Two variables move together. That can be useful for prediction, triage or anomaly detection. It still leaves open why they move together, which direction influence runs, and whether changing either one would alter the other.

An observation belongs to a regime

Return to the account-review incident. The confidence field and hold state are measured after applications have passed through data selection, missing-field handling, model scoring and policy evaluation. The observed distribution is therefore not raw reality. It is reality as filtered by a configured process. A different sampling window, threshold, missing-data rule or routing policy could strengthen, weaken or reverse the association.

This is true whether the dependency is summarised by a correlation coefficient, a regression weight, a conditional probability, a benchmark score or mutual information. An observed relationship belongs to the regime that generated and measured it. It cannot be detached from that regime and promoted into a universal mechanism.

The smallest formal distinction

Let X be low confidence and Y be a hold. The observational difference is:

The causal difference asks another question:

P(Y | X) and P(Y | do(X)) answer different questions. They coincide only under conditions that must be defended, not assumed. Causal inference is the discipline of making those conditions explicit and testing what follows from them.[3]

Worked example: a reversal hidden by aggregation

Suppose a new review method is tested on low-complexity and high-complexity applications. Within low-complexity cases, the new method succeeds 90 per cent of the time while the old method succeeds 80 per cent. Within high-complexity cases, the new method succeeds 30 per cent while the old method succeeds 20 per cent. The new method is better in both groups.

Now change only assignment. Give the new method 90 high-complexity cases and 10 low-complexity cases. Give the old method 10 high-complexity and 90 low-complexity cases. The new method then records 36 successes out of 100, while the old method records 74. The aggregate comparison reverses the within-group result. No arithmetic trick occurred. Case complexity altered the composition of the observed groups.

Bradford Hill’s classic discussion of association and causation did not offer a mechanical checklist that converts correlation into cause. It offered considerations that matter in context, including temporality, consistency, dose response and alternative explanation.[5] Modern causal methods sharpen the counterfactual question, but the same discipline remains: specify the contrast, expose the assignment process and state which assumptions identify the effect.

flowchart TD
  A["The field that caused the hold"] --> B["The trace is not the lever"]
  B --> C["The lever is not the machine"]
  C --> D["Interventions that lie"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#D95545,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Suppose the intervention is sound and changing X changes Y. We now have evidence that X is a difference-maker for Y in the tested conditions. We still do not know whether X is part of the organised process that realises Y.

Causation changes an alternative history

Interventionist accounts make causal claims answerable by well-formed changes. Roughly, X is causally relevant to Y when an intervention that changes X, without reaching YY.[4] This definition forces precision. What is set? Which pathways are held fixed? At what time? Across which population and operating range?

In the hold system, pinning confidence before policy evaluation changes a later path. The two runs represent alternative histories of the same application. A causal intervention reaches forward from a changed antecedent to a changed consequent. It may reveal direct, mediated, enabling or inhibiting influence, depending on the design.

A cause is contrastive, ranged and policy-relevant

“Does confidence cause a hold?” is still too broad. Cause is always relative to a contrast. Are we comparing 0.40 with 0.80, missing with present, or model-derived with human-entered confidence? The effect can differ across those contrasts. It is also ranged. Moving a score from 0.40 to 0.50 may cross a threshold and change the path, while moving it from 0.70 to 0.80 changes nothing. A global sentence can therefore conceal a local mechanism.

The intervention also has a policy meaning. Setting confidence to a value the evidence cannot support may estimate a mathematical effect while describing an operationally impossible action. A valid causal question must distinguish manipulability for inquiry from permissibility for deployment. Researchers may clamp an internal variable to learn how a model works. An operating system may be forbidden to fabricate that value. The causal result survives; the available decision does not.

Finally, effects can depend on context. A confidence change may matter only when the policy version contains a threshold, the account is eligible for automated review, and the authority grant is fresh. This is not a defect in the claim. It is its domain. A useful causal statement names the intervention, contrast, operating range and effect modifier. Without those fields, “X causes Y” is too compressed to guide engineering or governance.

Constitution changes the question

Mechanistic explanation asks how a phenomenon is produced by entities and activities organised so that they carry the process from start to finish.[6] A constituent is more than something that happened before the target. It participates in the target’s realisation at the chosen boundary and time grain.

Consider a heartbeat. A fright can cause the heart rate to rise, but fright is not a component of cardiac contraction. A hammer can cause a clock to stop, but the hammer is not part of timekeeping. Training data can cause a model to acquire a parameter configuration, but the historical data stream is not automatically a component of each later inference. A power outage can stop an AI service, but the outage does not thereby belong inside the reasoning mechanism.

