Responsibility Attribution¶
Core Idea¶
Responsibility attribution is the process of mapping an observed outcome back onto the agents or factors whose actions and omissions produced it, and apportioning credit or blame among them. Its defining structure is a directed assignment from effects to responsible sources, gated by counterfactual and normative tests: an agent is held responsible to the degree that the outcome would have differed had it acted otherwise (counterfactual dependence), that the act was within its control, and that it could foresee the consequence. [1] The pattern is the assignment operation itself, distinct from any particular bias in how it is performed and distinct from the causal facts it consumes as input. Where bare causation answers "what produced this?", attribution answers the further question "who or what is to be held to account, and in what proportion?" — converting a sprawling causal history into a bounded ledger of credit and blame. [2]
The operation has a striking property that recurs wherever it appears: it deliberately stops short of the full causal chain. A complete causal account regresses indefinitely ("the spark caused the fire, but the dry brush caused the spark's effect, but the drought caused the dryness…"), whereas attribution halts at agents or factors that could have acted otherwise and that were positioned to bear sanction, reward, or repair. This selective truncation is not a defect of the operation but its central function: it is what makes responsibility actionable. [2]
How would you explain it like I'm…
Who Did It?
Pinning Down Who's Accountable
Assigning Responsibility
Structural Signature¶
Responsibility attribution encodes a structural pattern: observed outcome → causal back-tracing → counterfactual/control/foresight gating → apportioned assignment over accountable sources. It separates two things that folk reasoning fuses — the fact of causal contribution and the judgment of accountability — and names the gated mapping that runs from the first to the second. [2]
Recurring features:
- Directed assignment from outcomes back to responsible sources
- Counterfactual test: would the outcome have differed had the agent acted otherwise?
- Control and foreseeability gates on causal contribution
- Apportionment of credit or blame across multiple contributors
- Truncation of the causal regress at agents who could have acted otherwise
- Separation of "what caused it" from "who is answerable for it"
- Weighted allocation over a small set of accountable agents
The structural insight is robust: a juror apportioning liability across multiple negligent parties, a reinforcement-learning agent distributing reward for a delayed outcome across the actions that produced it, an incident-review team tracing a system failure back through contributing factors, and a historian debating whether structural forces or individual choices drove an outcome are all running the same operation. [3] Each takes an effect, traces it back through a causal field, and projects a weighted assignment onto a bounded set of sources, with the gating tests determining how much weight each source receives.
What It Is Not¶
Responsibility attribution is not causal explanation. Establishing that X caused Y is a factual claim about the world; attributing responsibility for Y to an agent adds normative gates (control, foreseeability) that causation alone does not supply. A falling branch can cause an injury without bearing any responsibility for it, because a branch satisfies no control or foreseeability test. [2] The prime consumes causal facts as input but produces a different kind of output: an allocation of accountability, not a description of mechanism.
Nor does the prime claim that responsibility tracks causal contribution proportionally. A minor causal contributor can bear major responsibility (a saboteur who flips one switch), and a major causal contributor can bear little (a load-bearing wall that "causes" a collapse it was overloaded beyond spec). The gating tests can amplify or suppress causal weight, so the mapping from causation to responsibility is deliberately non-linear. The prime names the existence and shape of this mapping, not a particular formula for it.
The prime also does not assert that any particular attribution is correct. It is the operation, not the verdict. Two competent observers running the operation with different counterfactual baselines, different theories of control, or different foreseeability standards can reach opposite assignments from identical facts — and both are doing responsibility attribution. Treating the prime as if it guaranteed a unique answer confuses the structure with one of its outputs.
Finally, responsibility attribution is not the same as a bias in performing it. The well-known tendency to over-weight salient agents and under-weight situational forces is a distortion of the operation, not the operation itself. The prime is the general assignment process that such biases corrupt; naming the process is precisely what lets us see the bias as a deviation rather than as the nature of the thing. [3]
Broad Use¶
Law: Doctrines of causation, negligence, and apportionment assign liability across multiple contributing parties. Tort law's "but-for" causation is an explicit counterfactual test, and comparative-negligence regimes apportion damages by fault percentage across plaintiffs and defendants. [4] Criminal law layers further gates — mens rea encodes foreseeability and intent, and doctrines of duress or incapacity remove control.
