuity says that experience is lived as an ongoing course. Process continuity says that causally relevant state crosses from one episode into the next. Metaphysical continuity says what kind of thing the stream ultimately is. A probe report can inform the first, and interventions can test the second. Neither, by itself, settles the third. The present argument needs only the middle claim: if omitted transitional variables improve prediction after focal content is held constant, then the bead model is causally incomplete.

The same mistake appears in ordinary explanation. We often infer that the mind contains only what can be pointed to, because nouns are easier to report than relations. James noted the asymmetry: language has names for colours and objects, yet the felt “and”, “but”, “towards” and “not quite” of thinking are harder to hold still.[1] A method built from checkboxes inherits that bias.

The first correction is therefore methodological: do not equate a sample of contents with a sample of the process. Before asking whether consciousness is continuous, discrete or segmented, specify what the instrument can represent.

flowchart LR
  A["The camera that invents beads"] --> B["What carries a thought forward"]
  B --> C["Segments without mental atoms"]
  C --> D["Sampling without freezing"]
  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

What carries a thought forward

James’s stream becomes clearer when treated as a mechanism rather than a metaphor. Four features do the work: variation in pace, transitive feeling, fringe and selection. None requires the mind to be smooth at every scale. Together they explain how a changing process can remain one intelligible course.

Flights and perchings are changes of pace

Some moments are easy to stabilise: a red cup, an inner sentence, a remembered room. Other moments are mostly passage: recognising that one idea contradicts another, sensing where a sentence is going, abandoning a line of reasoning before a new one has arrived. James called the first substantive and the second transitive. The terms mark different rates and functions within thought, not two kinds of mental substance.

This distinction matters because a transition cannot always be inspected without being changed. Trying to hold a passing tendency still may turn it into a new focal object. The measurement then substitutes “thinking about the transition” for the transition itself.

The fringe is a constraint, not a second object

Return to the forgotten name. The face is focal. The sense that a candidate is wrong is not another image beside the face. It is a relation between the present content, remembered context and an anticipated fit. The fringe can include familiarity, expectation, relevance, ownership, emotional colouring and the sense of a thought’s destination.

A useful way to model this is to distinguish the content currently reportable from the constraints currently active. Let X_t be focal content, F_t the fringe of felt relations, C_t deliberate or automatic constraints, and H_t the carried history of the episode. A bead model predicts the next content from the current content alone:

What the equation commits us to

The bead model hides a strong assumption. Once the focal content X_t is known, earlier history and surrounding relations contribute nothing further to the distribution of X_t+1. In statistical language, the present content screens off the past. That may be a useful approximation for some tasks, but it should be tested rather than inherited from the reporting format.

The transition-bearing model relaxes that screening assumption. It predicts that two episodes with the same focal label can diverge because their fringe, active constraint or carried history differs. The cleanest test is not to ask people whether the fringe “caused” the next thought. It is to hold the focal content roughly constant, alter one contextual relation, and examine the successor distribution. A reminder, unresolved goal, emotional revaluation or task instruction can provide the intervention.

This model also has a failure condition. If movement and fringe reports add no repeatable prediction beyond focal content, and controlled changes to the proposed constraints leave successor probabilities unchanged, the richer account loses its practical advantage. It may still describe how thought feels, but it has not earned a causal role. Phenomenological richness and mechanistic usefulness must be demonstrated separately.

Selection creates a channel through abundance

At any moment, many sensations, memories and possible continuations are available. Interest and attention select among them. Selection does more than spotlight an already complete bead. It changes which relations are amplified, which tendencies mature and which paths close.

This helps explain why continuity is compatible with novelty. A conversation can turn sharply when a remark changes its emotional meaning, yet the turn still belongs to that conversation because the new direction inherits unresolved relevance from the previous moment. Continuity is not sameness of content. It is constrained succession with carried stakes.

Continuity does not mean smoothness

A stream can turn, narrow, bifurcate and even be interrupted. James’s claim does not require a perfectly uniform medium or a subject that never loses track. It requires attention to the relations by which one phase is taken up as following, correcting, completing or displacing another. A sudden alarm may replace an inner melody in a fraction of a second, yet the shock is experienced as an interruption of what was happening, not as the birth of an unrelated universe.

That distinction removes a common false choice. Thought need not be either an unbroken blur or a sequence of sealed atoms. It can be locally segmented and globally path-dependent. Stable phases make inspection possible; transitive phases make direction possible. Selection continually reshapes the reachable future without requiring a separate observer outside the stream.

The same distinction should govern comparisons with artificial systems. A token stream, chat transcript or persistent persona may look continuous. To claim a Jamesian analogue, one must identify a state that carries unresolved relations across time, intervene on it, and show that successor possibilities change. Fluency alone is not the bridge.