Causal relevance is necessary for many constitutive explanations, but causal relevance alone is too permissive. Historical causes, external disruptors and background enablers can all change the phenomenon. Constitution adds parthood or inclusion, concurrent participation, organisation and a declared system boundary.

The account hold as an organised event

The hold example becomes clearer when the target is typed precisely. The target is not “the model says hold”. It is “an authorised hold exists for account A, for purpose P, under policy version V, and has been independently read back”. That target spans several activities.

The scorer supplies evidence. The policy gate combines that evidence with rules and current state. The authority kernel checks identity and permission. The typed action contract limits what may be written. The system of record realises the effect. Readback distinguishes requested action from verified outcome. Remove the proposal and the gate may have no candidate action. Remove authority and the proposal cannot become an effect. Remove the write and there is no hold. Remove readback and the system cannot establish whether the effect occurred.

None of those observations says that every named component is indispensable in every implementation. A different architecture may realise the same governed relation. The constitutive claim is therefore not “these vendor services are the essence of holding”. It is that some organised set must preserve the relevant roles: evidence, policy interpretation, authority, effect and verification.

One factor can occupy several relations

The categories are not mutually exclusive labels for variables. A policy gate can correlate with holds across logs, causally alter them when its rule is changed, and participate constitutively in the governed decision. The distinction lies in the relation tested, not in the object’s permanent type. The same component may also be an effect at another timescale. A deployment decision caused the gate to exist; during execution, the gate helps realise the hold.

This matters when explanations cross levels. A component state can be caused by neighbouring components while helping constitute a system-level activity. Craver and Bechtel describe many apparent cases of top-down causation as hybrids: constitutive relations connect levels, while ordinary causal relations operate within a level.[8] Whether that account covers every case is contested, but its discipline is valuable. Before drawing an arrow between “system” and “component”, specify whether the arrow denotes temporal influence, part-whole inclusion or a shorthand combining both.

A good diagram therefore uses different graphical grammar for different relations. Solid directed arrows can mark executed causal paths. Containment or nested boundaries can mark constitution. Dotted lines can mark hypotheses or uncertain mappings. Treating all three as identical arrows creates a visual version of the conceptual error this paper is trying to prevent.

Necessity and sufficiency do not settle constitution

Researchers often reach for two familiar tests. If removing a factor abolishes the phenomenon, call it necessary. If adding the factor produces the phenomenon, call it sufficient. Both tests are useful, but neither maps cleanly onto constitution.

A factor can be necessary without being a constituent. Oxygen may be necessary for a person to complete a reasoning task, but ambient oxygen is not normally represented as a step in the argument. Electricity may be necessary for a server process, yet a study of token selection can reasonably treat mains generation as background. Necessity depends on what alternatives the environment supplies and where the boundary is drawn.

A constituent can also fail to be necessary. Redundant circuits, replicated services and compensatory biological pathways permit the target to continue after one part is removed. Nor need a constituent be sufficient. A policy gate alone cannot create a governed hold because it lacks evidence, authority, effect and readback. Sufficiency usually belongs to an organised set under background conditions, not to an isolated part.

The more useful question is contribution under organised alternatives. Does changing the candidate alter the target in the predicted way? Can another realiser substitute while preserving the same role? Does the effect disappear when the proposed organisation is broken but the parts remain? Does the explanation predict recovery, failure and untested branches? These questions replace a binary necessity test with a causal contribution profile.

That profile is still relative to grain. A database write may be necessary for a millisecond-scale operational hold, while institutional authority is necessary at the process scale and legal mandate at the organisational scale. The factors do not compete for one true level. They answer different explananda. The discipline is to keep the target fixed long enough to know which one is being explained.

Intervention is powerful because it replaces passive observation with a controlled difference. It can also create false confidence when the change is destructive, reaches several pathways, crosses the wrong boundary or tests only one redundant realiser.

Output-preserving replacement

Four ways an intervention overclaims

A destructive intervention proves too much. Removing a clock’s hands stops readable time, but so does crushing the clock. The first change targets a likely constituent; the second demonstrates only vulnerability. An ablation that destabilises every downstream activation may show that a component is structurally load-bearing without revealing its ordinary computational role.

Redundancy proves too little. Two pumps can jointly maintain flow. Remove either one and output remains normal. A single-component ablation would falsely suggest that neither matters. The relevant intervention may need to remove both, perturb coordination, or increase demand until reserve capacity becomes visible.

A fat-handed intervention changes several routes at once. Patching a neural activation can alter the intended feature, layer normalisation, downstream scale and distributional validity. A drug can affect several receptors. A policy toggle can alter routing and logging together. The observed effect cannot be assigned to one path without controls.

A boundary error turns environment into mechanism or mechanism into background. Network latency may be external to a model’s token computation but internal to a deployed decision loop. Human confirmation may be outside a narrow algorithm boundary but constitutive of an authorised business action. There is no boundary-free answer.

