Levels, Boundaries and the Grain of Explanation

At 02:13, a deterioration alert on a hospital ward turns amber. The risk model has read the latest observations and crossed its configured threshold. A nurse taps “seen” while dealing with another patient. The sound stops. The alert remains on one screen but does not enter the handover board. At 03:07, the patient needs urgent attention. The next morning, the review meeting begins with a familiar question: Why did the model fail?

The case is synthetic, and every number is illustrative. Its structure is common. A visible component produces a visible output, so explanation rushes towards that component. Yet a replay shows that the model produced the signal expected by its specification. The consequential failure occurred across a threshold, an interface, an acknowledgement convention, a handover process, a staffing pattern and an institutional measure that rewarded rapid queue clearance. Which of these belongs inside the system being explained?

There is no useful answer in the abstract. For the question “Why did the probability value change?”, the model may be the right system. For “Why was no action taken?”, the model is almost certainly too small. For “Why did this pattern recur across wards?”, even the nurse-model loop may be too small. The explanatory boundary must be earned by the contrast we want to change.

The aperture metaphor guards against two opposite errors. Reductionism assumes that the smallest visible component is automatically the deepest explanation. Holism assumes that every context belongs inside every explanation. Both avoid the real work. A good explanation states what is inside, what remains outside, which variables cross the boundary and what intervention would reveal that the chosen grain is wrong.

Part I The event stays still while the boundary moves

Begin with a simple manipulation. Replace the risk model with a perfect oracle, but leave the acknowledgement rule, handover board, staffing and escalation authority unchanged. Suppose the same non-response occurs. The replacement has improved one component while preserving the wider causal organisation. That result does not prove the model was irrelevant. It shows that model quality was not sufficient for the decision outcome.

Now perform the complementary intervention. Restore the original model, but require an acknowledged alert to enter a shared handover state until a named clinician closes it. Suppose the non-response disappears across varied workloads. The component has not changed; the loop has. The useful explanatory grain has moved because the intervention that changes the target lives outside the component.

A grain has four coordinates

“Level” is dangerously overloaded. David Marr separated the computational task, the representation and algorithm, and the physical implementation of an information-processing system.[2] That is a distinction among kinds of explanation, not a simple ladder from small to large. A team can have an algorithmic organisation. A single chip can be discussed at a computational level. Moving from algorithm to hardware is not the same operation as moving from person to team.

Mechanistic levels are different again. They relate the behaviour of a whole mechanism to the organised activities of its parts. Organisational levels distinguish individuals, dyads, teams and institutions. Spatial and temporal scales distinguish nanometres from kilometres, or milliseconds from months. These axes can cross. The right analysis may concern an institution-level policy, implemented through a team-level routine, realised in person-tool loops, observed over weeks.

A boundary is not a physical outline

A screen may sit centimetres from a person yet remain outside the mechanism if its contents are ignored. A remote registry may sit across a network yet belong inside the task if every valid action depends on its current state. Physical enclosure, network address and organisational ownership are therefore clues, not answers. The aperture follows the dependence that the explanation must preserve.

Four observations make membership more plausible. First, removing the entity changes the target under otherwise matched conditions. Second, its state is available at the time the process needs it rather than reconstructed afterwards. Third, the wider process constrains the entity and is constrained by it, so the coupling is reciprocal rather than a one-way supply of information. Fourth, the relation remains dependable across the contexts covered by the claim. Extended-mind and distributed-cognition arguments become practically useful when translated into these tests rather than settled by whether an object feels internal or external.[6][7]

None of the four observations is sufficient alone. A weather feed can alter a decision without becoming part of the decision-maker. A notebook can store state yet be irrelevant to the current task. Reciprocal interaction can also be noisy rather than functional. Membership is a joint claim about difference-making, role and organised dependence. This is why the same artefact can be inside the aperture for one question and outside it for another.

This is the practical lesson of extended and distributed cognition without the metaphysical shortcut. Clark and Chalmers argued that reliably coupled external resources can sometimes count as parts of a cognitive process.[6] Hutchins analysed navigation teams as cognitive systems whose properties cannot be recovered by inspecting one navigator in isolation.[7] Hollan, Hutchins and Kirsh brought that boundary shift into human-computer interaction.[8] The disciplined conclusion is conditional: coupling can make an external resource constitutively relevant to a task, but proximity or usefulness alone does not.

