WorksComputer Science
From Type to Token: Petri Nets, Individual Token Philosophy, and Individualized Causal Effects
Abstract
Structural Causal Models and Bayesian Networks are the dominant framework for representing causal relations. They are fundamentally type-level formalisms since their estimands, Average Treatment Effects and Conditional Average Treatment Effects, are properties of distributions over populations, not of individual units. When treatment effects exhibit unbounded heterogeneity, these estimands collapse. What persists is the structural, token-level core of causation. We argue that Petri Nets under the Individual Token Philosophy are the natural representation of these token-level causal structures. We develop a formal translation from Bayesian Networks to Petri Nets, providing a two-layer account of individualized causality. The PN layer provides a structural representation of causal mechanisms that does not rest on distributional assumptions. In turn, a conformal prediction layer quantifies causal effects with finite-sample coverage guarantees. Together they constitute a complete account of individual causal effects.
Introduction
What does it mean for an event to be the cause of another specific event, rather than of that type of event? The philosophy of causation has long distinguished type causation, as in “smoking causes cancer”, from token causation, like in “Alice’s smoking caused her cancer”. The former is a claim about regularities across populations. The latter is a claim about a specific, concrete event.
Pearl’s framework of Structural Causal Models (SCMs) and Bayesian Networks (BNs) is the most influential formal approach to causation in both philosophy and the empirical sciences [13, 16]. It provides a unified theory of causal structure (represented by directed acyclic graphs), causal intervention (using the operator), and causal identification (through the -calculus). Yet this unification comes at a cost. The SCM framework is fundamentally type-level in its inferential output. Its estimands, the Average Treatment Effect (ATE) and the Conditional Average Treatment Effect (CATE), are expectations over distributions, hence properties of populations, not of individuals.
The SCM framework can express individual-level quantities. A unit is represented by a realization of the exogenous variables, and the unit-level effect is a genuinely token-level object [13, 9]. We claim that the machinery of the framework, its estimands, its identification results, and its inferential practice, is organized around distributions, and the individual appears in it only as an anonymous realization to be averaged over. The difference we press is one of emphasis on operational semantics, and the formal results below (Theorems 1 and 2) confirm that the two frameworks agree on their subjects.
The tension between these two approaches surfaces with particular force when treatment effects exhibit unbounded heterogeneity. When , the conditional expectation no longer represents any actual individual in the population. The type-level estimand collapses. What remains is the Individual Causal Effect (ICE) , a counterfactual quantity defined entirely at the token level, for the specific individual .
This paper argues that Petri Nets (PNs) under the Individual Token Philosophy (ITP) [19] are the natural formalism for this token-level causal structure. The argument proceeds in four steps. First (§2), we develop the philosophical diagnosis: SCMs contain a token-level primitive (the operator) but wrap it in probability, yielding type-level conclusions.
In §3, we construct the formal translation from BNs to PNs, assuming the monotonicity of activation patterns. We show how -separation maps to flow-separation in the translated net, how maps to a transition-enabling operation, and how the individual is represented by a concrete marking under ITP.
In Section §4, we develop the individualization argument. The ICE is the difference between two reachable markings from under alternative interventions, and the set of valid controls for estimating is formally characterized by equality of local subnets.
In Section §5, we prove the main formal results, correspondence theorems between BN semantics and PN semantics, and a structural foundation for ensuring coverage, using conformal guarantees [20]. We conclude in §6 by situating the framework in relation to Causal Kripke Models [21], the literature on singular causation [1, 8], and the internal ITP/CTP debate in PN theory [19, 7].
From Type to Token Causation
The type/token distinction
The philosophy of causation has recognised two fundamentally different kinds of causal claims since Lewis [1973]. Type causation is a claim about kinds of events, typically analyzed in probabilistic or nomological terms. The relata are event types, and the claim is verified by statistical regularities across populations. Token causation (also called singular causation) is a claim about specific, individual events, typically analyzed via counterfactuals. In this case, the relata are concrete event tokens and the claim is verified, if at all, by what would have happened to this event under different antecedent conditions.
Hall [2004] influentially argues that ‘causation’ names two distinct concepts: dependence (counterfactual, token-level) and production (mechanistic, process-level). This matters for our purposes because, as we argue, Pearl’s SCM framework contains both concepts but does not formally separate them, and this conflation is precisely what fails under unbounded heterogeneity.
The connection between formal causal models and the type/token distinction has received increasing attention. Erasmus [2023] shows that Causal Bayes Nets (CBNs) are abstract enough to fit any token-level theory of causation, but only under certain modelling assumptions that are not always satisfied. When those assumptions fail, accounts of actual causation based on CBNs face fundamental limitations. Our contribution is to formally specify those assumptions and to identify the formalism that obtains when they are dropped.
