FPF Reference

CausalUse-CAL: Causal-Use Questions, Causality-Ladder Rungs, Identification and Realizability

C.28 · Cluster C.IV - Composite & Macro-Scale · Stable · Calculus (C) · Normative unless explicitly marked informative · Part C - Kernel Extension Specifications

About this pattern

This is a generated FPF pattern page projected from the published FPF source. It is canonical FPF content for this ID; it is not a fpf-memory product feature page.

How to use this pattern

Read the ID, status, type, and normativity first. Use the content for exact wording, the relations for adjacent concepts, and citations to keep active work grounded without pasting the whole specification.

Type: Calculus (C) Status: Stable Normativity: Normative unless explicitly marked informative

Plain-name. Causal-use calculus.

Intent. Govern live causal questions and decision-bearing causal use: improvement, intervention effect, causal fairness, counterfactual comparison, causal policy optimality, realized counterfactual-rung evidence, identified counterfactual estimate, or simulation-only counterfactual output.

Governed object. A causal-use question or claim together with the support basis and follow-on records needed to use it lawfully: causality-ladder rung, estimand or contrast, evidence support basis, identification posture, counterfactual sampling realizability posture, supported use, unsupported use, and next move.

Not a physical ontology. C.28 governs how FPF authors, reviewers, and operators use causal models, evidence, and counterfactual reasoning in records, policy claims, fairness claims, method comparisons, and work plans. It does not define physical causality in general and does not replace local domain science.

Use C.28 when a claim is being used causally:

Keywords

  • causal-use question
  • causality ladder
  • association
  • intervention
  • counterfactual
  • Pearl Causal Hierarchy
  • Structural Causal Model
  • causal diagram
  • causal estimand
  • identification
  • counterfactual sampling realizability
  • causal evidence support basis
  • target trial
  • causal fairness
  • off-policy causal evaluation
  • causal-RL evaluation.

Relations

C.28builds onEvidence Graph Referring (C-4)
C.28builds onTrust & Assurance Calculus (F–G–R with Congruence)
C.28builds onDecision Theory (Decsn-CAL)
C.28builds onExplore–Exploit Governor (E/E‑LOG)
C.28builds onAgentic Tool‑Use & Call‑Planning (C.Agent‑Tools‑CAL)
C.28builds onQuantum-Like Modeling Lens
C.28builds onTemporal Claim Adequacy: State Readings, Temporal Trends, and Intervention-Sensitive Temporal Change
C.28builds onBias-Audit & Ethical Assurance
C.28builds onMulti‑Method Dispatcher & MethodFamily Registry
C.28builds onParity / Benchmark Harness
C.28coordinates withU.EvidenceRole: The Evidential Stance
C.28coordinates withU.MethodDescription: The Recipe for Action
C.28coordinates withSignature Stack & Boundary Discipline
C.28coordinates withRole-Method-Work Alignment (Contextual Enactment)
C.28coordinates withMM-CHR — Measurement & Metrics Characterization
C.28coordinates withTelemetry-Driven Refresh & Decay Orchestrator
C.28explicit referenceMM-CHR — Measurement & Metrics Characterization
C.28explicit referenceTemporal Claim Adequacy: State Readings, Temporal Trends, and Intervention-Sensitive Temporal Change
C.28explicit referenceTrust & Assurance Calculus (F–G–R with Congruence)
C.28explicit referenceEvidence Graph Referring (C-4)
C.28explicit referenceParity / Benchmark Harness
C.28explicit referenceDecision Theory (Decsn-CAL)
C.28explicit referenceExplore–Exploit Governor (E/E‑LOG)
C.28explicit referenceAgentic Tool‑Use & Call‑Planning (C.Agent‑Tools‑CAL)
C.28explicit referenceQuantum-Like Modeling Lens
C.28explicit referenceMulti‑Method Dispatcher & MethodFamily Registry
C.28explicit referenceBias-Audit & Ethical Assurance
C.28explicit referenceU.EvidenceRole: The Evidential Stance
C.28explicit referenceSignature Stack & Boundary Discipline
C.28explicit referenceBoundary Norm Square (Laws / Admissibility / Deontics / Work-Effects)
C.28explicit referenceContract Unpacking for Boundaries

Content

Use This When

Use C.28 when a claim is being used causally:

  • "method A improves result";
  • "users who received intervention X had better outcomes";
  • "this practice is fair";
  • "the agent chose optimally";
  • "the model simulates what would have happened";
  • "the system can collect counterfactual data";
  • "this benchmark shows a causal method is better";
  • "this policy should be deployed because it would have changed the outcome".

Use C.28 especially when the claim must distinguish:

  • observed association;
  • intervention or action effect;
  • counterfactual comparison;
  • direct counterfactual-rung data collection;
  • identified counterfactual estimate;
  • simulation-only counterfactual output;
  • causal policy class;
  • causal fairness use;
  • causality-ladder parity in method comparison.

Not this pattern when. If no causal use is claimed, keep the work in the neighboring pattern: C.16 for measurement, C.27 for temporal trend or rate-change adequacy, B.3 for assurance posture, A.10 for evidence graph reference, G.9 for ordinary parity, C.11 for local choice, C.19 for pool policy, C.24 for call planning, or C.26 for a surviving quantum-like modeling cue after ordinary causal explanations have been tried.

Activation boundary. C.28 activates at CausalUseActivation: causal wording changes what the claim makes admissible for publication, choice, deployment, assurance, audit, benchmark, or support treatment. The trigger is admissible downstream use, not the presence of a causal-looking word. If the wording is only exploratory prose and no stronger use is made, rewrite to association, trend, measurement, or simulation-only wording and stop.

Exploratory causal-looking prose is not a CausalUseActivation by itself. A note may say that a relation is plausible, worth probing, or suggested by traces and still remain in C.16, C.27, A.10, C.11, C.19, C.24, G.5, or G.9 until the text makes a stronger causal use admissible. The moment the text makes publication, choice, deployment, assurance, audit, benchmark, or support treatment depend on causal support, C.28 owns the causal-use boundary.

What Goes Wrong If Missed

A weak causal phrase gets promoted into a strong causal use.

Correlation becomes intervention effect. Interventional proxy becomes counterfactual fairness. A simulation becomes realized counterfactual-rung evidence. A benchmark compares methods across different causality-ladder rungs and still publishes one scalar superiority claim. An agentic policy is called optimal without saying whether it is a natural behavior policy, an interventional policy, or a counterfactual policy.

The practical error is laundering: the reader sees causal language but cannot recover what rung, estimand, evidence basis, and supported use are actually admissible.

What This Buys

C.28 gives FPF one cheap first stop for causal use.

The first useful result is not a heavy record. It is one small causal-use triage that says whether causal use is present, which causality-ladder rung is being used, what comparator or counterfactual is in play, what evidence posture supports it, and what the next move is.

Durable cards and profiles appear only when the claim needs them. The pattern buys stronger causal discipline without turning every causal word into a paperwork exercise.

First-Minute Questions

C.28 in 60 seconds is the operational entry into CausalUseTriageRecord:

  1. Detect whether the claim reaches CausalUseActivation: it changes what publication, choice, deployment, assurance, audit, benchmark, or support treatment is admissible.
  2. Stop with nextMove.cheapStop if the claim only reports association, trend, description, measurement, or simulation-only output.
  3. If causal use is live, fill targetCausalityLadderRung, comparatorOrCounterfactualRef, and evidencePosture.
  4. Fill supportedUse: CausalUseSupportStatement and unsupportedUse: CausalUseUnsupportedStatement as one action pair.
  5. Fill nextMove: CausalUseNextMove: choose cheapStop or escalate only when the claim is decision-bearing, publication-bearing, assurance-bearing, fairness-bearing, benchmark-bearing, or reusable.