Moral psychology: People run counterfactual and intentionality tests to decide who is blameworthy for a harm, and experimental work shows these tests are graded rather than binary — blame scales with perceived control, foreknowledge, and the availability of alternative actions. [3] Folk attribution fuses cause and blame by default, which is exactly why naming the gated structure has clarifying power.
Reinforcement learning (non-obvious): The temporal credit-assignment problem distributes responsibility for a delayed reward across the earlier actions that brought it about — the same structure, formalized. Eligibility traces, advantage estimation, and discounting are mechanisms for back-projecting credit through a chain of actions toward a delayed outcome. [5] This is the prime's most precise computational incarnation: a directed assignment from an effect (reward) to its sources (prior actions), weighted by how much each plausibly contributed.
Organizational behavior: Blameless post-mortems and root-cause analyses attribute an incident to contributing factors rather than scapegoats, deliberately resisting the pull toward the most proximate operator. The discipline's core move is to keep tracing past the person who "pressed the button" to the systemic conditions that made the error likely.
History and social science: Debates over structural versus individual causes are disputes about how to attribute responsibility for large outcomes — whether to assign a war, a famine, or a financial collapse to identifiable decision-makers or to impersonal forces. These are not disagreements about facts so much as about where the attribution operation should halt.
Clarity¶
A core function of "responsibility attribution" is to separate the fact of causal contribution from the judgment of responsibility, which folk reasoning routinely fuses. Once the two are pried apart, a practitioner can coherently say that an agent caused an outcome yet bears reduced responsibility for it (no control, no foresight), or that an agent contributed only trivially to the causal field yet bears the bulk of responsibility (full control, full foresight, clear duty). [1] This is impossible to state cleanly without the prime, because ordinary language slides between "you did it" (causal) and "it's your fault" (normative) as though they were one claim.
The prime also clarifies that responsibility can be shared and apportioned rather than pinned on a single party. The question shifts from "who is to blame?" — which presupposes a unique culprit — to "how is responsibility distributed across the contributors, and what gates set each share?" This reframing is what allows comparative-fault law, multi-party root-cause analysis, and graded moral judgment to operate at all; without it, every outcome demands a single owner.
Manages Complexity¶
Responsibility attribution collapses a tangled causal web into a tractable allocation over a small set of accountable agents, telling us where to direct sanction, reward, or repair. A real outcome has effectively unbounded causal antecedents; the prime's gating tests prune that field to the handful of sources that could have acted otherwise and were positioned to be held to account. [2] The result is a bounded ledger that institutions can actually act on — a defendant to sue, an action to reinforce, a process to fix.
Crucially, it bounds the otherwise infinite regress of "but what caused that?" by stopping at agents who could have acted otherwise. This truncation is what makes the operation usable: rather than tracing a fire back through the spark, the dryness, the drought, and the climate, attribution halts at the camper who left the fire unattended — because that is the node where control and foreseeability were present and where intervention is possible. The prime thus does double duty: it reduces dimensionality (web to short list) and it terminates a regress (chain to actionable node).
Abstract Reasoning¶
Recognizing the pattern makes its components inspectable and contestable. Once attribution is seen as a gated mapping rather than an atomic judgment, each gate becomes a separate lever for reasoning: counterfactual dependence (would it have happened anyway?), control (could the agent have done otherwise?), foreseeability (could the agent have known?), and apportionment weights (how is the residual divided?). [3] Disagreements can then be localized to a specific gate — two parties may agree on the facts and the counterfactual yet differ entirely on foreseeability — which makes the dispute tractable instead of a clash of bare intuitions.
This decomposition also enables principled reasoning about shared and diminished responsibility, and it exposes characteristic distortions. The over-weighting of salient, dispositional agents at the expense of situational forces — the pattern the fundamental attribution error names — is visible as a systematic miscalibration of the control and counterfactual gates. [6] Naming the normal operation is what turns the bias from an inscrutable quirk into a predictable failure mode of a known process.