The SCM as an ambiguous framework
Pearl’s SCM framework has three layers that operate at different levels. The structural equations are type-level. They describe how variables relate in general, for the entire population. The function is the same for all individuals. The operator is operationally token-level: ‘I intervene and set to in this specific system.’ The physical interpretation is of a concrete manipulation of a particular system. But the output of , the interventional distribution , is immediately type-level again, an expectation over the noise averaging over all individuals who share the same structural equations.
The tension arises because uses a token-level primitive to produce type-level conclusions, skipping exactly the quantity that is often the object of interest. This is not to say that the framework cannot recover that quantity: fixing the exogenous realization yields the unit-level effect, as noted in §1. The point is that this move runs against the grain of the formalism, whose semantics and standard theorems are stated for distributions, whereas in the process semantics developed below the individual is the default object and the distribution is the derived one. Baetu [2025] argues that knowledge of singular causation depends epistemically on type-level population contrasts. We accept the epistemic dependence while denying the ontological one, since is a property of an individual , not of a population distribution. The population serves as epistemic scaffolding without being a constituent of the causal fact.
The modal-relativist analysis of Hitchcock [2024] adds further texture. What counts as the correct causal model depends on which background possibilities are held fixed, a specification that is context-dependent and not fixed by the model itself. The ‘local SCM’ of individual , which can be formalized in our framework of Section §3, as a subnetwork of the translated PN.
The empirical trajectory: from type to token by necessity
Assessing the answer to a specific query using the full population is suboptimal, only the subset of observations that share the local mechanism of are relevant [4]. This implicitly recognizes that has its own mechanism, distinct from the population mechanism. In the causal setting, this means that the global SCM does not apply to the individual .
In Delbianco and Tohmé [2026], this is formalized via conformal prediction. defines its own proxy population by structural similarity, and inference is calibrated on that local class. The conformal guarantee is distribution-free but requires exchangeability within the local class. This can be further developed, formalizing local conformal prediction for individual causal effects under distribution shift, precisely the estimation problem that the quantitative layer of the present framework, introduced in § 4, is designed to support.
A limit is reached when . Then, the Conditional ATE collapses. The conditional expectation no longer represents any individual. The only meaningful causal quantity that survives is , a purely token-level quantity. The SCM retains its structural role, but its probabilistic layer ceases to be informative.
Two token philosophies in Petri Nets: an internal mirror
Remarkably, Petri net theory has independently arrived at the same fundamental distinction. van Glabbeek et al. [ [19] identify two competing philosophies for the causal semantics of PNs:
The Individual Token Philosophy (ITP): it matters which specific token participates in each transition. If token fires transition and token fires , their causal histories will be distinct even if they originate from the same place. Causal dependence is tracked at the level of individual tokens.
The Collective Token Philosophy (CTP): only the number of tokens in each place matters, not which ones. This abstracts away individual identity. This is analogous to type causation since what matters is that some token of the right type is present.
When heterogeneity is unbounded, this CTP-analogue fails. The ICE requires ITP. The trajectory of the specific token () is what matters.
ITP was introduced to distinguish causal semantics of Petri nets [19]. What ITP supplies is exactly the formal resource that token causation demands, persistent identity and traceable causal history for the entities the formalism reasons about. Whether that resource can be made to represent individuals, rather than merely tokens, requires a bridge principle, which we state explicitly in § 3 when the individual is introduced as a marking.
Petri Nets as a Token-Level Causal Formalism
Basic formalism
A Petri Net is a tuple where is a finite set of places, is a finite set of transitions (with ), is the flow relation, and is the initial marking, assigning to each place a number of tokens.
A transition is enabled under a marking if every input place has at least one token, i.e., for all such that , . When fires, it consumes one token from each input place and produces one token in each output place, yielding a new marking . We write .
The set of markings reachable from is denoted . Note that this firing rule is conjunctive: a transition with multiple output arcs deposits a token in every output place. Exclusive branching between alternative outcomes is represented not by a single transition with multiple outputs, but by conflict, distinct transitions competing for the same input token, as is standard in free-choice nets [19].
Under the ITP, tokens have identity. Two tokens in the same place are distinguishable by their causal history, the sequence of transitions whose firings produced them. The marking under ITP records which individual tokens, with their corresponding histories, occupy each place.
The translation
A Bayesian Network (BN) is a directed acyclic graph together with a joint probability distribution over the variables , where each node represents a random variable and each directed edge encodes a direct probabilistic dependence. Under the causal Markov condition, every variable is conditionally independent of its non-descendants given its parents, allowing the joint distribution to factorize as
where denotes the set of parents of in the graph.
Given a BN with binary variables under the Markov condition, we define a translation satisfying the following condition:
Definition 1 (Monotone support condition). For , let
such that
while
Then, the components of are defined as follows:
Places: For each node , a place . A token in represents . Its absence represents .
Transitions: For each variable with parents and each minimal element we create a transition .
Arcs: For the transition :
An input arc from whenever .
An output arc to .
