From Observation to Explanation: Correlation, Causation and Constitution

The trace is not the lever

Begin with the least controversial observation: two measurements differ together. That fact may be valuable. It may support forecasting, triage or anomaly detection. It does not yet tell us which difference produced the other.

An observation belongs to a regime

Suppose application holds occur more often when evidence confidence is low. The statement is incomplete until we name how the cases entered the data. Was confidence measured before the policy gate ran? Did the service request extra evidence only for risky cases? Were manually reviewed cases excluded? Did a threshold compress a continuous score into two bins? The observed relationship belongs to this measurement and selection regime, not to an unlabelled world.

Correlation here means any statistical dependence, not only Pearson’s coefficient. A conditional probability, mutual information estimate, regression coefficient, embedding similarity or benchmark score can all expose dependence. The form changes; the epistemic limit does not. A pattern can be excellent for prediction while remaining radically ambiguous about production.

The smallest useful causal question

A causal question compares outcomes under alternative assignments. In potential-outcome notation, each case has a possible outcome under treatment and under no treatment, although only one is observed. In a structural causal model, the do operator represents replacing the normal assignment rule for a variable with an external setting. These are different formalisms for the same practical shift: stop asking what happened among naturally occurring groups and ask what would differ under a specified change.[2][1]

Interventionist explanation adds a useful discipline: the change must be connected to a counterfactual pattern, not merely followed by an outcome. A proposed cause earns explanatory relevance when suitably changing it would change the outcome under specified background conditions.[4] That still leaves open whether the changed item is a trigger, an external condition or a constituent of the process.

The notation compresses a vital design choice. What exactly is set? For whom? At what time? Compared with what alternative? Over what outcome window? “Does confidence cause holds?” is still too vague. “For applications eligible for automated review, what is the 30-second effect on the probability of an authorised hold when the policy input is set to 0.80 rather than its computed value?” is an estimand.

Randomisation can make assignment independent of many rival causes. Observational identification can also be defensible when a causal graph, design and domain assumptions justify adjustment, instruments, discontinuities or other strategies. The point is not that experiments are the only route. It is that a causal estimate is an observational result plus an identification argument. The assumptions are part of the claim, not a footnote.[3]

A reversal you can calculate mentally

The worked example also exposes a limit. Even after the intervention establishes that enabling the cache reduces latency, it has not explained what latency is, how request serving is organised, or whether the cache is part of the mechanism for this service. Causal success finishes one question and opens another.

A causal estimate has an address

A result such as “X changes Y” is incomplete until its address is supplied. Which population was eligible? Which version of X was set? What was the comparator? Over what time window was Y measured? Which other pathways were held fixed? A causal effect belongs to that package. Moving the number to a new population, interface or system configuration is a transport claim that needs fresh support.

This also separates a population effect from the cause of one episode. An average intervention can change the rate of an outcome while leaving many individual cases unchanged. Conversely, a factor can matter decisively in one case even when its average effect is small because it rarely becomes active. Average causal effect, actual cause and mechanism are three outputs, not synonyms.

The lever is not the machine

A cause answers why an event or difference occurred. A constitutive explanation answers how a capacity, state or process is realised by organised parts and activities. The two often meet inside mechanisms, but they are not interchangeable.

Which “why” are you asking?

Ask why this application was held. A stale retrieval index, a low confidence input and a policy threshold may form an actual causal path. Ask instead how the deployed service has the capacity to place an application on hold. The answer now describes a configured organisation: evidence retrieval, model proposal, policy decision, authority, state transition and readback. The incident cause is one trajectory through that organisation. The constitutive explanation is the organisation that makes such trajectories possible.

Mechanistic accounts in science commonly describe entities and activities organised so that they produce a phenomenon.[5] The constitutive question is selective. A rack screw is inside the server but usually irrelevant to the decision capacity. A power feed is causally necessary for operation but may be an enabling condition rather than a component of the computational mechanism. A training corpus caused the checkpoint to acquire its parameters but is not thereby part of the runtime inference process.

A cause of the episode need not be a component of the capacity. Conversely, a component may be constitutively relevant even when its contribution is distributed and no single episode can be attributed to it alone.