A null effect has several meanings

Suppose an ablation produces no measurable change. Four interpretations remain open. The candidate may truly be irrelevant. The system may compensate through redundancy. The intervention may have missed the causally active aspect of the candidate. Or the outcome measure may be too coarse to register the change. A fifth possibility appears in adaptive systems: the effect may emerge only later, after the system updates around the perturbation.

These alternatives imply different next tests. To probe redundancy, combine perturbations or increase demand. To probe intervention validity, measure the local state before and after the change rather than trusting the command. To probe measurement sensitivity, inspect trajectory, latency, calibration or error type instead of final accuracy alone. To probe adaptation, follow the system over the timescale on which compensation occurs.

The mirror problem applies to a large effect. Dramatic failure may reflect an ordinary constitutive role, a shared bottleneck, an invalid state, or collateral damage. Effect magnitude ranks consequences, not interpretations. This is why component importance scores cannot by themselves become mechanism diagrams. They need a model of pathways and controls that can separate the rival reasons for the same score.

A useful report therefore pairs every positive or null result with a lowering condition. State what finding would reduce confidence in the preferred interpretation. For example: “The constitutive hypothesis weakens if a norm-matched random direction produces the same behavioural effect, if another seed lacks the mapping, or if restoring the downstream state rescues behaviour without restoring the candidate.” Negative evidence becomes informative when the rival outcomes are planned before the result is seen.

Matched interlevel experiments

One influential approach to constitutive relevance is mutual manipulability. If changing a component changes the whole phenomenon, and changing the whole through a distinct route changes the component in the predicted way, the paired result supports an interlevel relation.[7] Later work refines this idea using matched interlevel experiments and causal betweenness.[12]

The attraction is clear. Constitution becomes experimentally tractable rather than a matter of visual resemblance. Yet critics argue that mutual manipulability can conflate causal and constitutive relations, presuppose the very part-whole relation it aims to establish, or remain vulnerable to fat-handed intervention.[9][10] A recent critique argues that matched interlevel experiments may reduce constitution to causal mediation and lose the distinct content of realisation claims.[13]

stateDiagram-v2
  [*] --> Claim
  state "An observation belongs to a regime" as Claim
  state "The perfect warning lamp" as Test
  state "The smallest formal distinction" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

The aim is not to force every investigation to the deepest layer. It is to prevent a surface result from being written as a causal result, or a causal result from being written as a constitutive explanation.

The three-cut protocol

The protocol below treats an explanation as a typed claim. Each cut removes a different ambiguity.

The trace cut records the observational regime: population, measurement, selection, timing and dependency. It asks whether the pattern survives reasonable remeasurement and stratification. Passing this cut licences an association or predictive claim.

The history cut specifies an intervention or identifying strategy: what is set, which incoming causes are severed, what is held fixed, and which later outcome is compared. Passing this cut licences a bounded causal claim.

The assembly cut tests whether the candidate factor participates in the target’s ongoing organised realisation. It records the boundary, time grain, component role, redundancy, reciprocal constraints, replacements and rival mechanisms. Passing this cut licences a constitutive hypothesis, not a metaphysical verdict.

The worksheet before the experiment

Before collecting more data, write one sentence for each field. Target: the exact phenomenon to explain. Contrast: the alternative state or outcome. Boundary: what counts as inside the candidate system. Time grain: the event window over which the target exists. Relation: trace, history or assembly. Rival generator: another process capable of producing the same evidence. Decision: what action will change if the claim survives.

These fields prevent a common failure in AI evaluation. A team begins with a behavioural score, notices an activation, adds an ablation, and gradually changes the target from “predicts the answer” to “is the reasoning process”. The evidence may improve at every step while the explanandum drifts. Freezing the target and contrast makes that drift visible. If the target must change, create a new record rather than silently rewriting the old one.

The rival generator is especially important. For a correlation, it may be confounding or selection. For a causal result, it may be a collateral path, invalid intervention or mediated effect. For a constitutive claim, it may be a redundant or alternative realiser. Building the rival is often more informative than collecting another confirming example because it creates a discriminating intervention.

The final field is the prohibited inference. This is not ceremonial caution. It protects downstream decisions from inheriting a stronger claim than the study tested. A trace record might prohibit causal language. A causal record might prohibit claims of uniqueness or constitution. An assembly record might prohibit transfer to another architecture, scale or metaphysical property. The strongest trustworthy result is the one whose boundary travels with it.

Worked audit: a feature that predicts refusal

An interpretability team discovers a direction in a language model’s activations that predicts refusals with 94 per cent accuracy. The trace cut is promising. The team checks held-out prompts, alternative probes and prompt families. The feature remains decodable.