Initial marking: if is an active exogenous root variable, otherwise.
In , arcs from parent places to transitions are interpreted as persistent. That is, the presence of a parent node is tested without consuming its token. Otherwise a parent node with two children may not be able to activate both of them, spending a token for just one of them. This extends the PNs defined in 3.1 to what are known as contextual nets [12].1 Each transition encodes a minimal causally sufficient configuration, represented by the combination of active parents that may produce . The configurations supported in a BN are translated into reachable markings. In turn, individual firing sequences record their token-level causal realization.
is a structural translation. The Markov condition is used to construct the net, since it licenses reading each conditional as a local mechanism, and it is the resulting mechanisms. No independencies, factorizations, or probabilities appear in the translated net. The Markov condition survives in qualitative form, as the correspondence between -separation and flow-separation that will be established in Theorem 1 and that Corollary 1 makes precise, while the probabilistic layer is reintroduced, where quantification demands it, by the conformal layer of §4.
If the only payoff of were “now we have tokens”, the translation would amount to a change of notation, since any sufficiently rich formalism can encode a DAG. The payoff lies in properties that are key in PNs and not quite central in the BN formalism, yielding an operational semantics, under which mechanisms are executable and a causal history is a concrete run of the net. It also provides token identity, which grounds the representation of the individual in §3.4 below; concurrency, which distinguishes mechanisms that fire independently from those that merely happen to be uncorrelated; and the possibility of cycles, which BNs exclude by construction (§6). Each of these is put to work in what follows, and we return to their joint significance in §6.
A different line of work in theoretical computer science relates Petri nets and Bayesian networks in the opposite direction. Bruni et al. [ [3] equip Petri occurrence nets with probability distributions and use Bayesian-network inference machinery to resolve confusion, the interplay of conflict and concurrency that confuses non-deterministic choice. This is solved via a decomposition of the net into structural branching cells. Their target is a compositional probabilistic semantics for concurrent systems in general. The correspondence they establish is with the algorithms for eliminating nondeterminism. The translation developed here instead starts from a given causal BN and asks what token-level structure preserves its causal semantics, so that individual-level counterfactuals become representable. The two projects are compatible rather than competing. Their apparatus for resolving confusion could in principle be combined with to give a fully probabilistic treatment of concurrent choice within the translated net, a direction we leave for future work.
The operation in the PN
In Pearl’s framework, over a BN consists of: (i) removing all incoming edges to node , and (ii) fixing . The PN translation is defined as follows:
Definition 2 (do_PN). The intervened PN is obtained by:
Removing all arcs and disable every transition that would determine as output (severing the mechanism that normally determines ).
Setting and for all
This definition mirrors Pearl’s mutilation, and Theorem 2 certifies that the rewriting is faithful. The operation is token-level in both its operational semantics and its output. It modifies the concrete marking of individual . The interventional distribution in the BN is the type-level abstraction of this token-level operation, obtained by marginalizing over noise realizations.
The individual as a marking
Under ITP, the individual query with observed feature vector has a natural formal representation as a concrete marking of :
This marking is the causal state of ’s world, a concrete world. This concreteness is precisely what ITP makes possible and what the CTP reading of a net abstracts away. In the BN, the corresponding object is a complete instantiation of the variables, which the probabilistic semantics treats as one anonymous realization among many rather than as a first-class object of the formalism.
This representation rests on a bridge principle that deserves to be made explicit, as announced in §2. ITP guarantees that tokens have identity and causal history, but it does not by itself say that a token is an individual. We therefore adopt the following representational postulate: the individual is represented by the marking , and the identity of the tokens in tracks the identity of ’s actual features together with the causal histories that produced them. The justification of this postulate is that it supplies exactly what the ICE requires. is defined by re-running the same individual under alternative interventions, and “the same individual” is formally meaningful only in a semantics where identity persists across firings. CTP cannot ground the postulate, since it identifies markings that differ only in which tokens occupy which places. ITP is the weakest PN semantics that can achieve this.
The two counterfactual markings that define the ICE are:
The ICE is then represented structurally as:
where indicates whether carries a token in the terminal marking reachable from .
Notice that when reachability is non-deterministic, the terminal marking depends on an adequate definition of as follows:
The Individualization Argument
The ICE as a token-level quantity
The progression ATE CATE ICE is often presented as statistical refinement, from a population estimand to a conditional one, and then to an individual one. But this reading obscures the philosophically salient fact. The ATE and CATE are parameters of distributions (type-level objects). The ICE is a counterfactual fact about a concrete individual (a token-level object). A distributional parameter is estimated from population data. A counterfactual fact about is a property of , and the population serves only as epistemic scaffolding for its estimation.
This distinction determines what kind of formalism is needed. A distributional parameter can be represented in a probabilistic structure. A counterfactual fact about an individual requires a formalism where that individual is a first-class object. PNs under ITP provide exactly this. is the individual, transitions are its causal mechanisms, and reachable markings are its counterfactual worlds.