First Output

The first output is a CausalUseTriageRecord:

CausalUseTriageRecord:
  causalUse: yes | no | unclear
  targetCausalityLadderRung?: CausalityLadderRung
  comparatorOrCounterfactualRef?
  evidencePosture: CausalEvidencePostureTriageValue
  supportedUse?: CausalUseSupportStatement
  unsupportedUse?: CausalUseUnsupportedStatement
  nextMove: CausalUseNextMove
CausalUseNextMove:
  cheapStop:
    stopNoCausalUse |
    publishAssociationOnly |
    rewriteAsTrendOrAssociation |
    keepSimulationOnlyModelUse |
    downgradeCausalWording |
    abstainFromCausalUse |
    rerouteToNeighborPattern
  escalateOnlyIfLoadBearing:
    openLocalCausalUseQuestionCard |
    openDurableCausalUseQuestionCard |
    buildCausalIdentificationProfile |
    buildCounterfactualSamplingRealizabilityProfile |
    planCausalUseEvidenceDesign |
    openCausalFairnessUseAuditCard |
    openCausalMethodRungParityRecord
CausalEvidencePostureTriageValue =
  observationalAssociationSupportBasis |
  interventionalActionSupportBasis |
  realizedCounterfactualSampleSupportBasis |
  identifiedCounterfactualEstimateSupportBasis |
  simulationOnlyCounterfactualOutputBasis |
  missing

cheapStop values are terminal or downgrade actions. They close the local causal-use question for now by saying what weaker use remains admissible, which neighbor owns the remaining non-causal question, or that causal use is declined. escalateOnlyIfLoadBearing values are record-opening actions. They are admissible only when the supported/unsupported-use boundary cannot safely carry the reader's next action by itself.

If this first output cannot be written honestly, the causal-use claim is not ready.

CausalUseSupportStatement is one concrete causal-use action the current support makes admissible, such as publish association-only wording, use a bounded interventional estimate for a named decision, deploy only under a named policy constraint, run a fairness audit under a named causal estimand, or compare methods only inside one declared causality-ladder rung. It is not a confidence label, graph name, method name, or generic "evidence exists" phrase.

CausalUseUnsupportedStatement is the matching concrete causal-use action the current support does not make admissible, such as intervention-effect wording, realized counterfactual sample wording, causal fairness certification, causal policy optimality, cross-rung benchmark superiority, or release/deployment use. The supported and unsupported statements travel as a pair so the reader can act without inferring the boundary from prose tone.

The triage record may be the final causal-use carrier. Triage lines are enough when they block the overclaim and tell the reader what weaker use remains admissible. Do not open a local card merely because the word "cause", "effect", or "counterfactual" appears.

The triage evidencePosture field is the first-pass alias for CausalEvidenceSupportBasis | missing. If a claim escalates beyond triage, the value must dock to CausalEvidenceSupportBasis; missing becomes unsupportedUse, CausalUseSupportVerdict = unsupported, or abstain.

Problem Frame

FPF already has strong neighboring patterns for measurement, evidence, assurance, temporal claims, decisions, exploration, call planning, fairness, method dispatch, parity, and quantum-like modeling. None of those neighbors should become the general authority for causal use.

C.28 exists because causal use cuts across those neighbors. The same sentence can be:

  • a measurement description handled by C.16;
  • a temporal trend handled by C.27;
  • an assurance claim handled by B.3;
  • an evidence graph reference handled by A.10;
  • a decision record handled by C.11;
  • a pool-policy record handled by C.19;
  • a call-planning record handled by C.24;
  • a fairness audit handled by D.5;
  • a parity report handled by G.9;
  • a quantum-like residual handled by C.26;
  • or a causal-use claim governed here.

The first pattern task is therefore not to classify wording for its own sake. It is to recover the live causal question, the target causality-ladder rung, the support basis currently available, and the cheapest truthful next move. Sometimes that move is to downgrade the claim to association, temporal change, metric-only fairness, or simulation-only use. Sometimes it is to open identification, realizability, evidence-design, fairness, policy-evaluation, or benchmark-parity work. C.28 exists to keep those moves distinct and to stop teams from acting as if a stronger causal basis had already been earned.

Problem

Causal language is easy to overclaim because ordinary prose hides the difference between association, action, counterfactual comparison, realized counterfactual sample, identified estimate, and simulation.

Three collapses are especially dangerous:

  1. Rung collapse. Observational association, interventional action/effect, and counterfactual comparison are treated as one causal strength.
  2. Support collapse. Observed data, experimental data, direct counterfactual-rung samples, identified estimates, and simulations are treated as one evidence basis.
  3. Use collapse. A result that supports one use, such as association reporting, is reused for another use, such as causal fairness, policy optimality, or method superiority.

C.28 prevents those collapses by making rung, support, and use explicit before stronger claims are admitted.

Forces

ForceTension
Causal safety vs cognitive affordabilityFPF must block causal laundering without forcing every causal word into a full causal dossier.
Rung clarity vs ordinary languageOrdinary language says "improves", "causes", "fair", or "would have"; FPF must recover whether that means association, intervention, or counterfactual comparison.
Identification vs realizabilityA counterfactual estimand may be identifiable from other data but not directly sampleable, or directly sampleable under action constraints but not generally available.
Graph/formalism precision vs reader usabilitySCM, DAG, ADMG, SWIG, SCM twin network, AMWN, and counterfactual graphical model names matter, but they must not bury the first practical move.
Domain plurality vs one FPF patternSCM/PCH, potential outcomes, target-trial emulation, causal ML, transportability, causal representation learning, causal RL, and causal fairness must all remain recognizable without making C.28 a one-school vocabulary.
Neighbor fit vs authority creepNeighbor patterns need causal-use hooks, but they must not redefine causal-use question, rung, estimand, identification, or realizability.

Solution

Use a three-level causal-use escalation:

  1. Start with CausalUseTriageRecord.
  2. Escalate to LocalCausalUseQuestionCard or DurableCausalUseQuestionCard only when the claimed use needs a reusable causal-use record.
  3. Add profiles or specialized records only when the claim triggers that exact need: identification, realizability, evidence design, fairness, policy evaluation, transportability, estimation validity, causal-variable representation, or parity.

The default move is cheap. The heavy move is triggered.

Record or profile kindOrdinary sizeTrigger
CausalUseTriageRecordone short record; usually 5-8 lines covering activation, rung, comparator/counterfactual, evidence posture, supported/unsupported use pair, and next moveany live causal wording or suspected causal laundering
LocalCausalUseQuestionCardone small card; usually one causal-use question, one rung, optional comparator/estimand, one support basis, one supported/unsupported use pair, and one next movethe team needs a reusable local record but not a publication/release/fairness/benchmark/assurance object
DurableCausalUseQuestionCardone durable card with causal-use kind, estimand, timing/outcome when needed, assumptions, rival causes, support basis, supported/unsupported use pair, next move, and reopen/exit conditionthe claim is decision-bearing, publication-bearing, fairness-bearing, benchmark-bearing, assurance-bearing, or reusable
heavy profile or specialized recordonly the fields needed for the named triggered question or work item; absent fields remain absent rather than becoming implied dossier requirementsidentification, realizability, target-trial emulation, parameter estimation, transportability, off-policy evaluation, causal representation, evidence design, fairness audit, or causal parity is materially needed

Semantic Authority and Consumer Carry-Through Boundary

C.28 is the semantic authority for causal-use objects and causal-use support roles. Neighbor patterns keep their local authority and consume only the causal-use pieces they need: measurement, evidence path, assurance, fairness, decision, exploration, call-planning, dispatch, parity, and refresh records do not become causal-use owners by carrying C.28 fields.

