C.28.MR - Derive an Intervention Consequence by Mechanism Replacement

Type: Method pattern Status: Draft Normativity: Normative

C.28.MR:1 - Problem frame

Use this pattern when a causal model is available and the working question asks what follows if a mechanism is replaced: a controller uses a different rule, a variable is held at a chosen value, or an action occurs at a specified event. The model may be small enough to calculate by hand.

The subject is the intervention query in that model. The Method constructs the modified model and derives the requested value, distribution, contrast or bound under its assumptions. The practitioner needs the subject meaning of the variables and enough mathematical or computational help to solve the relevant equations.

First useful move. Point to the equation that represents the proposed change. Replace that equation and re-evaluate the variables needed for the question. If a display merely reports voltage, changing its displayed number leaves the supply equation in place. Changing the supply command instead changes the voltage and the current through the retained load equation.

Use an adequate existing derivation directly. Observational reporting can finish without this transformation. If only an observed joint distribution is available, recover a causal-model assumption or an identification result through C.28 before claiming a particular intervention consequence. Identifying an effect from data is a separate task from deriving it in a supplied model.

C.28.MR:2 - Problem

An observed value tells us something about how the existing mechanisms operated. An intervention changes a mechanism. Conditioning on the observed value can therefore answer a different question from imposing that value.

A second difficulty appears when a mathematical or software operation has an unclear physical interpretation. Assigning a value to an indicator, replacing a controller and forcing a physical quantity can change different mechanisms. Reusing the same variable name hides that difference.

Even a clearly specified replacement can change the solvability of a model or its subsequent dynamics. The useful result requires both the model transformation and a derivation for the question actually asked.

C.28.MR:3 - Forces

C.28.MR:4 - Solution

Choose the intervention and question → recover the mechanisms and input law → replace the selected mechanisms → solve the affected relations → obtain the requested consequence → return its interpretation and useful limit.

These moves form a derivation. A few equations can carry the complete small result.

C.28.MR:4.1 - Specify what is changed and what is asked

Name the modeled variable or mechanism, the proposed replacement and the result needed for use. Include the time or horizon when it changes the question. “Set the alarm to zero now” and “keep its output at zero for the next hour” describe different interventions in a model with later feedback.

For a constant intervention, use do(X=x) to mean replacing the equation for X by the constant x. For a changed rule, state that rule and the information it may use: for example, a controller’s command as a function of the readings available before the command. C.28:4.10 distinguishes the relevant action-policy families when that choice matters.

Specify a comparison only when one is wanted. A value under one intervention, a difference between two interventions and a difference from natural behavior have different required inputs. A question about the same historical case under another action also needs the factual observations and the resulting information about that case’s underlying inputs. Infer those inputs using the factual mechanisms and observing procedure before replacing the mechanism. Carry the inferred distribution or compatible possibilities into the alternative calculation, preserving the dependence that makes it the same modeled case. C.28 governs what supports that counterfactual use; a population intervention average can answer a different question.

C.28.MR:4.2 - Recover the model that gives the replacement meaning

Write or recover each relevant mechanism in a form such as X_i = f_i(parents, U). The function tells how the modeled quantity is determined by other modeled quantities and by U, the inputs left outside these mechanisms. Recover their domains and the joint distribution of U, or their given values for a deterministic calculation.

Keep the subject interpretation with the equations. An equation’s left-hand side is selected by the causal account. Algebraically rearranging I=V/R to V=IR preserves the equality but does not decide which physical component an intervention replaces. In :5.2 the supply controls V and the resistor relates V to I.

Preserve common inputs and their dependence. If two disturbances arise from one shared event, retain that event or their joint law when calculating both effects. A simulation must generate one compatible joint input for a case; independent resampling of its components would change the model.

Recover only the mechanisms needed to determine the queried consequence and their relevant conditions. A missing actuator relation may prevent a physical consequence while still allowing a useful comparison inside an explicitly assumed model. State that assumption where the result uses it.

C.28.MR:4.3 - Construct the modified model

Replace each targeted mechanism by its stated constant or function. Retain the other mechanism functions and the supplied input law. Then let their output values follow from the new inputs to those functions.

For a randomized replacement, give its randomization law and its dependence on existing inputs. An independent randomizer is one possible specification. If the proposed change also alters the environment or population, include that change in the compared model instead of silently treating it as part of a constant intervention.

Keep a replacement’s scope visible in time. In a discrete-time model, identify the affected update steps. In an event-driven model, identify the triggering condition and the state change it initiates. A.3.3.TR helps construct the relevant continuation rule from interactions, events and subject laws when that rule is missing.

At the physical interpretation, ask whether the proposed action also changes another modeled mechanism. Changing a supply command and attaching a device that forces its output can have different consequences for current limits and other connections. Add the relevant effects before using the modified model to answer that physical question.

C.28.MR:4.4 - Solve what the intervention query requires

For an acyclic model, evaluate the retained and replaced functions in dependency order. For a time-stepped model, propagate from its initial conditions through the selected horizon. For simultaneous equations or continuous dynamics, use the solution method and conditions appropriate to those relations.