Knowledge Transfer¶
The reinforcement-learning credit-assignment apparatus — discounting responsibility back through a chain of actions toward a delayed outcome — transfers to organizational root-cause analysis, where proximate operators are often the wrong place to stop. An RL agent does not assign all credit to the final action before the reward; it propagates credit backward along the trajectory. The same logic, applied to incident review, warns against stopping at the operator who made the last move and pushes the analysis toward the upstream actions and conditions that set up the failure. [5]
Conversely, legal apportionment rules transfer to multi-agent AI accountability, where harm emerges from interacting components and no single module "did it." Comparative-fault doctrine — the machinery for splitting liability across multiple contributors by relative fault — offers a ready template for distributing responsibility across the models, data pipelines, and human operators that jointly produce an automated harm. In both directions the transfer is conceptually grounded rather than merely metaphorical: the shared structure is a weighted, gated assignment from an effect onto a bounded set of contributing sources.
Examples¶
Formal/abstract¶
Reinforcement learning (temporal credit assignment): An agent navigates a maze and receives a reward only at the goal, many steps after the moves that actually mattered. The learning problem is to assign credit for that delayed reward across the earlier actions — to discover that the turn taken twenty steps ago, not the final step into the goal, was the decisive choice. Algorithms solve this with discounting and eligibility traces: credit flows backward along the trajectory, weighted by how much each action plausibly contributed to the eventual reward. The proximate action (the last step) gets some credit, but a pivotal earlier action can get more. Mapped back: This is responsibility attribution in its purest formal incarnation. The reward is the outcome; the actions are the candidate sources; discounting is the apportionment weighting; and the refusal to dump all credit on the final action mirrors the refusal, in moral and legal attribution, to stop at the most proximate agent. The counterfactual gate appears as the question every credit-assignment method implicitly asks: how much would the reward have differed had this action been otherwise?
Comparative-fault apportionment: A multi-car collision involves a speeding driver, a driver who ran a yellow light, and a municipality that left a signal malfunctioning. A court does not seek the single cause; it apportions fault — say 50%, 30%, 20% — and divides liability accordingly. Each share is set by running the gates: but-for causation establishes who is in the causal field at all, foreseeability and duty establish who could and should have prevented the harm, and the relative weights divide the residual. A party whose conduct was causally necessary but entirely unforeseeable in its consequence draws a small share or none. Mapped back: The structure is identical to the RL case: an outcome (the collision) is traced back to a bounded set of sources (three parties), gated (causation, foreseeability, duty), and apportioned (fault percentages). What differs is only the substrate — courtroom rather than maze — and the gates' normative content.
Applied/industry¶
Blameless post-mortem (software operations): A production outage is triggered when an on-call engineer runs a deployment script that wipes a database. The naive attribution stops at the engineer: human error, the cause is found. A blameless post-mortem refuses this truncation and keeps tracing: the script had no confirmation prompt, the staging and production environments were indistinguishable in the terminal, the runbook was outdated, and an earlier alert had been silenced. Responsibility is distributed across these contributing factors, and the repairs target the system (add a guardrail, differentiate environments) rather than the person. Mapped back: This is attribution with the control and foreseeability gates applied deliberately against the proximate agent. The engineer was causally proximate but had little real control given a system designed to make exactly this error likely; responsibility flows upstream to the conditions and the decision-makers who shaped them — the same backward propagation the RL example formalizes.
Multi-agent AI accountability: An automated lending system denies a qualified applicant. The harm emerges from interacting components: a training dataset with historical bias, a model that amplified it, a threshold set by a product manager, and a human reviewer who rubber-stamped the output. No single component "decided" to discriminate. Borrowing legal apportionment, an accountability review distributes responsibility across the components and the humans who configured them, weighting each by control and foreseeability — the data team that knew the bias existed bears more than the reviewer who lacked the information to catch it. Mapped back: This applies comparative-fault machinery to a non-human causal field. The structure holds: an outcome is traced to a bounded set of sources, each gated by what it could control and foresee, and the residual responsibility is apportioned rather than pinned on whichever component is easiest to point at.
Structural Tensions¶
T1: The truncation that makes attribution usable also makes it contestable. Halting the causal regress at "agents who could have acted otherwise" is what converts an infinite chain into an actionable verdict, but where exactly the regress halts is a choice, not a fact. Stopping at the on-call engineer and stopping at the executive who underfunded tooling both terminate the same chain at different nodes. The very feature that gives the operation its power — selective truncation — is the feature that lets two honest analysts reach opposite assignments from identical causal facts.