The local SCM of as a subnetwork
The empirical procedure of Delbianco and Tohmé [2021, 2026] constructs, for each query , a set of relevant controls that share the local causal mechanism of . In the language of Peters et al. [2016], this is equivalent to discovering the local SCM of , the restriction of the global SCM to the structural neighborhood of .
This operation has a precise PN representation. The local SCM of is the reachable subnetwork from in :
Definition 3 (Local Net).
where , are the transitions that can be enabled in some marking of , and .
The relevant controls are, in PN terms, the individuals whose initial marking generates the same reachable subnetwork. Under ITP, this is a property of individual tokens. Two individuals that have the same feature values but arrived at them via different causal trajectories may have different LocalNets and therefore be invalid controls for one another, even if superficially similar in covariates.
Equality of LocalNets is a strong criterion, and weaker alternatives are conceivable: isomorphism of the local subnets rather than identity, or agreement only up to a bounded radius around the treatment and outcome places. We work with exact equality because it is the weakest condition under which the invariance result of Proposition 1(b) holds without further assumptions.
Two counterfactuals as intervened markings
The fundamental problem of causal inference, that we cannot simultaneously observe and , has a particularly clear formulation in the PN framework.
Starting from the concrete marking , the two interventions and produce two mutilated nets with modified markings and . These are two causally possible worlds for , generated by the same mechanistic structure under different interventions. The impossibility of observing both simultaneously is a structural property of the PN since the same marking cannot evolve in two different directions in a single run.
The conformal prediction approach [20] estimates the unobserved counterfactual by using , individuals with compatible LocalNet under intervention , as an empirical approximation of ’s alternative world. The coverage guarantee of the conformal interval rests on the exchangeability of non-conformity scores within , which in turn rests on those individuals sharing ’s LocalNet.
What the PN formalism adds: structure without distribution
Unlike in the BN/SCM framework, where structure and distribution are entangled, in the PN, the separation is explicit. The claim ‘ causes through mechanism ’ is represented by the existence of a transition with an input arc from and an output arc to , a structural claim, verifiable without distributions. When outcome distributions are unknown or unbounded, this structure remains informative. This motivates a two-layer account:
The structural layer (PN under ITP) determines which causal mechanisms are relevant for (LocalNet()), which interventions have effect (which transitions are enabled by do_PN), and which counterfactuals are pertinent ( vs. ). This layer requires no probability distributions.
The quantitative layer (conformal prediction) determines how much ’s outcome changes under the alternative intervention, with coverage guarantee . This layer is also distribution-free but uses data, the relevant controls , as empirical reference.
The two layers articulate as follows. The structural layer delimits the domain of validity of the quantitative layer (only individuals with compatible LocalNet are valid controls), and the quantitative layer operationalizes the counterfactual that the structural layer defines qualitatively. This is illustrated in Figure 1.

Figure 1. The two-layer account of the Individual Causal Effect. The structural layer (PN under ITP) determines which mechanisms and controls are relevant to without appeal to distributions; the quantitative layer (conformal prediction) uses that structural constraint to deliver a finite-sample coverage guarantee for .
A caveat guards this claim against overstatement. What the structural layer dispenses with is distributional assumptions for representing causal structure. Quantifying the magnitude of still requires data, as the quantitative layer makes explicit. The proper contrast is therefore not “with versus without distributions” but between structural representation and statistical quantification, and the contribution of the PN formalism is to make the boundary between the two exact.
Formal Results
Preliminary definitions
Definition 4 (Path in ). A path in is an alternating sequence where each consecutive pair of places is connected by an arc in . A place is a collider node in if both adjacent arcs in are incoming: and , with and having output arcs to . It is a non-collider otherwise.
Definition 5 (Flow-active path under ITP). A path is flow-active relative to a set of places if:
Every non-collider satisfies
Every collider satisfies , or some place in , the set of places reachable by firing transitions from , belongs to
Condition (2) requires ITP: under ITP, a token in a collider place has an identifiable causal history (it arrived via one of the input transitions), so knowing the collider’s state opens information about its sources. Under CTP this is not possible, and condition (2) cannot be implemented.
Definition 6 (Flow-separation). Places and are flow-separated by in under ITP if there is no flow-active path between and relative to .
d-Separation and flow-separation
Definition 7 (Edge effectiveness).
A BN is edge-effective if for every edge there exists such that .
That is, if every edge in the BN occurs in at least one minimal sufficient configuration.
Edge-effectiveness requires every graphical edge to participate in at least one minimal sufficient activation mechanism. It is weaker than faithfulness, the property that the probability distribution in BN has no conditional independencies beyond those encoded by -separation in the graph [18], since it imposes no restrictions on conditional independencies among non-adjacent variables or on cancellations along multiple paths. It just rules out edges that are inert with respect to the positive-support semantics represented by .