Object or decisionC.28 ownsNeighbor may carryNeighbor must not do
Causal-use kind and rungCausalUseClaimKind, CausalityLadderRung, causal-use question, comparator/counterfactual, estimand, supported use, unsupported usecausalUseSpec?, causalActionUseSpec?, method dispatch spec, parity record, fairness audit cardInfer causal-use kind from local vocabulary alone or publish a stronger rung without C.28 support
Causal evidence support basisCausalEvidenceSupportBasis and its five valuesEvidence path refs in A.10, evidence-role specializations in A.2.4, consumer fields in B.3, C.19, D.5, G.5, and G.9Mint another support-basis value set, add assumption-only/no-support values, or let simulation-only output become realized evidence by name
Identification and realizabilityCausalIdentificationProfile, CounterfactualSamplingRealizabilityProfile, their verdicts, and supported/unsupported useEvidence, assurance, decision, exploration, call-planning, fairness, dispatch, and parity refs to those profilesTreat identification as direct sampling, or treat direct-sampling infeasibility as absence of all possible causal support
Graph and calculus namingCausalGraphRepresentationKind, GraphSeparationCriterionKind, CausalInferenceCalculusKind, StructuralCausalModel, CausalDiagramRefNamed graph/calculus refs when the neighbor records the causal-use support basis and cited formalismUse generic graph prose where a graph formalism or calculus is load-bearing
Assurance consequenceCausalUseSupportVerdict as causal-use action grammarB.3 degrade/block/abstain consequences for F-G-R/CL assuranceLet assurance prose certify causal identification, realizability, or fairness
Fairness, policy, and parity specializationCausal-use question/rung/estimand/support basis/support verdict for fairness, policy, and causal method comparisonD.5 ethical/fairness audit card, C.11 choice result, C.19 pool policy, C.24 call plan, G.5 method dispatch spec, G.9 parity report with local refs to consumed C.28 supportCollapse metric disparity, policy replay, method dispatch, or benchmark score into a causal-use verdict

A neighbor may quote the C.28 values it consumes for by-value readability. Quoting the values is not ownership. A neighbor owns only its local record and must cite C.28 when the causal-use question or causal-support basis is live.

Compact crosswalk:

Field or decision slotQuestion answeredTypical valuesDo not confuse with
CausalityLadderRungWhat kind of causal question or use is being claimed?observation, intervention/action, counterfactual comparisonthe evidence source or the method family
CausalEvidenceSupportBasisWhat support posture is being used for that causal use?observational association, interventional action, realized counterfactual sample, identified counterfactual estimate, simulation-only outputthe rung itself, a raw evidence-role name, or a no-support verdict
supportedUse / unsupportedUseWhat may the reader do next, and what must they not do?CausalUseSupportStatement, CausalUseUnsupportedStatementa confidence score, a graph name, a method name, or a local pattern owner

Rung-support-use examples:

RungSupport basisSupported useUnsupported use
observationalAssociationRungobservationalAssociationSupportBasisassociation report, descriptive risk comparison, probe selectionintervention-effect claim, causal fairness certification, policy optimality
interventionalActionRunginterventionalActionSupportBasisdeclared action-effect use inside assignment, follow-up, and outcome limitscounterfactual sample claim, cross-population policy claim without transportability
counterfactualComparisonRungidentifiedCounterfactualEstimateSupportBasisidentified or bounded counterfactual estimate under assumptions and profile refsrealized sample wording or assumption-free counterfactual certainty
counterfactualComparisonRungsimulationOnlyCounterfactualOutputBasisbounded model-supported simulation userealized counterfactual sample evidence, intervention-effect evidence

Causality-Ladder Rung

CausalityLadderRung is a controlled value set:

CausalityLadderRung =
  observationalAssociationRung |
  interventionalActionRung |
  counterfactualComparisonRung
  • observationalAssociationRung means passive observation, natural behavior, association, or seeing-only posture.
  • interventionalActionRung means do(x), intervention, action setting, experiment, policy change, or action-effect posture.
  • counterfactualComparisonRung means counter-to-fact comparison, unit-history-conditioned comparison, potential-outcome contrast, or imagining posture.

A stronger rung is not supported by weaker data unless a CausalIdentificationProfile, CounterfactualSamplingRealizabilityProfile, or bounded-use statement says exactly what is supported and what is not.

Causal-Use Claim Kind

CausalUseClaimKind is the controlled value set for the local causal-use claim being made:

CausalUseClaimKind =
  causalEffectClaim |
  counterfactualComparisonClaim |
  causalFairnessClaim |
  causalPolicyClaim |
  causalBenchmarkParityClaim |
  causalEvidenceSupportClaim |
  causalAssuranceSupportClaim
  • causalEffectClaim means a result is used as an effect, improvement, harm, or intervention/outcome claim.
  • counterfactualComparisonClaim means a counter-to-fact, potential-outcome, or unit-history-conditioned comparison is being used.
  • causalFairnessClaim means fairness is claimed through a causal path, intervention, counterfactual, or causal estimand rather than only a metric.
  • causalPolicyClaim means a policy, action rule, exploration rule, or agentic strategy is claimed as causally preferable.
  • causalBenchmarkParityClaim means causal methods are compared for parity, superiority, or benchmark consumption.
  • causalEvidenceSupportClaim means an evidence path is being used as causal-use support.
  • causalAssuranceSupportClaim means an assurance tuple or support verdict is being used for a causal-use claim.

Simulation-only causal use stays inside the existing claim-kind set. simulationOnlyCounterfactualOutputBasis is a support/use posture, not a new CausalUseClaimKind. Use the relevant claim kind, usually counterfactualComparisonClaim, causalPolicyClaim, causalBenchmarkParityClaim, or causalEvidenceSupportClaim, and set CausalEvidenceSupportBasis = simulationOnlyCounterfactualOutputBasis with bounded model-supported use and unsupported use. Bounded model-supported simulation use does not become realized counterfactual sample evidence or intervention-effect evidence. Do not mint a separate simulation-only claim kind merely to avoid naming the support posture.

Encoding rule: choose the causal-use claim kind by the question being answered, then choose simulationOnlyCounterfactualOutputBasis as the support basis and write CausalUseSupportStatement / CausalUseUnsupportedStatement for the bounded simulation use.

Causal-Use Cards

Use a local card when the claim needs a small working record:

LocalCausalUseQuestionCard:
  causalUseQuestionRef: U.CausalUseQuestion
  targetCausalityLadderRung: CausalityLadderRung
  causalUseClaimKind?: CausalUseClaimKind
  comparatorOrCounterfactualRef?
  estimandRef?
  causalEvidenceSupportBasis: CausalEvidenceSupportBasis
  supportedUse: CausalUseSupportStatement
  unsupportedUse: CausalUseUnsupportedStatement
  nextMove: CausalUseNextMove

Use a durable card when the claim is decision-bearing, publication-bearing, fairness-bearing, benchmark-bearing, assurance-bearing, or reusable:

DurableCausalUseQuestionCard:
  causalUseQuestionRef: U.CausalUseQuestion
  targetCausalityLadderRung: CausalityLadderRung
  causalUseClaimKind: CausalUseClaimKind
  causalInterventionSpecRef?
  comparatorOrCounterfactualRef?
  estimandRef: U.CausalEstimand
  potentialOutcomeContrastRef?
  targetTrialProtocolRef?
  assignmentOrInterventionWindowRef?
  causalFollowUpWindowRef?
  outcomeMeasureRef?
  causalAssumptionSetRef
  rivalCauseSetRef?
  causalEvidenceSupportBasis: CausalEvidenceSupportBasis
  causalIdentificationProfileRef?
  counterfactualSamplingRealizabilityProfileRef?
  causalParameterEstimationProfileRef?
  causalTransportabilityProfileRef?
  causalVariableRepresentationRef?
  falsificationOrNegativeControlRef?
  sensitivityAnalysisRef?
  rivalCauseStressTestRef?
  supportedUse: CausalUseSupportStatement
  unsupportedUse: CausalUseUnsupportedStatement
  nextMove: CausalUseNextMove
  reopenOrExitCondition

The durable card is not the default. It is the record used when a weak causal note would be unsafe.

Causal Evidence Support Basis

CausalEvidenceSupportBasis is a controlled value set:

CausalEvidenceSupportBasis =
  observationalAssociationSupportBasis |
  interventionalActionSupportBasis |
  realizedCounterfactualSampleSupportBasis |
  identifiedCounterfactualEstimateSupportBasis |
  simulationOnlyCounterfactualOutputBasis

This is the C.28-governed value set for causal evidence support basis. causalAssumptionOnlySupport and noCausalEvidenceSupport are not values of CausalEvidenceSupportBasis: assumption-only posture belongs in causalAssumptionSetRef plus supported/unsupported use; no-support posture belongs in CausalUseSupportVerdict, unsupportedUse, or abstain.