Check the transformed relations themselves when a replacement can change existence or select among several solutions. A solver’s first returned branch may leave other admitted answers. If all admitted solutions agree on the queried quantity, that agreement can answer the question. If their answers differ, retain the possible answers, derive a useful bound, or expose the condition needed to choose among them. If the model has no relevant solution for inputs the query includes, return that obstruction and repair the model or question.

For probabilistic results, the selected solutions must define the requested random quantity under the input law. Existence of one numerical trace supplies less than a uniquely determined distribution. A general proof is unnecessary when a direct finite calculation settles the present query.

For a dynamically triggered jump, recover the event-order or flow/jump choice if it changes the result. A model’s steady-state solution alone does not determine what happens during a transient. The hybrid-system approach discussed in :11 supplies one current treatment with explicit conditions; the subject model still determines which treatment is appropriate.

C.28.MR:4.5 - Calculate the consequence under the retained conditions

Let Y_g(u) be the queried output obtained with replacement g and input u. For a finite input population with probabilities p(u), its mean is

E[Y_g] = sum_u p(u) Y_g(u).

Use the corresponding integral for a continuous input law when it exists. For a contrast between g and h in the same population, evaluate both under that population’s law. A paired simulation can reuse each joint input u for both computations. Preserve the dependence of inputs within each computation.

A symbolic expression, an inequality or a bound can be sufficient. Numerical precision should serve the receiving choice or explanation. If an estimate is required, distinguish its computational or sampling error from uncertainty about the causal model and its parameters.

C.28.MR:4.6 - Return the result and reopen only what changes it

Return the consequence together with the replacement, model assumptions and comparison or horizon needed to interpret it. The equations and answer already produced may suffice. Use C.28’s reusable support form only when another use needs a referenced conclusion.

Keep the mathematical result conditional on the supplied causal account. When the receiving question concerns an actual device or organization, connect the proposed action to the modeled replacement and ask which unmodeled interaction could change this result. Obtain further evidence only when that unresolved question can change the use. B.5.MPC connects the physical account, mathematical construction, computation and observation when this passage needs work.

When a parameter, intervention scope or mechanism changes, revisit the dependent derivation. Preserve unaffected relations and calculations. A changed event time can require another continuation; a changed source of disturbance can require another joint input law. Stop when the requested result or a useful supported limit is available.

C.28.MR:5 - Archetypal Grounding

C.28.MR:5.1 - Observing an alarm and forcing its output

Let H mean binary high load and S mean a binary alarm. Take independent uniform inputs U_H,U_S on [0,1] and the supplied mechanisms

H = 1[U_H < 0.5]

S = 1[U_S < 0.1 + 0.8 H].

Here 1[condition] is one when the condition holds and zero otherwise. In high-load cases, the alarm sounds with probability 0.9. In low-load cases, it sounds with probability 0.1.

The observational probability of no alarm is 0.5*0.1 + 0.5*0.9 = 0.5. High load with no alarm has probability 0.5*0.1 = 0.05. Thus P(H=1 | S=0)=0.05/0.5=0.1: the observation changes what we infer about the load.

For do(S=0), replace the second mechanism by S=0. The equation for H and the input law are retained. Therefore P(H=1 | do(S=0))=0.5. Forcing this indicator removes its information about H without changing H in the supplied model.

The result answers a current-load question under the stated absence of an S-to-H influence. If alarm output controls subsequent cooling, a later-load query needs that later mechanism and the duration of the override. The current calculation remains usable for the current question; the new temporal question has an additional required input.

C.28.MR:5.2 - Change a display or a supply command

Consider an ideal regulated supply and a resistive load. C is the command in volts, V the delivered voltage, D the displayed voltage, R a fixed resistance, I the load current and W its power. In the operating region being modeled,

C=12 V; V=C; D=V; R=6 ohm; I=V/R; W=V I.

The natural result is I=2 A and W=24 W.

Replacing the display equation by D=6 V leaves V=12 V; current and power remain 2 A and 24 W. Replacing the command equation by C=6 V instead gives V=6 V, I=1 A and W=6 W. The two changes share a numeral but replace different mechanisms.

Suppose the supply’s current limit is now relevant. Use the revised positive-command model V=min(C, I_max R) with I_max=1.5 A. The 12 V command now gives 9 V at the load and 13.5 W. The 6 V command still gives 6 W. Relative to natural behavior, the calculated power reduction is now 7.5 W, instead of 18 W in the earlier operating-region model.

This is a model comparison. Using this result for a physical supply requires an account of its regulation, load and limiting behavior. Altering a graphical interface value is a physical intervention only when the interface actually controls the modeled command.

C.28.MR:5.3 - A definite intervention answer with an indefinite baseline

Take real-valued variables with mechanisms X=Y and Y=X. Every equal pair satisfies the baseline equations. Without another condition, the model does not determine the baseline value of Y.