Theorem 1 (-Separation Correspondence). Let be a Bayesian network satisfying the Markov condition and edge-effectiveness. Let be its PN translation under ITP. For any disjoint sets :
To prove this result, we need a previous one:
Lemma 1 (Reachability correspondence). For any , is a descendant of in the BN graph , where BN satisfies the conditions of Theorem 1 if and only if in .
Proof. By construction, implies a transition with an input arc from and output arc to , for any configuration compatible with and . Hence whenever is a child of . The general case follows by induction on the length of the directed path from to in .
Then we have: Proof of Theorem 1. By induction on path length. The base case () covers the three fundamental path types.
(i) Fork (, with a non-collider). If : condition (1) of flow-activity fails, path is blocked; in BN, follows from the Markov condition applied to the fork. If : path is flow-active. In BN, the fork path is open and induces dependence.
(ii) Chain (, with a non-collider). Analogous to the fork case by the symmetry of blocking conditions.
(iii) Collider (, with a collider). If and no descendant of is in , condition (2) of Definition 5 fails and the path gets blocked. In BN, the unconditioned collider blocks dependence. If : under ITP, the token in has identifiable causal history, it arrived via or , so knowing ’s state opens information about and . Condition (2) is satisfied. The path is flow-active. In BN, conditioning on the collider opens dependence. This is an instance of Berkson’s paradox [17].
For the inductive step, consider an arbitrary path of length . Each internal place is classified, by the direction of its two incident arcs, as a collider or non-collider. By construction of , this classification agrees with the corresponding node’s status in the BN path. By the induction hypothesis, the flow-activity conditions for the shorter subpaths obtained by removing an endpoint agree with the -separation conditions. For collider places, agreement of condition (2) additionally requires Lemma 1, which guarantees that in coincides with the descendant set of in . Since a path is active exactly when every non-collider lies outside and every collider is either already in or has a descendant in , the conjunction of these local conditions yields the equivalence for the full path. □
Corollary 1 (Soundness). Under the Markov condition, flow-separation in under ITP implies conditional independence in the BN distribution: if and are flow-separated by , then .
Theorem 1 fails under CTP. The collider case requires tokens to have identifiable causal histories, a property that CTP explicitly abstracts away. Under CTP, conditioning on cannot open the information channel between and , so the correspondence breaks precisely where the ITP/CTP distinction matters. This confirms that type-level semantics (CTP) are formally inadequate for token-level causal reasoning.
A natural objection is that flow-activity is defined so as to mimic -separation, making Theorem 1 true by construction. But notice, first, that the clause that does the distinctive work, condition (2) on colliders, is not an arbitrary stipulation but is forced by the execution semantics of the net under ITP. A token in a collider place carries an identifiable causal history, so conditioning on the collider genuinely reveals information about its sources. The definition records a fact about the semantics, and that the same clause is unimplementable under CTP shows it is not free to stipulate. Second, the theorem should be read as an adequacy result, certifying that the token-level semantics recovers exactly the independence structure of the type-level formalism. Its value is the guarantee that nothing of the BN’s qualitative content is lost by the translation.
Correspondence for the do() operator
Theorem 2 (do() Correspondence under monotone support). *Consider BN with binary variables satisfying the Markov condition. Let be a marking representing individual . For any variable and , let denote the mutilated BN, and denote
the PN intervention. Then, for every outcome variable ,
That is, belongs to the support of the interventional distribution if and only if can be marked in the intervened PN.
Proof. Consider the mutilated DAG associated with . Since the graph is acyclic, its variables can be arranged in order.
The proof is by induction. For root variables, the claim follows directly from the initial marking. The intervened variable is fixed to in the mutilated , while the marking of is fixed to the same value in the intervened .
Now consider a non-root, non-intervened variable , and assume that the correspondence has been established for all its predecessors. Suppose first that
By the factorization of the mutilated BN, there must exist a parent configuration having positive probability under the intervention such that
Hence . Since is finite, there exists a minimal element satisfying . By construction of , there is a transition whose input places are precisely the places for which .
Because has positive support, all the parents required by can occur simultaneously. By induction, the corresponding places can therefore be simultaneously marked in a reachable marking of the intervened PN. The transition is consequently enabled. Its firing produces a token in , and hence there exists with .
Conversely, suppose that becomes marked in some reachable marking of the intervened PN. Since is neither a root nor the intervened variable, the token in must have been produced by some transition , where . For this transition to be enabled, every place corresponding to a component must have been simultaneously marked. By the induction hypothesis, these active parent states occur jointly in a parent configuration having positive support under the intervention. This configuration necessarily satisfies . Since and is upward closed by the monotone support condition, it follows that and therefore
Since the parent configuration itself has positive interventional probability, the Markov factorization yields
The induction therefore establishes the equivalence for every variable in the mutilated graph.