Simulation-only output never becomes realized counterfactual-rung evidence by name alone. It may support model-based use only when assumptions, validation, and supported/unsupported use are declared.

realizedCounterfactualSampleSupportBasis does not mean observing two incompatible outcomes for the same unit in one realized world. It means physically obtaining samples from the declared target counterfactual distribution under the profile's constraints.

CausalEvidenceSupportBasis names a support posture. It is distinct from an evidence source, an A.2.4 evidence role, and an A.10 evidence path. Some support bases are direct empirical postures, such as observational or interventional support. Other support bases are inferential postures, such as identified counterfactual estimate support. Do not read this value set as only a raw evidence-source kind.

realizedCounterfactualSampleSupportBasis does not mean observing two incompatible outcomes for the same unit in one realized world. It means physically obtaining samples from the declared target counterfactual distribution under the profile's physical, ethical, operational, unit-history, and graph constraints.

Identification Profile

CausalIdentificationProfile answers whether a causal or counterfactual estimand can be expressed from available data plus assumptions, graph representation, and inferential calculus.

CausalIdentificationProfile:
  causalUseQuestionRef: U.CausalUseQuestion
  estimandRef: U.CausalEstimand
  targetCausalityLadderRung: CausalityLadderRung
  sourceCausalEvidenceSupportBasis?: CausalEvidenceSupportBasis
  structuralCausalModelRef?: StructuralCausalModelRef
  causalDiagramRef?: CausalDiagramRef
  causalGraphRepresentationKind?: CausalGraphRepresentationKind
  graphSeparationCriterionKind?: GraphSeparationCriterionKind
  causalInferenceCalculusKind?: CausalInferenceCalculusKind
  causalAssumptionSetRef
  availableDataRegimeSetRef: AvailableCausalDataRegimeSetRef
  realizedCounterfactualDataRefs?: RealizedCounterfactualDataRefSet
  counterfactualDataIdentificationMethodRef?: CounterfactualDataIdentificationMethodRef
  counterfactualDataBoundRef?: CounterfactualDataBoundRef
  causalBoundOrPartialIdentificationRef?
  falsificationOrNegativeControlRef?
  sensitivityAnalysisRef?
  rivalCauseStressTestRef?
  verdict: identified | nonidentified | bounded | unknown
  supportedUse
  unsupportedUse

Identification is inferential support. It is not direct physical sampling.

Realized counterfactual data may change an identification route, tighten a bound, or change which assumptions are still needed. When it does, the profile names the data refs, identification method, and bound ref that changed the result. It does not erase the distinction between identification and direct sampling; the profile must still state what is identified, bounded, unknown, or not identified.

Counterfactual Sampling Realizability Profile

CounterfactualSamplingRealizabilityProfile answers whether samples from a counterfactual-comparison target distribution can be physically obtained through admissible actions under physical, ethical, operational, unit-history, and graph constraints.

CounterfactualSamplingRealizabilityProfile:
  causalUseQuestionRef: U.CausalUseQuestion
  targetCounterfactualDistributionRef
  targetCausalityLadderRung: counterfactualComparisonRung
  structuralCausalModelRef?: StructuralCausalModelRef
  causalDiagramRef?: CausalDiagramRef
  causalGraphRepresentationKind?: CausalGraphRepresentationKind
  graphSeparationCriterionKind?: GraphSeparationCriterionKind
  causalInferenceCalculusKind?: CausalInferenceCalculusKind
  graphChildInterventionConstraintRef?
  sameUnitConflictCheck
  ancestorRegimeConflictCheck
  physicalConstraintSetRef
  ethicalConstraintSetRef
  operationalConstraintSetRef
  unitHistoryAvailabilityRef?
  counterfactualSamplingActionSetRef
  counterfactualRandomizationCapabilityRef?
  counterfactualSamplingWorkPlanRef?
  verdict: realizable | nonrealizable | bounded | unknown
  supportedUse
  unsupportedUse

Realizability is operational. It asks what work can be done, by which system, with which action primitives, under which constraints.

Applied Causal-Inference Profiles

Target-trial and potential-outcomes claims use TargetTrialProtocolRecord and U.PotentialOutcomeContrast when the causal-use claim is an applied intervention-effect claim.

TargetTrialProtocolRecord:
  causalUseQuestionRef: U.CausalUseQuestion
  targetPopulationRef?
  eligibilityCriteriaRef?
  treatmentStrategySetRef
  treatmentAssignmentProcedureRef?
  timeZeroAlignmentRef?
  causalFollowUpWindowRef
  outcomeMeasureRef
  potentialOutcomeContrastRef?
  estimandRef: U.CausalEstimand
  causalAnalysisPlanRef?

Target-trial emulation from observational data adds a mapping/reporting record. TargetTrialEmulationMappingRecord owns the fit between the protocol and the observed data; TargetTrialProtocolRecord alone does not state emulation adequacy.

TargetTrialEmulationMappingRecord:
  targetTrialProtocolRef: TargetTrialProtocolRecord
  observationalDataSourceRef: ObservationalDataSourceRef
  eligibilityMappingRef: TargetTrialEligibilityMappingRef
  treatmentStrategyMappingRef: TargetTrialTreatmentStrategyMappingRef
  assignmentOrTimeZeroMappingRef: TargetTrialAssignmentOrTimeZeroMappingRef
  followUpMappingRef: TargetTrialFollowUpMappingRef
  outcomeMappingRef: TargetTrialOutcomeMappingRef
  emulationGapRef?: TargetTrialEmulationGapRef
  residualConfoundingAssessmentRef?: ResidualConfoundingAssessmentRef
  sensitivityOrAdditionalAnalysisRef?: TargetTrialSensitivityOrAdditionalAnalysisRef
  supportedEmulationUse: CausalUseSupportStatement
  unsupportedEmulationUse: CausalUseUnsupportedStatement

Numerical causal estimates use CausalParameterEstimationProfile when estimation validity is live:

CausalParameterEstimationProfile:
  estimandRef: U.CausalEstimand
  causalIdentificationProfileRef?
  estimatorRef
  nuisanceModelSetRef?
  orthogonalScoreRef?
  crossFittingPlanRef?
  positivityOrOverlapCheckRef?
  sensitivityAnalysisRef?
  uncertaintyIntervalRef?
  supportedEstimateUse
  unsupportedEstimateUse

Transported support uses CausalTransportabilityProfile:

CausalTransportabilityProfile:
  causalUseQuestionRef: U.CausalUseQuestion
  sourcePopulationRef
  targetPopulationRef
  sourceContextRef?
  targetContextRef?
  selectionDiagramRef?
  domainShiftAssumptionSetRef?
  transportFormulaOrBridgeRef?
  supportedTransportUse
  unsupportedTransportUse

Off-policy causal evaluation uses OffPolicyCausalEvaluationProfile when a policy is evaluated from data generated by another behavior or logging policy:

OffPolicyCausalEvaluationProfile:
  evaluationPolicyRef
  behaviorPolicyRef
  causalUseQuestionRef: U.CausalUseQuestion
  sequentialHorizonRef?: SequentialPolicyHorizonRef
  adaptivePolicyClassRef?: AdaptivePolicyClassRef
  unitHistoryConditioningRef?: UnitHistoryConditioningRef
  confoundingAssumptionSetRef?
  supportOrOverlapCheckRef?
  policyTransportabilityRef?: CausalPolicyTransportabilityRef
  offPolicyEstimatorRef?
  uncertaintyIntervalRef?
  supportedPolicyUse
  unsupportedPolicyUse

Causal representation learning uses CausalVariableRepresentationRecord when high-level causal variables are learned, selected, abstracted, or represented from lower-level observations rather than given by the domain:

CausalVariableRepresentationRecord:
  causalUseQuestionRef?: U.CausalUseQuestion
  structuralCausalModelRef?: StructuralCausalModelRef
  causalVariableSetRef
  representationSourceRef
  abstractionOrSelectionMethodRef?
  interventionValidityRef?: CausalRepresentationInterventionValidityRef
  mechanismInvarianceRef?: CausalRepresentationMechanismInvarianceRef
  abstractionFidelityRef?: CausalRepresentationAbstractionFidelityRef
  counterfactualQueryPreservationRef?: CausalRepresentationCounterfactualQueryPreservationRef
  representationShiftRef?: CausalRepresentationShiftOrOODRef
  validationRef?
  supportedCausalVariableUse
  unsupportedCausalVariableUse

Causal Graph Representation Names

Use names that causal inference specialists can recognize:

CausalGraphRepresentationKind =
  causalDirectedAcyclicGraphRepresentation |
  acyclicDirectedMixedGraphRepresentation |
  singleWorldInterventionGraphRepresentation |
  structuralCausalModelTwinNetworkRepresentation |
  ancestralMultiWorldNetworkRepresentation |
  counterfactualGraphicalModelRepresentation

When graph separation or graphical calculus is part of the causal-use support, use controlled values rather than open prose:

GraphSeparationCriterionKind =
  dSeparationCriterion |
  mSeparationCriterion |
  singleWorldInterventionGraphSeparationCriterion |
  ancestralMultiWorldNetworkSeparationCriterion |
  counterfactualGraphSeparationCriterion

CausalInferenceCalculusKind =
  doCalculus |
  ctfCalculus |
  potentialOutcomeCalculus |
  gFormulaCalculus

CausalGraphRepresentationKind, GraphSeparationCriterionKind, and CausalInferenceCalculusKind are formal-support classification values, not minted model objects. They classify the formal support form being used for causal support. Concrete ...Ref fields point to actual models, diagrams, proof objects, assumptions, or epistemes and must be present when that formal support form is load-bearing. For example, StructuralCausalModelRef cites a concrete SCM object, while structuralCausalModelTwinNetworkRepresentation classifies a representation form.

StructuralCausalModel is the causal model kind with endogenous variables, exogenous variables, structural assignments, and intervention semantics. structuralCausalModelTwinNetworkRepresentation means the SCM twin-network representation used in counterfactual reasoning with shared exogenous variables. It is not a deep-learning twin network.

Acronyms such as SCM, DAG, ADMG, SWIG, and AMWN may appear as source/plain labels and bridge notes. FPF Tech values expand the source name when expansion reduces alias risk.

Causal Use Evidence Design

Use CausalUseEvidenceDesignRecord when the causal-use claim needs evidence planning, evidence graph support, experiment or quasi-experiment design, counterfactual randomization, mixed-design accountability, or simulation validation.

CausalUseEvidenceDesignRecord:
  causalUseQuestionRef: U.CausalUseQuestion
  targetCausalityLadderRung: CausalityLadderRung
  estimandRef?
  causalInterventionSpecRef?
  targetTrialProtocolRef?
  potentialOutcomeContrastRef?
  causalIdentificationProfileRef?
  causalParameterEstimationProfileRef?
  counterfactualSamplingRealizabilityProfileRef?
  causalTransportabilityProfileRef?
  causalVariableRepresentationRef?
  causalEvidenceSupportBasis: CausalEvidenceSupportBasis
  causalEvidenceWorkRefs?
  causalEvidenceRoleRefs?
  causalEvidenceMethodRef?
  causalEvidenceWorkPlanRef?
  structuralCausalModelRef?
  causalDiagramRef?
  causalGraphRepresentationKind?: CausalGraphRepresentationKind
  graphSeparationCriterionKind?: GraphSeparationCriterionKind
  causalInferenceCalculusKind?: CausalInferenceCalculusKind
  counterfactualGraphicalModelClassRef?
  causalAssumptionSetRef
  counterfactualModelAssumptionSetRef?
  simulationValidationRef?
  falsificationOrNegativeControlRef?
  sensitivityAnalysisRef?
  rivalCauseStressTestRef?
  decisionThresholdAffected?: yes | no | unclear
  causalEvidenceDecisionImpactRef?: CausalEvidenceDecisionImpactRef
  evidenceValueOrProbeWorthinessRef?: EvidenceValueOrProbeWorthinessRef
  causalEvidenceCostRiskRef?: CausalEvidenceCostRiskRef
  supportedUse
  unsupportedUse

This record does not replace A.10 or B.3. It gives them causal-use structure.

Stronger causal evidence is worth planning only when it can change a choice, deployment decision, fairness posture, assurance consequence, or benchmark conclusion enough to justify its cost, risk, and delay. If stronger support would not change the next action, keep the weaker supported use explicit and stop.

Verdicts

CausalUseSupportVerdict is the action grammar:

  • supported means proceed only under the named supported use.
  • bounded means proceed only inside the named limit and record causalBoundedUseReason.
  • unsupported means downgrade the claim or remove causal use.
  • abstain means no causal-use conclusion and records causalAbstainReason.

No verdict is allowed to silently strengthen the claim beyond its evidence support basis.

Causal Action Policy Class

Use CausalActionPolicyClass when a decision, exploration policy, call plan, or agentic strategy depends on causal rung:

CausalActionPolicyClass =
  naturalBehaviorPolicy |
  interventionalPolicy |
  counterfactualPolicy
  • naturalBehaviorPolicy follows observed or natural behavior.
  • interventionalPolicy chooses an action or do(x).
  • counterfactualPolicy acts conditioned on natural action, unit history, or counterfactual response.

This distinction matters for C.11, C.19, and C.24; it does not make those patterns the authority for causal evidence, identification, or realizability.

CausalActionPolicyClass is a classification value for policy-use posture: natural behavior, interventional action, counterfactual policy, mixed policy, or unknown policy. It is not the policy object, not U.Policy, not C.19 pool policy, and not the executable policy used by an agent.

Local U.* Docking

U.CausalUseQuestion names the question whose answer would make a causal use admissible: association use, intervention-effect use, counterfactual-comparison use, causal fairness use, causal policy use, causal evidence support use, causal assurance use, or causal parity use. It governs the question-to-use relation, not the evidence path, estimator, policy object, graph object, or local neighbor pattern.

U.CausalEstimand names the target quantity, contrast, distribution, or functional answer shape for a U.CausalUseQuestion. It binds the question to what would have to be estimated, identified, sampled, bounded, or emulated. It is not the estimator, not the observed metric, not the graph, not the policy object, and not the support verdict.

The card/profile family docks to those heads this way: triage decides whether a U.CausalUseQuestion is live; local and durable cards stabilize the question, U.CausalEstimand, and supported/unsupported use boundary; profiles and specialized records state what support basis, formal support form, operational work, assumptions, and admissible use the question-estimand pair can carry.

Local name cards:

NameKindPlain senseMust not mean
U.CausalUseQuestionquestion object for causal-use admissibilitythe question-to-use object stabilized by triage/local/durable cards before a claim is used causallya whole research project, evidence path, graph, estimator, policy object, or neighboring-pattern reroute
U.CausalEstimandtarget causal quantity, contrast, distribution, or functionalthe answer-shape object linked to a U.CausalUseQuestion before estimation, identification, sampling, bounding, or emulation is judgedestimator, metric reading, support verdict, policy object, or causal graph

Lexical tripwires:

PhraseUse instead when load-bearing
"causal evidence"name CausalEvidenceSupportBasis, A.10 evidence path refs, CausalUseSupportRecordRef, and CausalUseSupportStatement / CausalUseUnsupportedStatement
"counterfactual data"distinguish realized counterfactual data refs, realizedCounterfactualSampleSupportBasis, identifiedCounterfactualEstimateSupportBasis, and simulationOnlyCounterfactualOutputBasis
"policy optimality"name causalPolicyClaim, CausalActionPolicyClass, OffPolicyCausalEvaluationProfile, CausalUseSupportStatement, and unsupported unqualified optimality
"fairness evidence"distinguish metric/evaluation fairness from causalFairnessClaim with rung, estimand, support basis, support record/verdict, and supported/unsupported fairness use
"method improves"name whether the claim is association, intervention effect, counterfactual comparison, or parity result, then name rung, support basis, and supported/unsupported use
"what would have happened"name counterfactual comparison support, realized counterfactual sample support, identified estimate support, or simulation-only bounded model use