Replace the first mechanism by X=1. The retained equation now gives Y=1. That intervention consequence is determined. A difference from the baseline remains undetermined because the baseline’s value or distribution has not been supplied.

If the requested result is only Y under the intervention, return one. If the requested result is its change from natural behavior, return the missing baseline condition. Resolving the first question does not require inventing that condition.

C.28.MR:6 - Bias-Annotation

Small equation examples favor deterministic reasoning and can hide shared disturbances, partial observation and delayed effects. Carry the joint input law and intervention horizon into stochastic or dynamic applications. The directness of a software assignment can also bias the modeler toward treating any variable as independently controllable. Recover the physical or organizational mechanism that would implement the proposed change.

The examples illustrate derivation under supplied assumptions. They do not estimate how often those assumptions hold in another practice. A user can retain a useful conditional result while investigating a particular assumption that matters to its application.

C.28.MR:7 - Conformance Checklist

• The intervention and queried consequence identify the same modeled subject, population and relevant time.
• The replaced equation or rule is recoverable, including any inputs used by a policy or randomizer.
• Other mechanisms remain in force; their values are recomputed where the replacement affects them.
• The joint input law and common causes needed by the query survive the computation.
• The transformed model determines the reported answer, or the result retains the relevant alternatives, bound or obstruction.
• A comparison uses the required baseline and conditions; an intervention outcome alone is not reported as an estimated change.
• The result’s physical use connects the actual action to the modeled replacement and any consequential additional effect.
• The amount of calculation, recording and new evidence serves the receiving question.

C.28.MR:8 - Common Anti-Patterns and How to Avoid Them

C.28.MR:9 - Consequences

The practitioner obtains a reproducible model-based consequence and can locate which proposed change produces it. This supports comparisons of control rules, experimental proposals and organizational mechanisms under their stated assumptions.

The method also exposes useful partial results. A fixed intervention outcome may be available when a baseline contrast is not. A bound may already separate two choices. These results preserve progress without requiring a complete empirical programme.

The main effort lies in recovering the mechanisms and conditions needed to answer the query. More elaborate equations can increase solving cost while leaving a decisive omitted interaction untouched. A small model is preferable when it already resolves the receiving question under adequate assumptions.

C.28.MR:10 - Architectural Rationale

The construction works on mechanisms because the question concerns what would change if a different way of determining a variable operated. Statistical conditioning instead uses information about values generated by the current ways. Both operations are useful; :5.1 shows when they answer different questions.

Retaining functions while recomputing values carries the effects of a local replacement through the model. Keeping the input law retains the population and common causes against which those effects are compared. This makes the difference attributable to the specified modeled change rather than an unnoticed change of background.

Acyclic evaluation is the simplest executable case. Coupled constraints and dynamic events require attention to the transformed system’s solutions. Query-specific determinacy keeps that requirement proportionate: the relevant question can have one answer even when another part of the model is unresolved.

Direct physical enactment, empirical identification and model-based derivation provide different needed results. A supplied model is enough for the conditional derivation. Applying that consequence to the world additionally needs a suitable connection between action, model and observations. This arrangement lets a theoretical construction contribute immediately while exposing the particular additional result its physical use needs.

C.28.MR:11 - SoTA-Echoing

The corrected Pearl, Glymour and Jewell primer, p.55 provides the classical mechanism-replacement account and explains why an intervention with additional direct effects needs a richer model. It supports the ordinary constant-intervention construction here.

Bongers, Forré, Peters and Mooij (2021), Foundations of Structural Causal Models with Cycles and Latent Variables, particularly :2.4 and :3.4, shows that solvability can change under intervention. The model transformation and its induced distribution therefore need separate consideration. Its perfect-intervention results do not automatically establish the properties of every replacement policy.

Zane et al. (2025), A Counterfactual Semantics for Hybrid Dynamical Systems, :3-4, treats instantaneous, state-triggered interventions as transformations of flow and jump constraints. Under stated assumptions it preserves existence, uniqueness and measurability over the modeled horizon. This informs the event and solvability boundary; it does not impose one hybrid-system formalism on every causal query.

The synthesis retains ordinary substitution as the small executable case, adds the solution question where it matters, and connects the result to the subject interpretation. Domain-specific causal discovery, identification and estimator choices remain governed by their methods and evidence through C.28.

C.28.MR:12 - Relations

• C.28 recovers the causal-use question, distinguishes intervention and counterfactual reliance, and combines the support results needed by a receiving use. This member supplies the derivation inside a given model.
• C.28:4.10 distinguishes action-policy regimes when the replacement’s available information changes the causal question.
• A.3.3.TR constructs continuation rules from interactions, operations and subject laws when time or events matter to the replacement.
• B.5.MPC connects the physical account, mathematical construction, computation and observation used to interpret the consequence.
• B.5.RR revises the affected reasoning when a premise or mechanism changes.
• C.29.1 checks transfer of a mathematical result into another representation or use; the current member constructs the intervention result that can later be transferred.

C.28.MR:End