T2: Causal weight and responsibility weight pull in different directions. The gates can amplify a trivial causal contributor into the principal bearer of responsibility (a saboteur) or suppress a dominant causal contributor to near-zero responsibility (an overloaded wall). This non-linearity is essential — it is what encodes intent and control into the verdict — but it means attribution can never be read off a causal diagram. Practitioners who anchor on causal magnitude systematically misattribute, and practitioners who ignore causal magnitude entirely lose contact with what actually produced the outcome.
T3: Apportioning responsibility can dilute it to the point of evaporation. Distributing blame across many contributors is fairer than scapegoating a single party, yet beyond some point shared responsibility becomes felt responsibility by no one — every contributor reasonably regards their slice as too small to own. The same apportionment that protects the proximate operator from unjust blame can leave a harm with no party motivated to repair it. The operation that spreads responsibility correctly can thereby undermine the accountability it was meant to secure.
T4: Lowering the salience of individuals to find systemic causes can erode deterrence. Blameless analysis correctly resists stopping at the proximate human and traces failure to systemic conditions, which improves learning and psychological safety. But a regime that never assigns individual responsibility removes the incentive that makes individuals careful in the first place. The shift from "who erred?" to "what conditions produced the error?" is a genuine advance for learning and a genuine risk for deterrence, and the same case can be read either way depending on which the institution most needs.
T5: The counterfactual gate depends on a baseline that the operation cannot itself supply. "Would the outcome have differed had the agent acted otherwise?" requires choosing what the agent would have done instead, and that comparison case is not given by the facts. Against a baseline of "did nothing," an intervener looks responsible; against a baseline of "did the reasonable thing," the same intervener may look blameless. The gate that does the most work in the operation is the gate whose input is most underdetermined, so the verdict can swing entirely on an unstated and contestable baseline.
T6: Foreseeability is judged after the outcome is known, distorting the very gate it sets. Whether an agent could have foreseen a consequence is assessed by observers who already know the consequence occurred, and hindsight inflates judged foreseeability. The gate meant to limit responsibility to the knowable is applied through a lens that makes outcomes look more knowable than they were, systematically pushing attribution toward harsher assignments. The operation builds in a gate to protect agents from being blamed for the unforeseeable, yet the standard way of applying that gate quietly defeats its purpose.
Structural–Framed Character¶
Responsibility Attribution is a framed prime on the structural–framed spectrum: it names the process of mapping an observed outcome back onto the agents or factors whose actions and omissions produced it, and apportioning credit or blame among them. Its defining structure is a directed assignment from effects to responsible sources, gated by counterfactual and normative tests of dependence, control, and the like.
The placement reads framed. The concept descends from moral psychology and law, importing a moral-legal vocabulary of credit, blame, and culpability, and it is normatively loaded by default — to attribute responsibility is already to assign desert. There is a structural counterpart in the bare credit-assignment operation, as in reinforcement learning's propagation of reward back to actions, which gives the mapping a partly recognizable form; but the full framing presupposes normative agents and imports a moral perspective onto the bare causal chain. Overall, it reads framed.
Substrate Independence¶
Responsibility Attribution is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its structure — a directed assignment from effects back to sources, gated by counterfactual and control tests — is substrate-agnostic and has a precise computational incarnation in the reinforcement-learning temporal credit-assignment problem, alongside legal apportionment, moral psychology, and blameless organizational post-mortems. The transfer evidence is strong and explicit: RL credit assignment maps onto organizational root-cause analysis, and legal apportionment maps onto multi-agent AI. It lands at 4 rather than 5 because the structure presupposes agents and causes, so it does not reach physical substrates.
- Composite substrate independence — 4 / 5
- Domain breadth — 4 / 5
- Structural abstraction — 4 / 5
- Transfer evidence — 4 / 5
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
-
Responsibility Attribution is a decomposition of Causality
Responsibility attribution is the specific shape causality takes when the cause-effect relation is read backward for the purpose of apportioning credit or blame to agents. The causal apparatus -- antecedent, consequent, productive connection, modal robustness -- supplies the input: the counterfactual-dependence test asks whether the outcome would have differed had the agent acted otherwise. Attribution adds normative gates (control, foresight, fault) that select which of the causal contributors bear responsibility, transforming bare causal facts into a directed assignment from effects to morally answerable sources.