Neighbor Routing Table

If the live issue is...Use...C.28 role
measured value, score, scale, indicator, or metric definitionC.16Only active when the measure is used causally.
temporal trend, rate, acceleration, inertia, or rhythm wordingC.27Active when temporal wording is used as causal effect or intervention evidence.
evidence graph reference or provenanceA.10Carries evidence/provenance path and C.28 support-basis refs, not causal-use support authority.
assurance level, degrade, abstain, or trust postureB.3Consumes C.28 support verdicts and applies assurance consequences.
local decision among optionsC.11Provides causal action-policy hooks when value/regret/optimality depends on causal rung.
exploration/exploitation over live poolsC.19Provides causal data-collection or causal policy-learning hooks when live.
tool/call/enactment planC.24Provides optional causal action use spec when the call selects observation, intervention, counterfactual-rung evidence collection, or counterfactual policy conditioning.
bias and fairness auditD.5Provides causal fairness rung and supported fairness use.
method dispatch or selector-facing registryG.5Provides causal method/policy class declarations when causal methods are compared.
benchmark or method parityG.9Provides causal method rung parity.
quantum-like modeling cueC.26Receives only the residual QL cue after causal-use explanation has been tried.

Non-Goals

C.28 does not:

  • define physical causation or decide what causation is in the modeled world;
  • choose one causal school, such as SCM/PCH, potential outcomes, target-trial emulation, transportability, causal ML, causal RL, or causal fairness, for all FPF use;
  • certify a DAG, SCM, SWIG, AMWN, or other graph as true or sufficient causal support by naming it;
  • replace local domain science, domain intervention definitions, outcome definitions, or substantive rival-cause knowledge;
  • replace C.16 measurement and metrics characterization, including metric construction, calibration, and non-causal score interpretation;
  • replace A.10 evidence graph referring, provenance paths, evidence-role carriers, or evidence graph path discipline;
  • replace B.3 trust and assurance calculus, assurance tuples, F-G-R/CL consequences, or assurance publication posture;
  • replace D.5 bias audit and ethical assurance, causal-fairness audit responsibility, or human/group-impact review;
  • replace G.9 parity benchmark harness, causal-rung parity screen, or benchmark report structure;
  • replace C.11 choice, C.19 exploration/exploitation policy, or C.24 call-planning patterns; it only supplies causal-use support boundaries consumed by those patterns.

Cheap Downgrade Library

Use a downgrade sentence when a weaker admissible use is enough:

Each sentence below is an admissible cheapStop wording. It closes the causal-use question for the named weak case unless the author keeps a stronger publish, choose, deploy, assure, audit, benchmark, or support-treatment use alive.

CaseAdmissible downgrade wording
association-only case"Observed association only; supported use = association report; unsupported use = intervention-effect claim."
temporal-change-only case"Temporal change or trend is recorded; supported use = temporal/rate description; unsupported use = causal-effect claim until a causal-use support basis is named."
simulation-only case"Simulation-only counterfactual output; supported use = bounded model-supported exploration or explanation; unsupported use = realized counterfactual sample evidence or intervention-effect claim."
metric-only fairness case"Metric disparity or metric improvement is recorded; supported use = metric-level fairness or disparity report; unsupported use = causal fairness claim without a causal rung, estimand, support basis, and supported fairness use."
logged-policy bounded case"Logged-policy evidence supports only the declared behavior/evaluation regime; supported use = bounded off-policy evaluation under named overlap/transportability limits; unsupported use = unqualified optimal-policy claim."
cross-rung benchmark case"Methods answer different causal rungs or support bases; supported use = publish bridge/loss, degraded parity, or abstain; unsupported use = one scalar causal winner."

CausalUseBureaucracySniffTest

CausalUseBureaucracySniffTest keeps a causal-use carrier only when it changes the admissible next action or blocks a concrete overclaim. Keep the causal-use record only when at least one answer is "yes":

QuestionIf no
Did the record change the next action?Remove fields until only the action-changing line remains.
Did it block a concrete causal overclaim by naming the stronger use as unsupported?Use association, trend, simulation-only, or metric-only wording and stop.
Did it support one concrete decision, evidence-work, fairness, assurance, benchmark-parity, or deployment move by changing supportedUse or unsupportedUse?Keep the neighboring pattern and do not open a durable causal-use object.
Was there a cheaper nextMove.cheapStop that preserved the same admissible use boundary?Use the cheaper stop.
Is the problem only the word "causal" or "counterfactual", rather than an admissible causal use?Repair wording locally or reroute to the neighboring language/authoring pattern.

Authored-Unit Repair Relation

When the problem is only local authored wording pressure, local head repair, relational precision restoration, explanation faithfulness, or conservative retextualization may come first. C.28 opens at CausalUseActivation, when the wording makes publication, choice, deployment, assurance, audit, fairness, policy, or benchmark use depend on causal support.

Causal-Laundering Golden Cases

CaseExpected C.28 output
Association laundering: "users who received X improved, so X works."rung = observationalAssociationRung; support basis = observationalAssociationSupportBasis; supported use = association report; unsupported use = intervention-effect claim.
Intervention overclaim: "we changed X once, so the policy will work everywhere."rung = interventionalActionRung; support basis = interventionalActionSupportBasis inside assignment, context, follow-up, and transportability limits; unsupported use = cross-population or unbounded policy claim.
Simulation laundering: "the simulator shows what would have happened."claim kind = relevant existing CausalUseClaimKind; support basis = simulationOnlyCounterfactualOutputBasis; supported use = bounded model-supported use; unsupported use = realized counterfactual sample or intervention-effect evidence.
Metric-only fairness laundering: "fairness improved because the metric improved."supported use = metric-level fairness/disparity report; unsupported use = causal fairness claim unless causalFairnessClaim, rung, estimand, support basis, and supported fairness use are declared.
Policy replay overclaim: "logged replay says this policy is optimal."claim kind = causalPolicyClaim; support basis = off-policy causal evaluation with behavior/evaluation policy refs and overlap/support checks; supported use = bounded policy evaluation; unsupported use = unqualified optimality.
Cross-rung benchmark: "method A beats method B as a causal method."claim kind = causalBenchmarkParityClaim; use G.9 CausalRungParityScreen; supported use = within-rung parity or declared bridge/loss; unsupported use = one scalar causal winner when rungs/support bases differ.
Temporal-cause wording: "after launch, recovery got faster, so launch caused resilience."supported use = C.27 temporal/rate adequacy; unsupported use = causal-effect claim until C.28 names intervention timing, outcome window, assumptions, rival causes, and support basis.
QL escape: "ordinary probability is hard here, so the effect is quantum-like."supported use = causal-use triage and ordinary-neighbor explanation first; unsupported use = bypassing C.28 with quantum-like vocabulary; C.26 is retained only for residual quantum-like probe/frame/order/export/coarsening issue.

| Target-trial name-drop: "we emulate a trial, so the effect is identified." | supported use = target-trial claim only with protocol plus emulation mapping, data source, assignment/time-zero, follow-up/outcome mapping, residual confounding, and sensitivity/additional analysis; unsupported use = identification claim by target-trial label alone. | | Realized-counterfactual-data claim: "we observed both outcomes for the same unit." | supported use = samples from the declared target counterfactual distribution under the realizability profile's constraints; unsupported use = same-world incompatible-outcome wording for one unit. |

Archetypal Grounding

Tell. A causal-use claim is a promise about what a reader may do with a result. The claim is safe only when the rung, contrast, support basis, and allowed use are named.

Show (System). A product team observes that users who received an intervention had better outcomes. C.28 first records an observational association unless the team can name an intervention/action design, target trial protocol, identification profile, or evidence design that supports intervention-effect use. If the team only has observational association, the next move is to publish association or build evidence, not to claim causal improvement.