Two clocks of explanation

Causal explanations usually connect differences across time, even when the interval is tiny. A configuration at one moment changes an outcome later. Constitutive explanations are commonly treated as synchronous at the relevant grain: the organised parts and the realised phenomenon occupy the same operating episode. Petri Ylikoski argues that constitution is a synchronous, asymmetric dependence, while developmental explanations can combine causal and constitutive dependencies across time.[8]

The word “synchronous” should not be mistaken for a frozen instant. A handshake protocol, oscillation or recurrent computation unfolds. The appropriate grain may be a 200 ms cycle, a complete transaction or a conversation turn. The rule is more practical: historical production and present realisation must not be collapsed into one relation. Training caused a model state. Runtime computation realises an output. Deployment configuration determines which larger system realises an authorised action.

Intervention is necessary but not self-interpreting

Mechanistic research often uses interventions across levels. Change a component and observe the whole phenomenon. Change the whole operating condition and inspect component activity. Craver’s mutual manipulability proposal tried to capture this practice as a test of constitutive relevance.[6] The appeal is clear: a genuine component should matter to the whole, and the whole’s operating states should be reflected in its components.

The difficulty is equally clear. Interventions designed for causal relations between independent variables behave strangely when one variable is part of the other. Changing the whole without changing its parts may be impossible. Changing a part may affect many things at once. Baumgartner and Gebharter call attention to “fat-handed” interventions, and later work has continued to debate how matched interlevel experiments support constitutive inference.[7][9] There is no settled universal algorithm that turns a perturbation into a constitution verdict.

That does not make constitutive explanation arbitrary. It changes the standard of evidence. Ablation shows that a system can be broken at X; reconstitution tests whether the proposed organisation can make the capacity return. A strong case combines selective perturbation, organisation change, replacement, recovery and boundary comparison. No single positive result carries the full burden.

One incident, three explanations

Interventions that lie

An intervention is evidence only for the contrast it actually creates. It can mislead when it changes several pathways, targets a proxy, destroys the operating regime or assumes the system boundary it was meant to discover.

The clean scalpel and the sledgehammer

Consider a transformer head whose ablation reduces performance on a reasoning task. The result establishes that the intervention made a difference. It does not immediately establish that the head stores “the reasoning rule”. Zeroing the head may alter activation scale, residual balance, later attention patterns and decoding. The semantic interpretation requires additional tests: targeted activation patching, alternative baselines, path-specific interventions, restoration, task controls and replication across prompts.

Mechanism claims are especially vulnerable to destructive interventions. Pulling a power cable disables every software capacity, yet electricity is not a constituent of each algorithm in the same explanatory sense as its state transitions. Deleting a database index may change the result by causing timeouts rather than by removing the represented evidence. Necessity under damage is weaker than relevance under normal organisation.

Boundary tests are interventions on the explanation

A system boundary is not just a line on an architecture diagram. It controls which dependencies count as inputs, context, components and environment. Move episodic state from a process-local store to a network service. The causal dependency may remain. The constitutive claim may change at one boundary and remain stable at another. That is why “the model remembers” is often an untyped statement. The checkpoint, prompt, external store, world state and evidence ledger have different roles.

A useful boundary test asks whether the proposed phenomenon can still be specified when the candidate component is relocated, replaced or shared. If the same external service simultaneously supports many agents, is it part of each, part of a larger joint system, or infrastructure that causally enables all of them? Evidence will not answer until the target capacity and individuation rule are explicit.

When a kill switch proves too much

From internal effect to mechanistic explanation in AI

Causal representation learning asks learned variables to support intervention, transfer and changes of regime rather than merely compress observed regularities.[10] Causal abstraction offers a useful bridge from that demand to mechanistic interpretation. A high-level algorithmic model is faithful to a lower-level network only when interventions correspond across levels and preserve the relevant effects. Recent work has generalised this framework across many interpretability methods, including activation and path patching, causal tracing, concept erasure and sparse feature interventions.[11] The contribution is not that every successful patch discovers “the” mechanism. It is a language for asking whether a proposed high-level story commutes with lower-level interventions.