Children (4) — more specific cases that build on this
-
Scapegoating is a kind of Responsibility Attribution
Scapegoating is a specialization of responsibility attribution. Both operations map outcomes back onto sources, apportioning blame; scapegoating specializes by concentrating the assignment on a single marked target whose punishment discharges the collective's distress, regardless of actual causal contribution. The general structure of directed-assignment-from-effects-to-sources persists, but the counterfactual and normative gates are bypassed in favor of displacement and concentration. The assignment operation itself remains, while its quality controls collapse — making scapegoating a pathological specialization rather than a separate phenomenon.
-
Fundamental Attribution Error presupposes Responsibility Attribution
Fundamental attribution error presupposes responsibility attribution because it is defined as a characteristic distortion in how the assignment from outcome to agent is performed: over-weighting dispositional sources and under-weighting situational ones, with an asymmetric self-other pattern. Without the prior operation of mapping observed behavior back onto candidate causes and apportioning credit or blame, there is no assignment for the bias to skew. The error is parasitic on the attribution machinery; it specifies which input weights get systematically miscalibrated when agents perform the directed assignment from effects to responsible sources.
-
Great Man Theory presupposes Responsibility Attribution
Great Man Theory attributes major historical outcomes to the actions and characteristics of exceptional individuals rather than to structural or sociological forces, treating those individuals as causally efficacious in ways structural accounts cannot reduce. The position is a particular application of responsibility attribution: mapping outcomes back to agents, with counterfactual-dependence as the test. Responsibility attribution supplies the underlying directed-assignment-from-effects-to-sources operation. Great Man Theory specializes it by concentrating the assignment heavily on individual agents and privileging their counterfactual significance over structural variables in the historical case.
- Self-Handicapping is a decomposition of Responsibility Attribution
Responsibility attribution is the directed assignment from outcomes to responsible sources, gated by counterfactual and normative tests. Self-handicapping is the particular shape this operation takes when an agent strategically engineers the attribution in advance: by introducing a performance-degrading obstacle before the task, the agent ensures that any failure attribution will land on the obstacle rather than on ability. It is a structurally-particularized instance of attribution whose specific machinery is pre-emptive cause-planting to redirect the post-hoc assignment away from internal capability.
Path to root: Responsibility Attribution → Causality → Dependency
Neighborhood in Abstraction Space¶
Responsibility Attribution sits among the more crowded primes in the catalog (14th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.
Family — Experimentation & Validation (18 primes)
Nearest neighbors
- Responsibility Diffusion — 0.84
- Reputation — 0.83
- Experimental Design — 0.83
- Fundamental Attribution Error — 0.81
- Belief Formation — 0.81
Computed from structural-signature embeddings · 2026-05-29
Not to Be Confused With¶
Responsibility attribution must be distinguished first from the Fundamental Attribution Error, which is the closest and most easily confused neighbor. The fundamental attribution error is a specific bias in performing attribution — the systematic tendency to over-weight dispositional, agent-internal causes (character, intent) and under-weight situational, structural ones when explaining others' behavior. It is a characteristic miscalibration of the gates: control and counterfactual dependence are read as higher for salient individuals than the facts warrant. Responsibility attribution, by contrast, is the general assignment operation that this bias distorts. The relationship is that of a process to one of its failure modes: the prime names the normal, well-formed operation; the fundamental attribution error names a predictable way that operation goes wrong. One cannot even state the bias coherently without presupposing the prime, because "over-weighting dispositional causes" is meaningful only against a baseline of correctly weighted attribution. Treating them as the same concept collapses the operation into one of its pathologies, which is exactly the conflation the prime exists to prevent.
Responsibility attribution is also not Accountability, with which it is frequently run together in institutional language. Accountability is a standing relational obligation: a structural condition in which an agent is answerable to some principal for a domain of conduct, established in advance and persisting independently of any particular outcome. A treasurer is accountable for the funds before any money goes missing; the accountability relation is part of the role. Responsibility attribution, by contrast, is the episodic act of determining, after an outcome, who is answerable for that specific result and in what proportion. Accountability sets up who is in the frame to be assessed; attribution does the assessing. The two interlock — accountability structures often define whose conduct attribution will examine, and attribution outcomes feed back into accountability mechanisms (sanction, reward) — but they sit at different points in the cycle. Accountability is the standing answerability; attribution is the situated assignment that draws on it. A system can have crisp accountability relations and still perform attribution badly (assigning blame against the gates), and it can attribute responsibility accurately in a setting with no formal accountability structure at all (a historian apportioning blame for a long-dead regime's failures). The dividing line is temporal and modal: accountability is prospective and standing; attribution is retrospective and event-bound.