Corollary 2 (Semantic equivalence). For propositional (binary) variables under the Markov condition, every positivity claim of the form “ is possible under ” has an exact counterpart in terms of marking reachability in .
The definition of is Pearl’s mutilation written in PN vocabulary, and Theorem 2 certifies that the rewriting is faithful. In the BN the output of mutilation is a distribution, while in the net it is a concrete marking of a concrete individual, from which the counterfactual markings of §3 are obtained directly rather than by conditioning a population-level object.
We can infer, from Corollary 2 that two BNs sharing the same graph and the same pattern of zero/nonzero conditional probabilities can differ in the exact values of those probabilities while translating to the same with identical reachability sets. Since reachability is a qualitative notion that records whether is positive or not, if we want to recover the magnitude of the intervention probability, we need an additional layer in our formalism. This is precisely what the conformal layer introduced in §4 accomplishes.
Local nets and the structural foundation of relevant controls
Proposition 1 (LocalNet and causal invariance). Let and be two individuals with markings and in . Then:
(a) if and only if and generate identical sets of reachable places and fireable transitions.
(b) If , then for any intervention and any outcome place :
where represents non-deterministic reachability. In that case, individuals with the same LocalNet have the same qualitative causal effect structure under any intervention.
Proof. (a) is immediate from Definition 3. For (b): identical LocalNets imply identical sets of fireable transitions. The PN intervention removes the same arcs in both cases (arcs depend on network structure, not on marking). The resulting initial markings differ only in , and since all downstream transitions are the same in both LocalNets, the reachable marking sets are identical.
Proposition 2 (Structural foundation for conformal exchangeability). Let be the set of super-relevant controls of under intervention . Suppose that (i) for every it holds that , and (ii) the conditional distribution of the outcome given the local mechanism is the same across individuals sharing that LocalNet (local ergodicity). Then the non-conformity scores are exchangeable with , guaranteeing conformal coverage:
Proof. Conformal coverage requires exchangeability of non-conformity scores within the calibration set. Proposition 1(b) establishes that identical LocalNets imply the same qualitative causal mechanism. By local ergodicity, individuals governed by the same mechanism draw their outcomes from the same conditional distribution, so the distribution of residuals is the same for and for . Exchangeability follows, and the standard conformal coverage guarantee [20, 2] applies.
LocalNet equality is a purely structural condition, whereas exchangeability is a distributional property that depends also on how the data are generated. What Proposition 2 establishes is the division of labor. The structural condition identifies which individuals are candidate controls,
and the ergodicity condition is the minimal probabilistic supplement under which that structural criterion yields the coverage guarantee.
Proposition 2 establishes that if controls share ’s LocalNet, then coverage is guaranteed. The empirical challenge of verifying or approximating LocalNet equality from observational data is addressed by Proposition 2, which provides the formal criterion that those heuristics approximate, namely that two individuals are valid controls for if and only if their LocalNets coincide.
Discussion
What the translation of BN to PN adds
The main thesis defended in this paper is that PNs provide an operational, identity-preserving semantics in which individual causal histories and counterfactual executions are first-class objects, while these are not first-class objects in the standard probabilistic semantics of BN/SCM. SCMs already possess a semantics for individuals via their exogenous variables and potential outcomes. The translation into PNs makes operationally explicit what the SCM leaves implicit, and that this explicitness has consequences. Four properties, none of which appears naturally in an SCM, carry the weight.
First, individual history: under ITP the token reaching carries the full record of the firings that produced it, whereas the history of a unit is not an object of the SCM formalism at all. Second, executable mechanisms: a transition is something that fires, so ’s counterfactual worlds are generated by running the net, not by solving a system of equations whose individual-level form must first be estimated. Third, concurrency and cycles: both are native to PNs and foreign to DAG-based models, as discussed below. Fourth, control selection: LocalNet() turns “valid control for ” into a checkable structural criterion, the formal target of the heuristics of Delbianco and Tohmé [2021, 2026].
Relation to the structural account of actual causation
The Halpern–Pearl structural model account of actual causation [9] might appear to already supply what we are calling for. It is evaluated relative to a fixed context , a complete assignment of the exogenous variables, and is in this sense token-level: it answers a question about this instantiation of the model, not about a population. Two differences separate our proposal from theirs. First, actual causation is a qualitative, binary judgment, whether was a cause of in context , established via counterfactual dependence under contingencies. It has no counterpart to the quantity we are after, . It does not ask how much would have differed under the alternative intervention, only whether made a difference at all. Second, and more consequential for our purposes, the Halpern–Pearl semantics presupposes that the full structural model, functional equations and context included, is already given. It is a representational theory of what causation is once the individual model is fixed, not an epistemic or estimation theory of how to obtain that individual model from population data. This is exactly the gap that opens under unbounded heterogeneity, when , population regression cannot supply a reliable structural equation for , and refining the definition of actual causation does not by itself address this estimation problem. The two-layer architecture developed here, PN structure plus conformal quantification, is designed to fill precisely this gap, i.e. the structural layer specifies which mechanisms are relevant to (LocalNet()) without assuming a parametric form, and the quantitative layer supplies a finite-sample guaranteed estimate of the counterfactual magnitude itself. The two accounts are consequently complementary rather than competing. An actual-causation judgment could in principle be evaluated on the marking once it has been estimated, but the Halpern–Pearl framework does not itself solve the estimation problem that motivates the present paper.