Part II An aperture is a causal hypothesis

Herbert Simon’s idea of near decomposability offers a first reason why boundaries work at all. In many complex systems, interactions within a subsystem are stronger or faster than interactions between subsystems. Short-run behaviour can then be approximated by looking inside the subsystem, while longer-run behaviour depends on cross-boundary interactions.[1] A useful boundary is therefore not a sealed wall. It is an approximation justified by an asymmetry of dependence and timescale.

That asymmetry can disappear. A model service may be nearly decomposable from the ward workflow while it computes a score. It is not nearly decomposable when the target is timely clinical action. A team may be treated as a unit during a five-minute response, yet its behaviour over six months may depend on training, staffing, procurement and performance measures. The same physical entities admit different grains because the target and horizon have changed.

This criterion is deliberately decision-relative. It does not claim that reality contains only the selected level. Nor does it say that wider causes are unreal. It asks a narrower question: which model is sufficient for this decision without hiding a dependence that would reverse the intervention’s expected effect?

The intervention test

Interventionist accounts of causation treat causal knowledge as information about what would change if we altered one factor in an appropriately controlled way.[3] For grain selection, the intervention matters twice. First, it identifies which variables are causally relevant. Second, it tests whether a coarse description preserves what a finer description predicts when the system is manipulated.

Let M be a fine-grained model, α a map that compresses its states into a coarser model, i a fine-grained intervention and ω(i) the corresponding coarse intervention. A useful abstraction should make the following two routes approximately agree:

Read the left route as “intervene on the detailed system, then summarise what happened”. Read the right route as “summarise the system first, then intervene at the coarse level”. When the results agree for the interventions that matter, the coarse grain is causally faithful for that use. When they diverge, the abstraction has hidden a distinction the decision needs. Multi-level causal models and recent causal-abstraction work formalise versions of this requirement.[12][13]

Worked example: the smallest numerical case

Use three binary variables in the alert loop. A is whether an alert exists, S whether someone has merely seen it, and H whether responsibility persists through handover. Define effective escalation as E = A × S × H. Every variable is either 0 or 1. The multiplication is not a clinical model. It makes one logical fact visible: a signal that is seen but not carried forward still produces zero effective escalation.

A component-only abstraction that records only A cannot distinguish the first three rows. It can support a claim about signal production, not a claim about escalation. A loop-level abstraction that records E is useful for the action question. Yet even E may fail over a longer horizon if repeated alerts change attention, staffing or threshold policy. An abstraction is adequate only for a declared intervention set and horizon.

Do not confuse constitution with causation

A component may constitute a mechanism without being an independent cause at another level. Craver and Bechtel argue that many apparent top-down effects are better described as same-level causal relations mediated through part-whole organisation.[4] Later work on constitutive relevance refines the experimental logic used to decide whether an entity or activity is genuinely part of the mechanism.[5]

This matters for language. “The team caused the nurse to act” may hide several distinct claims. The team’s communication pattern may be the mechanism through which action was coordinated. A supervisor’s instruction may be a same-level social cause. The team-level state may summarise constraints realised by individual actions. Each is testable in a different way. A higher-level description does not become causal merely because it is useful, and a lower-level description does not become explanatory merely because it is detailed.

Three assays for a disputed boundary

A boundary dispute often persists because each side presents another description. Description can reveal candidates, but intervention must do the deciding. Three small assays cover many cases.

The assays are complementary. Replacement tests whether a named part monopolises explanation. Relocation tests whether the proposed boundary has misplaced a causal role. Insulation tests whether the candidate is closed enough for the intended horizon. A grain earns selection only against the intervention family relevant to the decision. Passing one assay does not grant universal independence.

Part III Four failures caused by the wrong grain

Wrong grains survive because they are convenient. Component logs are easier to collect than interaction histories. Individual ownership is easier to assign than distributed control. Institutional language can protect local decisions from scrutiny, while component language can protect wider design choices. The resulting boundary may reflect the database schema, reporting line or preferred remedy rather than the mechanism.

A review should therefore ask a diagnostic question before debating causes: What becomes invisible at the unit in which the evidence was recorded? Missing sequence points towards a process grain. Missing reciprocity points towards a dyad or team. Missing adaptation points towards a longer horizon. Missing permission or incentive points towards an organisational or institutional grain.