The mapping itself needs discipline. If an alignment function is allowed to become arbitrarily expressive, it can map a network to an algorithm in ways that cease to be informative. Current results on the non-linear representation dilemma make this break condition explicit.[13] Intervention faithfulness requires constraints on what counts as the same variable across levels. Otherwise, the interpretation can be fitted after the fact.

A depth-aware practice

The practical mistake is usually made before analysis. A team starts with a dataset or ablation and only later decides what kind of explanation it wanted. Reverse the order. Type the claim first, then design the evidence.

The observation–intervention–constitution assay

The proposed assay is a operational synthesis, not a new metaphysical theory. It treats explanation as three cuts through the same case. Each cut adds fields that the previous one could leave unspecified.

At the trace depth, the main adversaries are measurement error, selection and unstable distributions. At the lever depth, they are confounding, interference, ambiguous treatment and poor surgicality. At the assembly depth, they are arbitrary boundaries, historical causes masquerading as current parts, generic enabling conditions and interventions that destroy the organisation they were meant to reveal.

Every depth should emit both a permitted conclusion and a prohibited inference. This one design choice prevents claim migration. It also makes negative results useful. A failed randomisation does not erase an association. A failed reconstitution does not erase a causal effect. It tells you exactly where the explanatory ladder stops.

type ClaimDepth = "association" | "causal" | "constitutive";

type EvidenceStatus =
  | "observed"
  | "identified"
  | "intervened"
  | "triangulated"
  | "contested";

interface ExplanatoryClaim {
  targetPhenomenon: string;
  candidateDependency: string;
  depth: ClaimDepth;
  populationOrSystem: string;
  boundary: string;
  timeGrain: string;
  intervention?: {
    setter: string;
    comparator: string;
    assignmentLogic: string;
    offTargetPaths: string[];
  };
  organisationEvidence?: {
    componentRole: string;
    replacementTest: string;
    restorationTest: string;
    rivalBoundary: string;
  };
  evidenceStatus: EvidenceStatus;
  permittedConclusion: string;
  prohibitedInference: string;
}

// Invariants
// 1. causal requires an intervention or an identification argument.
// 2. constitutive requires a declared boundary and organised role.
// 3. every record must state what the evidence cannot establish.

The schema exposes a logic that prose often hides: the same observation cannot silently change its claim depth, and every deeper claim must carry additional fields.

{
  "status": "awaiting input"
}

What each result means

A positive association result permits prediction inside the validated regime. It does not justify changing X. A positive causal result permits an action claim for the stated intervention, population and outcome window. It does not prove that X is the only cause, the actual cause of every episode, or a component of the target system. A positive constitutive assessment permits design and research attention to X as part of a specified mechanism. It does not establish metaphysical identity or exhaust the mechanism.

A negative result is diagnostic. If the observation is weak, improve measurement. If causal identification fails, redesign assignment or collect different data. If constitution evidence fails, clarify the boundary, seek selective interventions, attempt replacement and restoration, or retreat to a causal claim. Retreating to the claim the evidence supports is progress, not defeat.

The decision this changes

The next time a metric, lesion, ablation or patch “explains” a result, do not ask whether the evidence is impressive. Ask which decision it is meant to support.

For an alarm, ranking rule or forecasting system, stable correlation may be enough. For changing policy, treatment or architecture, you need a causal estimand and a defensible intervention or identification strategy. For replacing a module, locating a cognitive boundary, interpreting an internal circuit or claiming that a machine property constitutes a mental capacity, you need an organised mechanism model with boundary, time grain, selective intervention and recovery evidence.

The practical consequence reaches beyond terminology. It changes incident reviews, model interpretation, scientific experiments and arguments about artificial minds. A report may correlate with an internal state. An intervention may show that the state causally controls the report. Neither alone establishes that the state constitutes experience, selfhood or agency. That further claim needs a declared theory, candidate boundary and measurement bridge.

Explanation advances when the question and the test fit. The archaeological discipline is simple: read the trace, pull the lever, expose the assembly, and record where the evidence stops.