Show (Episteme). A fairness report says one model is fair because a metric improved after a policy change. C.28 asks whether the fairness claim is associative, interventional, or counterfactual. If it is interventional-action-rung only, it cannot be published as counterfactual fairness without identification or realizability support.

Show (Policy). A team wants to deploy a causal policy learned from logged behavior data. C.28 records causalPolicyClaim, interventionalActionRung or counterfactualComparisonRung as appropriate, CausalActionPolicyClass, OffPolicyCausalEvaluationProfile, support/overlap checks, uncertainty, supported policy use, and unsupported policy use. If the behavior policy cannot support the target policy, the admissible output is bounded use or abstain rather than "the policy is optimal".

Show (Causal RL). An online learner uses behavior-policy logs and counterfactual data-fusion to choose a treatment, ranking, or action policy. C.28 records the natural behavior policy, evaluation policy, CausalActionPolicyClass, target rung, confounding/support assumptions, OffPolicyCausalEvaluationProfile, uncertainty, supported policy use, and unsupported policy use. The learner may publish bounded causal policy support only for the declared regime; it must not turn replay reward, exploration success, or counterfactual strategy output into an unqualified optimal-action claim.

Show (Evidence Work). A lab can physically run a counterfactual-rung sampling procedure by assigning compatible action regimes to matched units under ethical and operational constraints. C.28 separates CounterfactualSamplingRealizabilityProfile from CausalIdentificationProfile: the realized sampling work becomes U.Work with evidence carriers and guards, while identification remains the inferential route from assumptions, graph, calculus, and available data.

Show (Simulation-Only). A simulator produces "what would have happened" traces for a rollout decision. C.28 can allow useful model-supported use without calling the traces realized counterfactual-rung evidence: the record uses simulationOnlyCounterfactualOutputBasis, names counterfactualModelAssumptionSetRef, simulationValidationRef, supported simulation use, and unsupported use. The output may support rehearsal, sensitivity exploration, or model-based explanation inside declared limits; it does not support direct counterfactual sample wording or intervention-effect publication by vocabulary alone.

Show (Benchmark). A benchmark compares one observational predictor, one intervention optimizer, and one counterfactual strategy. C.28 does not ban the comparison, but it requires CausalMethodRungParityRecord through G.9: if rung, estimandRef, intervention/action basis, support basis, consumed C.28 support record/verdict, transportability, follow-up window, and estimation-validity basis are not comparable, the benchmark publishes bridge/loss, degraded use, or abstain instead of one superiority claim.

Bias-Annotation

C.28 is mainly a causal-discipline and anti-overclaim pattern for decision-bearing causal use. One part of that work is catching language laundering, but the larger job is to keep causal reasoning, evidence design, realizability, policy evaluation, fairness use, and benchmark parity from silently borrowing stronger support than they actually have.

Common biases:

  • Causal prestige bias. A result sounds more important when phrased causally, so weak evidence gets overused.
  • Simulation laundering. A simulated counterfactual is treated as observed or realized counterfactual-rung evidence.
  • Metric proxy bias. A fairness or performance proxy is treated as a causal result without rung and estimand.
  • Benchmark scalarization bias. A method comparison collapses different causal rungs, estimands, or transport assumptions into one score.
  • Graph sufficiency bias. A named graph is treated as enough without assumptions, data regime, calculus, and supported use.

The repair is not to ban causal language. The repair is to recover the live causal question, choose the weakest honest supported use, and then either downgrade, bound, design stronger evidence, open identification or realizability work, or abstain.

Conformance Checklist

CheckRequirement
CC-C28-0 Triage-only useFor triage-only use, causality-ladder rung is named or causal use is declined, supported/unsupported use is named, and a stronger causal-use claim is not implied.
CC-C28-1 Causality-ladder rung declarationEvery causal-use claim declares its target causality-ladder rung: observational association question, interventional action/effect question, or counterfactual comparison question.
CC-C28-2 Durable causal estimand disciplineEvery durable interventional/counterfactual-rung causal-use claim names causal-use question, comparator or counterfactual, estimand, assignment or intervention window, follow-up window, outcome measure, assumptions, rival causes, and supported/unsupported use.
CC-C28-3 No unsupported causality-ladder climbA claim at interventional-action or counterfactual-comparison rung is not supported only by weaker causality-ladder data unless CausalIdentificationProfile, CounterfactualSamplingRealizabilityProfile, or bounded-use treatment is cited.
CC-C28-4 Realizability is not identificationCausalIdentificationProfile and CounterfactualSamplingRealizabilityProfile remain distinct. One supports inference from other data; the other supports direct sampling through feasible physical actions.
CC-C28-5 Counterfactual data collection is workAny realized counterfactual-rung-data procedure is represented as U.Work enacted by U.System under RoleAssignment, with MethodDescription, WorkPlan, evidence carriers, and physical/ethical/operational guards.
CC-C28-6 Verdicts are action grammarsupported, bounded, unsupported, and abstain each change what the reader may do next.
CC-C28-7 No durable-card defaultEscalate from triage to local card to durable card and profiles only when the claimed use triggers the stronger object.
CC-C28-8 Heavy causal-use object payoffEvery selected heavy field or check changes a reader action, blocks a specific overclaim, or supports a concrete evidence/assurance/fairness/parity decision.
CC-C28-9 Semantic-authority splitC.28 owns causal-use value sets, identification/realizability profiles, graph/calculus naming, and support verdicts; neighbors may consume or quote them but must not define competing causal-use value sets.
CC-C28-10 Simulation-only bounded useSimulation-only output may support bounded model-supported use, but it never becomes interventional evidence or realized counterfactual sample evidence by vocabulary, validation, or role relabeling alone.
CC-C28-11 Decision-economics of evidenceA stronger causal evidence plan names the decision threshold, evidence value or probe-worthiness, and cost/risk posture when escalation is not already mandatory by safety, release, or assurance constraints.

Common Anti-Patterns and How to Avoid Them

Anti-patternSymptomRepair
Fill-all-cards defaultEvery mention of "cause", "effect", or "counterfactual" triggers a durable dossier.Start with CausalUseTriageRecord; escalate only when the claimed use requires it.
Causal certification theaterEvery field is filled, but no reader action, evidence design, downgrade, or unsupported use changes.Remove or weaken fields until each one changes a decision or blocks an overclaim.
Association as intervention"Users who received intervention X did better" is published as effect of X without action/assignment support.Publish association or build identification/evidence design.
Interventional proxy as counterfactual fairnessA policy-change metric is called counterfactual fairness.Declare interventional-action rung unless counterfactual estimand plus identification or realizability is present.
Simulation as realized counterfactual sampleModel output is described as realized counterfactual-rung support without direct sampling or validation.Use simulationOnlyCounterfactualOutputBasis and name supported/unsupported model use.
Graph-only causalityA DAG or SCM diagram is treated as sufficient support.Add assumptions, data regime, graph representation kind, calculus, and supported use.
Cross-rung benchmarkMethods are compared as peers while one answers association, another intervention, and another counterfactual comparison.Use CausalMethodRungParityRecord and degrade or abstain when parity is absent.
QL escapeCausal confusion is rebranded as quantum-like because ordinary probability feels hard.Use C.26 only after causal-use explanation and ordinary FPF neighbors have done their work.

Consequences

C.28 makes causal use slower only when the claim is strong enough to deserve it. Cheap causal triage remains cheap.

Positive consequences:

  • Causal claims become inspectable by rung, support basis, and supported use.
  • Counterfactual sampling realizability becomes operational rather than merely philosophical.
  • Identification and realizability no longer collapse.
  • Fairness, policy, and benchmark claims stop borrowing stronger causal force than their evidence supports.
  • Neighbor patterns receive narrow causal hooks without becoming general causal authorities.

Costs:

  • Authors must learn a small causal vocabulary.
  • Some attractive claims will be downgraded to association, bounded use, simulation-only, or abstain.
  • Higher-rung claims need more evidence, assumptions, or work-plan detail.

The cost is intended. It is cheaper than publishing an unsupported causal use.