Relation to Causal Kripke Models
Wang et al. [ [21] and the present work share a starting point, namely that the Halpern–Pearl framework [9, 13] requires extension. But the extensions go in complementary directions.
Wang et al. extend Halpern–Pearl upward: they add modal operators and possible-world semantics to reason about causation under epistemic uncertainty. Their question is ‘what does an agent know about whether caused ?’ The result is a modal logic of actual causality with sound and weakly complete axiomatization. Their framework is type-level at its semantic base (it inherits Halpern–Pearl’s causal models) but adds an epistemic layer for reasoning about alternative possible worlds.
The present work extends Halpern–Pearl downward, to the operational-token level where causal mechanisms are concrete firings over concrete individuals. The question is ‘what is the causal mechanism of individual ?’ The result is a process semantics (PN under ITP) that is token-level from its foundation.
The complementarity is the following. Causal Kripke Models are the appropriate formalism when the problem is uncertainty about which causal model applies (multiple possible worlds, partial agent knowledge). PNs under ITP are appropriate when the causal model of is known or estimable, and the question is the token-level effect of an intervention on that specific individual. A unified theory combining both extensions, modal operators over PN markings under ITP, is a natural direction for future work.
Inverting Baetu’s epistemic argument
Baetu [ [1] argues that knowledge of singular causation rests, tacitly or explicitly, on prior contrasts between population-level patterns: we cannot know that ‘Alice’s smoking caused her cancer’ without invoking type-level regularities about smokers and cancer. Token causality is epistemically dependent on type causality.
We accept this epistemic dependence while denying the ontological one. To estimate we need a reference population, the relevant controls . In this sense, knowledge of the ICE depends on population comparisons. But this epistemic scaffolding does not constitute the causal fact itself. The PN under ITP makes this distinction explicit. and its reachable markings are properties of the individual (token-level ontology). is the epistemic scaffolding that permits estimation when the counterfactual cannot be directly observed.
Hall’s two concepts and the PN formalism
Hall’s [ [8] dichotomy between dependence (counterfactual, token-level) and production (mechanistic, process-level) is often presented as a challenge for unified theories of causation. Any single formalism seems to capture one at the expense of the other.
PNs under ITP integrate both. The dependence dimension is captured by the difference in reachable markings: is an overtly counterfactual quantity. The production dimension is captured by token flow, a concrete sequence of transition firings, a causal process that carries tokens from the intervention place to the outcome place. Unlike Hall’s dichotomy, the PN formalism does not force a choice since a single run of the PN under ITP is simultaneously the mechanism of production (token trajectory through transitions) and the ground of counterfactual dependence (the reachable marking that defines what would have occurred under the alternative intervention).
Extensions and open problems
Continuous variables. The translation assumes binary variables. For continuous covariates , the formalism extends naturally via colored PNs (where tokens carry numerical values) or weighted PNs (where arcs carry weights). Transitions become functions transforming values rather than simply propagating tokens. ITP persists in this case, only the implementation changes.
Non-monotone mechanisms. When the conditions of Definition 1 are not satisfied, non-minimal activation patterns of may not belong to and thus, there might exist transitions that cannot be activated for certain configurations of . This requires an alternative characterization of admissible transitions.
Cyclic causation. PNs handle cycles naturally, unlike BNs, which are acyclic by construction. For feedback systems, where the outcome of an intervention retroacts on covariates, PNs offer a direct representation that BN require extensions to achieve.
The ergodicity condition. Proposition 2 relies on local ergodicity. Characterizing sufficient conditions for this assumption in terms of properties is an open problem connecting PN theory with the transportability literature [14] and invariant causal prediction [15].
Weakening LocalNet equality. As noted in § 4, exact equality of LocalNets could be relaxed to isomorphism or to agreement within a bounded radius of the treatment and outcome places. Determining which relaxations preserve Proposition 1(b), and at what cost in the coverage guarantee of Proposition 2, would widen the usable control sets in applications.
Conclusions
This paper has argued for a precise thesis: when the object of causal inquiry is the Individual Causal Effect , the appropriate formal framework is Petri Nets under the Individual Token Philosophy, not Bayesian Networks under probabilistic semantics. SCMs can define the ICE through realizations of their exogenous variables, and our correspondence theorems show that the two frameworks agree wherever both apply. Our central claim is that the PN formalism puts the individual, its history, and its executable mechanisms in the foreground, where the SCM framework reaches them only against the grain of its distributional semantics.