1. The component tunnel

In machine learning, the component tunnel appears when model metrics are treated as system outcomes. Accuracy, calibration and latency may be genuine properties of a model under a defined interface. They do not establish that the configured system will obtain authorised context, present uncertainty well, route the proposal to the right person, execute a permitted action or verify the effect. A model proposal is not a system decision, and a system decision is not a verified outcome.

The same error occurs in biology and psychology. A neural correlate is not automatically the mechanism of a whole cognitive capacity. A person’s knowledge score is not automatically the team’s coordination capacity. A policy statement is not automatically the institution’s operating behaviour. Detail at the wrong place can create the feeling of depth while leaving the decisive coupling untouched.

2. The aggregate mirage

Suppose every member of a team can solve a task alone. It does not follow that the team can solve it together under time pressure. Conversely, no member may hold all the required information, while the interaction pattern reliably produces the answer. Interactive Team Cognition treats team cognition as activity that unfolds through communication and coordination, not simply as a static property averaged across members.[9]

An aggregate is justified when the within-unit variation it discards does not alter the target intervention. Mean individual accuracy may predict a quiz score. It may fail to predict a handover because sequence, timing, role complementarity and repair behaviour matter. The remedy is not to declare “emergence” and stop. It is to identify the interaction variable that carries the additional causal information.

3. The context swamp

When a narrow account fails, explanations often jump straight to “culture”, “the organisation” or “society”. Wider context can be causally important. Rasmussen modelled risk as a control problem spanning government, regulators, company, management and work practice.[14] Leveson likewise argued that accidents can arise from inadequate constraints and interactions rather than a single failed component.[15] These approaches widen the aperture, but they do not license vague totalisation.

An institutional explanation earns its place when it names an enforceable constraint, incentive, information path or authority relation and predicts what changes under intervention. “Culture caused it” is too broad if replacing a metric, changing a staffing rule or altering escalation authority would discriminate among rival accounts. Context becomes explanation only when it carries a testable difference.

4. Timescale aliasing

A camera aliases a fast wheel when its sampling rate makes forward rotation look slow or reversed. Explanations alias systems when observations are taken at the wrong temporal grain. A minute-level study may miss fatigue. A quarterly average may erase the order in which coordination broke. A pre-post evaluation may call a system stable while teams are still learning how to work around it.

Biological individuality offers a useful parallel. Krakauer et al. propose identifying candidate individuals partly through information propagated from their past into their future, allowing boundaries that need not coincide with obvious physical walls.[11] The relevant lesson is methodological, not that organisations are organisms. A boundary can be discovered through temporal dependence, and that dependence can change with the observation window.

Part IV A grain-selection protocol

The protocol below turns the aperture metaphor into a repeatable investigation. It is designed for research questions, system incidents, human-AI evaluations and policy decisions. It does not require a complete causal graph. It requires enough discipline to expose where the current model depends on what it has left outside.

Worked example: choosing the grain for the ward alert

At component level, changing the model threshold changes alert frequency, so causal relevance is present. Yet the outcome “owned action within fifteen minutes” remains unstable under workload and handover variation. The component fails closure. At person level, training may improve acknowledgement quality, but the result still depends on whether responsibility persists across shift change. The person grain also fails closure.

The person-tool loop captures signal, acknowledgement and immediate escalation, but it does not include the handover relation that carries responsibility across people. A dyad may fit one patient-clinician interaction, yet the ward task is distributed among specialised roles. Team level captures the shared state, role allocation, mutual monitoring and temporal handover. In the synthetic replay, it is the first grain that preserves the decisive intervention across the target horizon.

Institutional factors still matter. Staffing rules and queue metrics can alter team behaviour. They remain explicit boundary debt unless the decision concerns recurrent performance across wards or a policy change. Then the target and horizon widen, and institution may become the minimum sufficient grain. The protocol does not crown a permanent level; it makes the move between levels auditable.

Worked example: when the dyad is the first sufficient grain

Consider a synthetic document-review task. A language model highlights clauses, proposes a risk label and supplies supporting passages. An analyst accepts, edits or rejects the proposal. The target is not model accuracy in isolation. It is a correctly justified final decision within twenty minutes, with enough evidence for a second reviewer to reconstruct the route.

At component grain, a better model raises proposal accuracy. Yet final errors remain clustered in cases where fluent explanations induce rapid acceptance. Replacing the analyst with another person changes the pattern. Hiding the model label until the analyst records an independent judgement changes it again. Requiring passage-level evidence reduces one error class but increases time. These interventions reveal a coupled mechanism involving proposal quality, interface order, human confidence, verification cost and the analyst's veto.