Rationale

FPF needs this pattern because causal language changes what a reader may do.

Temporal language can say that something changed. Measurement language can say that a score is higher. Assurance language can say that evidence is stronger or weaker. None of those alone says that an action caused a result, that a counterfactual comparison is supported, or that a causal policy should be deployed.

C.28 therefore uses a semantic-authority split:

  • C.28 owns causal-use question, rung, estimand, identification, realizability, causal evidence support basis, and causal-use verdict.
  • Neighbor patterns keep their own authority and cite C.28 only when causal use is live.
  • C.26 receives a causal exit: intervention, causal effect, causal fairness, causal policy, and counterfactual-rung-data realizability are ordinary causal-use questions before they are quantum-like modeling questions.

The pattern is not Pearl-only. SCM/PCH provides the rung discipline, but potential outcomes, target-trial emulation, causal ML estimation, transportability, causal representation learning, causal RL, and causal fairness all change the fields that FPF must preserve.

SoTA-Echoing

SoTA claimPractice implicationSource anchorsFPF adoption
Causal reasoning separates seeing, doing, and imagining.A claim must declare CausalityLadderRung before support is judged.Pearl/SCM/PCH: On Pearl's Hierarchy and the Foundations of Causal Inference.Adopted as observationalAssociationRung, interventionalActionRung, counterfactualComparisonRung.
Lower-rung data generally underdetermines higher-rung questions.No unsupported causality-ladder climb.Pearl causal hierarchy and identification tradition: On Pearl's Hierarchy and the Foundations of Causal Inference.Adopted as CC-C28-3.
Counterfactual sampling realizability is operational and partial.Some counterfactual-rung distributions can be directly sampled; some cannot; some are bounded.Raghavan and Bareinboim: Counterfactual Sampling Realizability, technical report.Adopted as CounterfactualSamplingRealizabilityProfile.
Counterfactual randomization is action/work over an SCM.Realized counterfactual-rung data collection needs U.Work, action primitives, graph constraints, and guards.Forney, Bareinboim, Pearl: Counterfactual Randomization.Adopted as CC-C28-5 and evidence-design fields.
Counterfactual data can change what is identifiable or bounded.Identification profiles must make realized counterfactual data, identification methods, and bound changes explicit rather than treating all data regimes as one scalar source.Raghavan and Bareinboim: Counterfactual Sampling Realizability; Forney, Bareinboim, Pearl: Counterfactual Randomization.Adopted as availableDataRegimeSetRef, realizedCounterfactualDataRefs, counterfactualDataIdentificationMethodRef, and counterfactualDataBoundRef.
Counterfactual graphical models require named graph forms and calculus.Graph form, separation criterion, and calculus must be visible for counterfactual support.Yang and Bareinboim: A Hierarchy of Graphical Models for Counterfactual Inferences; Correa and Bareinboim: Counterfactual Graphical Models.Adopted as CausalGraphRepresentationKind, GraphSeparationCriterionKind, and CausalInferenceCalculusKind; doCalculus and ctfCalculus are controlled calculus values, not free-form hooks.
Potential outcomes and target-trial emulation operationalize intervention-effect claims.Applied intervention claims need target population, eligibility, treatment strategies, assignment/time-zero, follow-up, outcome, contrast, estimand, and analysis plan.Rubin: Estimating Causal Effects of Treatments; Hernan/Wang/Leaf: Target Trial Emulation.Adopted as U.PotentialOutcomeContrast and TargetTrialProtocolRecord.
Target-trial emulation from observational data needs mapping and reporting, not only protocol naming.Eligibility, strategies, assignment/time-zero, follow-up, outcomes, residual confounding, and sensitivity/additional analyses must be mapped from observational data to the target trial.Hernan/Wang/Leaf: Target Trial Emulation.Adopted as TargetTrialEmulationMappingRecord.
Causal ML estimation is not the same as identification or prediction.Estimator, nuisance models, orthogonal score, cross-fitting, overlap/positivity, sensitivity, and uncertainty must be visible when estimation validity is claimed.Chernozhukov et al.: Double/debiased machine learning for treatment and structural parameters.Adopted as CausalParameterEstimationProfile.
Causal support may not transport across populations or domains without assumptions.Source and target populations/contexts, selection diagrams, domain-shift assumptions, and transport formula/bridge must be named.Pearl and Bareinboim: Transportability of Causal and Statistical Relations.Adopted as CausalTransportabilityProfile.
AI causal work often cannot assume causal variables are already given.Learned or selected causal variables need a representation record.Scholkopf et al.: Toward Causal Representation Learning.Adopted as CausalVariableRepresentationRecord.
Causal representation support depends on intervention validity, invariance, abstraction fidelity, query preservation, and shift handling.A learned representation must not silently become a causal variable for every query or domain.Scholkopf et al.: Toward Causal Representation Learning.Adopted as causal representation validation hooks in CausalVariableRepresentationRecord.
Sequential causal games and causal RL make counterfactuality policy-relevant.Natural behavior, interventional, and counterfactual policies, sequential horizons, adaptive policies, unit-history conditioning, and transportability must be distinguished.Maiti and Bareinboim: Sequential Causal Games; Bareinboim/Forney/Pearl: Bandits with Unobserved Confounders; Forney/Pearl/Bareinboim: Counterfactual Data-Fusion for Online Reinforcement Learners.Adopted as CausalActionPolicyClass and OffPolicyCausalEvaluationProfile hooks.
Causal fairness is not only metric choice.Fairness claims must declare causal rung, path/estimand where live, and supported fairness use.Plecko and Bareinboim: Fairness-Accuracy Trade-Offs: A Causal Perspective.Adopted through D.5 relation and CausalFairnessUseAuditCard.

Relations

  • C.16 governs measurement and metrics. C.28 activates only when a measurement is used causally.
  • C.27 governs temporal claim adequacy. C.28 activates when temporal change is used as causal effect, intervention evidence, or counterfactual comparison.
  • A.10 governs evidence graph referring. C.28 supplies causal evidence support basis and causal-use support refs for evidence paths.
  • A.2.4 governs evidence roles. C.28 requires the evidence-role distinctions that keep simulationOnlyCounterfactualOutputBasis, identifiedCounterfactualEstimateSupportBasis, interventional evidence, and realizedCounterfactualSampleSupportBasis from being confused.
  • A.6 / A.6.B / A.6.C govern boundary, deontic, promise, commitment, utterance, contract-language, and routed claim language. C.28 supplies only causal-use support when mixed boundary sentences claim causal effect or counterfactual support.
  • A.15 governs role, method, plan, and work alignment. C.28 supplies the causal-use semantics for intervention assignment, target-trial emulation, counterfactual sampling work, and causal evidence collection.
  • B.3 governs trust and assurance. C.28 supplies the causal-use verdict that B.3 can degrade, bound, or abstain over.
  • C.11 governs decision theory. C.28 supplies causal-use question and causal action-policy class when value, utility, regret, or optimality depends on causal rung.
  • C.19 governs explore/exploit pool policy. C.28 supplies causal rung and policy/regime fields when exploration collects causal data or learns causal policy.
  • C.24 governs agentic tool use and call planning. C.28 supplies causalActionUseSpec when calls select observation, intervention, counterfactual-rung evidence collection, or counterfactual policy conditioning.
  • D.5 governs bias audit and ethical assurance. C.28 supplies causal fairness rung, estimand, support, and supported fairness use.
  • G.5 governs method dispatch and MethodFamily registry. C.28 supplies causal method or policy class declarations when method dispatch compares causal methods.
  • G.9 governs parity and benchmarks. C.28 supplies causal method rung parity.
  • G.11 governs refresh orchestration. C.28 supplies causal-use support records whose realizability, identification, fairness, representation, off-policy, target-trial, and simulation-validation shifts can trigger refresh.
  • C.26 governs quantum-like modeling. C.28 is a required causal exit before QL retention when the live question is intervention, causal effect, causal fairness, causal policy, counterfactual comparison, or counterfactual-rung-data realizability.

C.28:End