Segments without mental atoms

Recent research complicates a simple reading of James. Continuous experience is often segmented into events. During films and narratives, patterns of brain activity remain relatively stable within an event and shift near meaningful boundaries, with different cortical regions integrating over different timescales.[5] Studies of spontaneous thought also find clusters of related content punctuated by larger transitions.[6]

Does this restore the beads-on-a-string model? Only if an event boundary implies that nothing causal crosses it. The evidence does not establish that. It supports a weaker and more useful claim: thought has episodes, transitions and multiple timescales.

The decisive test is not boundary detection

Imagine two generators that produce ten segments, each lasting about twenty seconds. Generator R resets at every boundary and chooses the next topic from a fixed distribution. Generator C also changes topic at a boundary, but an unresolved concern from the previous segment biases the next choice. A duration histogram, a topic-transition count and even a semantic trajectory can look similar. The difference lies in conditional dependence across the boundary.

To expose it, intervene just before the transition. Strengthen or remove the unresolved concern while leaving the current topic and sensory setting as similar as possible. If post-boundary choices change systematically in Generator C but not Generator R, the boundary did not erase the relevant history. This is the experimental meaning of process continuity. It requires no claim that experience is mathematically continuous at every instant.

The same logic applies to event segmentation findings. Detecting a neural or reported boundary shows that organisation changed. It does not show whether control state, affective value, task set or anticipatory structure crossed that boundary. A boundary becomes evidence for a mental atom only when suitable interventions fail to find any carried dependency. Until then, segmentation and continuity are compatible descriptions at different grains.

What current studies add

Sripada and Taxali used verbalised thought and text analysis to report a “clump-and-jump” organisation: semantically related stretches followed by larger changes.[6] Later work found that some transitions include unfocused or scattered intervals rather than clean jumps.[7] A 2026 study asked participants to mark thought transitions and found that the distribution of coherent segment durations was better fitted by an accumulation-to-threshold model than several alternatives. In that study, the mean segment lasted about twenty seconds, and lower fitted thresholds were associated with greater self-reported disorganisation and inattention.[11]

That result is suggestive, not final. The accumulated quantity is unknown. A participant may be detecting a naturally occurring transition, accumulating evidence that a gradual drift has become “different enough”, or partly learning the experimenter’s boundary criterion. The study itself identifies this alternative.[11] Still, it improves the question. Instead of asking whether thought is continuous or discrete in the abstract, we can ask what variable accumulates, what sets the threshold, and whether the threshold changes under intervention.

Neural evidence also points to organised transitions. In a 2025 think-aloud fMRI study, larger semantic shifts in spontaneous memory and future-oriented thought were associated with responses in default and control networks, resembling some responses seen at external event boundaries.[9] Yet the researchers also caution that verbalising, being recorded and lying in a scanner may alter the thought trajectory. The method opens the stream by adding another current.

Experience-sampling work in 2026 found that people could distinguish freely moving thought from deliberately directed thought across several settings, and that the two dimensions were negatively related without being simple opposites.[10] This matters for James’s selection mechanism. A thought can be both exploratory and intentionally oriented, as in drafting an essay where ideas move freely inside a chosen problem.

The measurement frontier

Every method trades temporal access against reactivity. Sparse random probes preserve natural activity but rarely land inside fast transitions. Continuous think-aloud captures sequence but introduces speech planning, self-monitoring and social filtering. Self-caught keypresses can target boundaries with less verbal load, yet require sustained monitoring and may favour participants who notice transitions readily. Descriptive Experience Sampling reduces retrospective distance through random beeps and later expositional interviews, but it remains labour intensive and idiographic.[2]

Recent methodological work reinforces two Jamesian lessons. First, individual patterns can differ enough that group averages misdescribe a given person; repeated within-person sampling improves person-specific prediction.[12] Second, numerical ratings alone often lose context, cause and temporal order, so open descriptions remain necessary.[13]

stateDiagram-v2
  [*] --> Claim
  state "The freeze-frame inspector" as Claim
  state "A sample is not a slice of the<br/>stream" as Test
  state "Flights and perchings are<br/>changes of pace" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Sampling without freezing

The practical consequence is a different measurement unit. Rather than ask only “What were you thinking about?”, sample three layers separately: the focal nucleus, the movement of the episode, and the constraints shaping that movement. I call the resulting method Transition-Sensitive Experience Sampling, or TSES.

TSES is not a replacement for established experience-sampling, think-aloud or neurophenomenological methods. It is a small design extension aimed at one failure mode: capture-induced atomisation. Its central rule is simple: do not require every valid sample to contain a nameable object.

The protocol combines two capture modes. Random probes estimate how often different modes occur without waiting for participants to notice them. Brief self-caught markers target transitions that random probes would miss. Neither is treated as ground truth. Agreement and disagreement between them become data.

How the protocol works

A minimal study uses three sessions. In the first, six to eight random probes establish a baseline. In the second, participants add a single keypress when they notice a transition, while random probes continue at a lower rate. In the third, the same design is repeated during a real task such as reading, planning or walking. Each random probe asks about the final undisturbed moment before the signal, not the explanation produced afterwards.