The dyad is the first sufficient grain when those interaction variables preserve the intervention effect across the declared workload and session length. It does not follow that the dyad is a unified subject or agent. Nor does it follow that the institution is irrelevant. If a queue target rewards speed over review, that incentive remains boundary debt. It enters the selected grain only when the question changes from one review episode to recurrent performance under the operating policy.

Worked example: when institution becomes mechanism

Now compare two synthetic service centres using the same model, interface, training and team routine. One measures cases closed per hour. The other measures verified resolution and re-opened cases. Over several weeks, the first centre develops a shortcut: uncertain cases are closed with generic notes and return later. No component replacement explains the divergence because the local machinery is matched.

Change the metric and authority rule in a staggered rollout. If shortcutting falls while model, staff and interface remain stable, the institutional constraint carries discriminating causal information. For the question “Why does the pattern recur across teams?”, institution is no longer background. It is part of the mechanism through a metric, reporting cadence and allocation of authority. For the question “Why was this single case mislabelled?”, the same institution may be unnecessarily broad. The contrast decides whether policy is mechanism or context.

This example also limits the protocol. Institution-level selection requires evidence of a realised path from rule to local behaviour. A policy document alone is not that path. Investigators must show how the rule changes information, incentives, resources or permissions, and how those changes survive plausible local variation.

const CORE = [
  "causalSufficiency",
  "contextStability",
  "horizonFit",
  "interventionMapping",
  "authorityFit"
];

/** @typedef {0|1|2|3} Rating */
/**
 * @typedef {Object} GrainCandidate
 * @property {string} name
 * @property {number} order
 * @property {Rating} causalSufficiency
 * @property {Rating} contextStability
 * @property {Rating} horizonFit
 * @property {Rating} interventionMapping
 * @property {Rating} authorityFit
 * @property {Rating} broaderResidual
 */

function selectMinimumSufficientGrain(candidates) {
  const ordered = [...candidates].sort((a, b) => a.order - b.order);
  const passing = ordered.filter(candidate =>
    CORE.every(key => candidate[key] >= 2) &&
    candidate.broaderResidual <= 1
  );

  if (passing.length > 0) {
    return {
      status: "pass",
      candidate: passing[0],
      scope: "Valid only for the declared outcome, intervention set and horizon."
    };
  }

  const weakest = ordered
    .flatMap(candidate => CORE.map(key => ({
      candidate: candidate.name,
      criterion: key,
      rating: candidate[key]
    })))
    .sort((a, b) => a.rating - b.rating)[0];

  return {
    status: "unresolved",
    nextTest: weakest
      ? `Test ${weakest.criterion} at ${weakest.candidate} grain.`
      : "Add at least one candidate assessment."
  };
}

A small experiment that could weaken the thesis

Create a synthetic alert task with two model qualities, two handover designs and two workload regimes. Randomise teams across conditions. Pre-register the target as owned action within a fixed horizon. If model quality alone explains the intervention effect across both handover and workload conditions, the team-grain claim for this task weakens. If the handover intervention changes outcomes while model quality is held fixed, the loop or team grain gains support. If neither generalises, the candidate boundary or outcome definition is wrong.

The experiment does not prove that a team is a mind. It tests a narrower proposition: whether team-level organisation carries causal information needed to explain and control the target. This separation matters throughout human-machine research. A system may require team-level analysis for performance without possessing team consciousness, unified agency or moral patienthood.

Conclusion The decision this changes

When an explanation begins, do not ask which level is deepest. Ask which contrast must change, over what horizon, through which feasible intervention. Then move the aperture until the model preserves that intervention without importing decisive dependence through an unrecorded boundary.

The practical decision is to fund and govern the smallest system in which the intended intervention is actually closed. Everything narrower is a component claim. Everything wider is context until it adds a mechanism that changes the decision. The boundary remains provisional, because new coupling, adaptation or evidence can reopen it.

This is not explanatory relativism. The world constrains which apertures survive intervention. A grain that fails under replacement, state transfer, temporal extension or context shift is not equally good. Yet no single scale is automatically privileged for every question. Even claims of macro-level causal advantage are measure- and intervention-dependent, as work on causal emergence and coarse-graining makes clear.[10][16] Better explanation comes from a declared aperture, a discriminating test and an honest record of what still crosses the line.