They then patch the direction from a refusal-producing run into a matched non-refusal run. Refusal probability rises. Under controls for patch magnitude, layer and prompt, the result supports causal relevance. Causal abstraction work formalises how high-level variables may align with low-level interventions, offering a route from correlation to mechanism claims.[15]

Does the direction constitute refusal? Not yet. The patch may alter several features. A distributed circuit may realise the behaviour with redundancy. The direction may be a convenient basis that changes across model seeds. A rival mechanism could route through another subspace while preserving output. The assembly cut therefore asks whether the mapping survives reparameterisation, replacement, downstream restoration and matched interventions across the proposed circuit. Causal representation research similarly stresses that useful representations should support interventions and changes in data-generating process, rather than only decode observations.[14]

The licensed sentence is then precise: “This activation direction is a stable causal handle on refusal under the tested prompt and patch regime.” The prohibited sentence is: “This neuron or direction is the refusal mechanism.”

Worked audit: the original hold incident

The worksheet now returns to the opening incident. The target is an authorised hold recorded and verified for one application. The contrast is the same application proceeding without that hold. The boundary includes scoring context, policy gate, authority check, typed write, system of record and readback. The time grain begins when current evidence is assembled and ends when the outcome is verified or declared unknown.

The trace record states that low confidence predicts holds in the incident sample. It also records the selection process: only applications reaching automated review are present, the field is measured before the policy decision, and the displayed threshold is policy-version dependent. The history record states that pinning confidence in staging changed the hold rate during exact input replay. Its isolation checks confirm that policy version, identity, account state and tool responses were held constant.

The assembly record then asks a harder question. When the policy gate is replaced by a pass-through, does confidence retain its effect? When the authority service denies permission, can any model proposal create the hold? When the write times out, can independent readback distinguish failure from an unknown outcome? When the scorer is replaced by a rule-based estimator that preserves the typed evidence relation, does the governed path continue? These interventions test roles and organisation rather than product names.

The resulting conclusion is layered. Low confidence was an incident predictor. The confidence field was a causal input to the tested policy path. A governed hold was realised by a wider evidence-to-authority-to-effect organisation. Training history, operator interruption and network power remained causal background under this boundary. This record is more verbose than “confidence caused the hold”, but it is also reusable. It tells engineers which component to repair, assurance teams which evidence to demand, and incident responders which outcomes remain unknown.

How to use the lab

Begin with high confounding and low direct effect. A large observed difference paired with a small interventional difference demonstrates why a predictive trace cannot be read as a lever. Increase direct effect. When the interventional difference grows, a causal claim becomes available. Then check whether the activity sits inside the target boundary, participates at the same time and survives a matched probe. Marking a collateral pathway blocks the constitutive grade even when the causal effect is large.

A positive causal result permits the sentence “setting X changed Y under this synthetic intervention”. A negative result means either that X has little direct effect in the tested range or that the experiment lacked sensitivity. It does not prove absence under every regime. A constitutive-candidate result permits a stronger hypothesis about organised realisation. It still cannot establish uniqueness, subjective experience or metaphysical identity.

/** @typedef {'trace'|'history'|'assembly'} Depth */
/**
 * @typedef {Object} DepthAudit
 * @property {string} target
 * @property {string} contrast
 * @property {string} boundary
 * @property {string} timeGrain
 * @property {number} observedDifference
 * @property {number} interventionalDifference
 * @property {boolean} validIsolation
 * @property {boolean} insideBoundary
 * @property {boolean} concurrentParticipation
 * @property {boolean} matchedProbe
 * @property {string[]} rivalGenerators
 */

/** @param {DepthAudit} a */
function classify(a) {
  const association = Math.abs(a.observedDifference) >= 0.05;
  const causal = Math.abs(a.interventionalDifference) >= 0.05 && a.validIsolation;
  const constitutiveCandidate = causal && a.insideBoundary &&
    a.concurrentParticipation && a.matchedProbe;

  if (constitutiveCandidate) return {
    depth: 'assembly',
    licensed: 'Constitutive hypothesis within the declared boundary',
    prohibited: 'Uniqueness, metaphysical identity or untested transfer'
  };
  if (causal) return {
    depth: 'history',
    licensed: 'Bounded causal effect under the tested intervention',
    prohibited: 'Constitution or sole-mechanism claim'
  };
  if (association) return {
    depth: 'trace',
    licensed: 'Association or predictive use in the observed regime',
    prohibited: 'Causal or constitutive claim'
  };
  return {
    depth: 'underdetermined',
    licensed: 'No positive dependency claim at current sensitivity',
    prohibited: 'Proof of absence in other regimes'
  };
}