The argument rests on a philosophical diagnosis and a technical correspondence. The diagnosis (§ 2) is that Pearl’s SCM framework contains a token-level primitive, the operator, but wraps it in distributional expectations and returns to type-level conclusions. When treatment effects have unbounded heterogeneity, the distributional layer collapses while the structural layer remains intact. The ICE is the quantity that survives, a token-level, counterfactual fact about a specific individual.
The technical correspondence (§ 3–§ 5) is that PNs under ITP reproduce the full causal semantics of BNs, -separation (Theorem 1) and the operator (Theorem 2), while being natively token-level. The ITP/CTP distinction in PN theory (§ 2) is the formal counterpart of the type/token distinction in philosophy of causation, and the collider case in the proof of Theorem 1 is where this distinction does formal work.
The individualization argument (§ 4) shows that the set of valid controls for estimating is formally characterized by equality of local subnets (). Proposition 2 shows that this structural condition is exactly what is needed for the conformal coverage guarantee to hold.
The framework yields a principled two-layer account of individualized causality: a structural layer (PN under ITP) that provides causal architecture without distributional assumptions, and a quantitative layer (conformal prediction) that provides finite-sample coverage guarantees without parametric assumptions.
Three directions stand out for future work. The most technically demanding is to replace the local ergodicity assumption in Proposition 2 with verifiable conditions on . The most conceptually interesting is to develop a unified framework combining the modal-epistemic extension of Wang et al. [ [21] with the token-operational extension of the present work. The most empirically immediate is to implement the criterion as an algorithmic guide for control selection, replacing the heuristic similarity measures currently used.
Declarations
Funding: This work was supported by project PGI UNS 24-E172, Secretaría General de Ciencia y Técnica, Universidad Nacional del Sur.
Competing interests: The authors have no competing interests to declare that are relevant to the content of this article.
Data, materials and/or code availability: No datasets were generated or analysed in the current study.
Author contributions: All authors contributed to the conception and formal development of the paper, and read and approved the final manuscript.
References
References
- [1]Baetu, T. M. (2025). The epistemic problem of singular causation. Synthese, 205:16.DOI
- [2]Barber, R. F., Candès, E. J., Ramdas, A., and Tibshirani, R. J. (2021). Predictive inference with the jackknife+. Annals of Statistics, 49(1):486–507.arxiv.org/abs/1905.02928
- [3]Bruni, R., Melgratti, H. C., and Montanari, U. (2020). Bayesian network semantics for Petri nets. Theoretical Computer Science, 807:95–113.DOI
- [4]Delbianco, F. and Thomé, F. (2021). A methodology to answer individual queries: finding relevant and robust controls. Behaviormetrika, 48(2):377–400.DOI
- [7]Goltz, U. and Reisig, W. (1983). The non-sequential behaviour of Petri nets. Information and Control, 57(2–3):125–147.DOI
- [8]Hall, N. (2004). Two concepts of causation. In Collins, J., Hall, N., and Paul, L. A. (Eds.), Causation and Counterfactuals, pp. 225–276. MIT Press, Cambridge.DOI
- [9]Halpern, J. Y. (2016). Actual Causality. MIT Press, Cambridge.
- [12]Montanari, U. and Rossi, F. (1995). Contextual nets. Acta Informatica, 32(6): 545–596.
- [13]Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press, Cambridge.
- [14]Pearl, J. and Bareinboim, E. (2014). External validity: From do-calculus to transportability across populations. Statistical Science, 29(4):579–595.
- [15]Peters, J., Bühlmann, P., and Meinshausen, N. (2016). Causal inference by using invariant prediction: Identification and confidence intervals. Journal of the Royal Statistical Society: Series B, 78(5):947–1012.arxiv.org/abs/1501.01332
- [16]Peters, J., Janzing, D., and Schölkopf, B. (2017). Elements of Causal Inference. MIT Press, Cambridge.
- [17]Sharma, R., Kaushik, M., Peious, S.A., Shahin, M., Vidyarthi, A., Tiwari, P., and Draheim, D. (2022). Why not to trust big data: Discussing statistical paradoxes, pp. 50–63 in International Conference on Database Systems for Advanced Applications, Springer.DOI
- [18]Spirtes, P., Glymour, C., and Scheines, R. (2001). Causation, Prediction, and Search. The MIT press.
- [19]van Glabbeek, R., Goltz, U., and Schicke, J.-W. (2021). On causal semantics of Petri nets. arXiv preprint arXiv:2103.00729.arxiv.org/abs/2103.00729
- [20]Vovk, V., Gammerman, A., and Shafer, G. (2005). Algorithmic Learning in a Random World. Springer, New York.
- [21]Wang, R., Ding, Y., Manoorkar, K., Tzimoulis, A., and Wang, X. (2025). Causal Kripke models. Synthese, 206:133.