The response should take less than twenty seconds. The nucleus field is optional. The movement field is categorical. Constraint ratings are short and separable: deliberate direction, automatic capture, external coupling, felt rightness or wrongness, unresolved tension, and confidence in the report. One open sentence preserves context and temporal order. Later interviewing can deepen a subset of samples, but the initial record remains unchanged.

A sample is an event packet, not a verdict

Each TSES record should preserve four times: the signal time, the estimated moment being described, the response start and the response completion. These timestamps separate the target episode from the reflection needed to report it. The raw wording is immutable. Later coding may append a movement category or inferred context, but it must not overwrite the participant’s description. This protects the distinction between experience, report and analyst interpretation.

The protocol also requires a declared prediction target. One study may ask whether movement mode predicts the next reported topic. Another may predict a behavioural switch, response latency or return to an unfinished task. Without a target, “the fringe mattered” becomes too elastic to fail. The smallest useful analysis compares three models on held-out samples from the same person: focal content alone; focal content plus movement; and focal content plus movement and constraints. Improvement must survive a second session and a changed probe schedule.

Group statistics can follow, but the first unit is the person-context pair. A category that is reliable during reading may be meaningless during walking. A participant who uses “shifting” for topic change may differ from one who uses it for a change in stance towards the same topic. Brief calibration interviews can identify these usages, while repeated sampling estimates whether they remain stable.[12] Open text is retained because fixed ratings cannot reliably recover why a movement occurred or which event came first.[13]

The protocol succeeds when it makes its own distortion measurable. Vary probe density across sessions, include a no-self-caught baseline, and ask after each session whether monitoring changed the course of thought. Reactivity should enter the analysis as a competing cause of observed transitions, not as a footnote.

type Movement = "stable" | "arriving" | "leaving" | "shifting" | "scattered" | "blank";

type TSESSample = {
  beepTime: string;                  // immutable receipt
  captureMode: "random" | "self-caught";
  nucleus?: string;                  // absence is valid data
  movement: Movement;
  from?: string;
  toward?: string;
  deliberateConstraint: 0 | 1 | 2 | 3 | 4;
  automaticCapture: 0 | 1 | 2 | 3 | 4;
  externalCoupling: 0 | 1 | 2 | 3 | 4;
  fringe?: ("rightness" | "wrongness" | "familiarity" | "tension" | "direction")[];
  confidence: 0 | 1 | 2 | 3 | 4;
  openContext?: string;
};

// Invariant: coding may append fields, but never replace the participant's raw sample.

What a result permits

A positive result is not “the stream is continuous”. A useful positive result is narrower: movement and fringe reports show repeatable within-person structure; they predict successor content or behaviour beyond focal content; and the pattern survives changes in probe density, wording and task. That would justify treating transition variables as part of the explanatory model.

A negative result also has several meanings. Participants may lack reliable access to the proposed dimension. The prompt may be too slow. The categories may impose alien distinctions. The phenomenon may be genuinely weak in that context. A negative result should therefore lower confidence in the measurement bridge before it lowers confidence in all transitional structure.

The protocol cannot establish that a fringe feeling is a distinct neural mechanism, that phenomenological continuity is metaphysically fundamental, or that an artificial system with similar dynamics has experience. It can establish whether omitted transition variables improve explanation and prediction.

Conclusion

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The camera that invents beads
  M->>E: The freeze-frame inspector
  E-->>M: compare rival explanations
  E->>D: Build an instrument capable of seeing passage
  D-->>Q: revise scope or act

Build an instrument capable of seeing passage

When a mind is measured as a list of contents, the list begins to look like the mind. James’s lasting contribution is to expose the circularity. Stable objects are easiest to name, so methods record them; what is recorded is then used to argue that consciousness consists of stable objects connected by nothing of consequence.

The decision to change is the unit of analysis. For studies of human thought, do not treat the probe answer as a complete state. Record whether the episode was stable or moving, what relation or tendency surrounded the focus, and what constraint shaped the next step. For computational models, distinguish a sequence that merely preserves output order from a process whose unresolved state causally changes successor possibilities. For human-machine work, measure the transition cost and carried concern that accompany context switches, not only the task before and after.

That change also alters how findings are communicated. Report “a boundary was detected” when a detector crossed its criterion. Report “carried state influenced the next episode” only after a cross-boundary intervention or predictive test. Report “continuous experience” as a phenomenological claim tied to a method and population, not as a synonym for every form of temporal dependence. These verbs keep observation, mechanism and ontology from sliding into one another.

This does not require denying event boundaries. It requires refusing a false inference from boundary to independence. A river can have eddies, rapids, channels and confluences. What makes it one course is not the absence of structure. It is that each structure changes the conditions of what follows.

The practical rule is therefore compact: before calling a mind discrete, build an instrument capable of seeing passage.