Finally, responsibility attribution is distinct from Responsibility Diffusion, which is in some sense its mirror image. Responsibility diffusion is the paradoxical weakening of felt responsibility that occurs when responsibility is spread across many parties — the bystander effect being the canonical case, where each member of a crowd feels less personal obligation to act precisely because others are present. Diffusion is a phenomenon about the experienced motivational pull of responsibility and how it dilutes under distribution. Responsibility attribution, by contrast, is the deliberate, often institutional assignment of responsibility — the operation of apportioning credit or blame across contributors. The relationship is adversarial: diffusion is largely what attribution exists to counter. When a harm has many contributors, diffusion threatens to leave it owned by no one (each contributor's felt share shrinks below the threshold of action), and explicit attribution — naming each party's apportioned responsibility — is the corrective that re-concentrates ownership. This is why the same multi-party situation can be described by both primes from opposite directions: T3's tension (apportionment diluting responsibility) is precisely the seam where the assignment operation of attribution shades into the felt-dilution phenomenon of diffusion. They are not the same because one is a normative-computational operation and the other is a psychological-motivational effect, but they meet at the point where responsibility is spread thin: attribution is the deliberate counter-move against diffusion's drift toward unowned outcomes.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.
Notes¶
Responsibility attribution operates at multiple scales and in both retrospective and prospective modes. Retrospectively, it assigns credit or blame for outcomes that have already occurred (a verdict, a post-mortem, a reward update). Prospectively, the same gated structure is used to allocate responsibility in advance — defining who will be answerable for which outcomes — which is where it most closely brushes against the standing-obligation sense of accountability. Keeping the retrospective operation distinct from the prospective allocation prevents the two from being conflated.
The gating tests are not independent. Control and foreseeability interact: an agent who deliberately puts itself beyond control (drunk driving, willful ignorance) can be held responsible as if it had control and foresight, via doctrines of prior fault. This means the apparent components of the operation are entangled, and a fully decomposed account has to handle the cases where an agent is responsible for the very degradation of its own gating conditions.
The prime carries an implicit assumption that there exist sources that "could have acted otherwise." In strongly deterministic or purely mechanical settings this assumption is philosophically fraught, yet the operation runs anyway — courts, RL agents, and incident reviews all attribute responsibility without resolving the free-will question, because the operation needs only a counterfactual ("had this node been different, the outcome would have differed"), not metaphysical libertarian freedom. The counterfactual gate is what lets the prime function across substrates that include deterministic machines.
A recurring failure mode is mistaking the most salient or most proximate contributor for the most responsible one. The reinforcement-learning framing is a useful corrective precisely because it makes the backward propagation of credit explicit and mechanical, immunizing the analyst against the pull toward the last actor in the chain.
References¶
[1] Fischer, J. M., & Ravizza, M. (1998). Responsibility and Control: A Theory of Moral Responsibility. Cambridge University Press. Grounds the control- and reasons-responsiveness ("guidance control") and foreseeability conditions that gate moral responsibility, supporting both the defining gated structure (841) and the coherence of caused-yet-not-responsible / minor-cause-yet-major-responsibility judgments (851). ↩
[2] Halpern, J. Y., & Pearl, J. (2005). Causes and explanations: A structural-model approach. Part I: Causes. British Journal for the Philosophy of Science, 56(4), 843–887. Counterfactual structural-model account of actual-cause selection: formalizes how a sprawling causal field is narrowed to selected causes (842, 852), the truncation of the causal regress (843), the separation of causal contribution from the accountability judgment built atop it (844), and causation without responsibility (846). ↩
[3] Malle, B. F., Guglielmo, S., & Monroe, A. E. (2014). A theory of blame. Psychological Inquiry, 25(2), 147–186. Path Model of Blame in moral psychology: blame as a graded social-cognitive judgment scaling with perceived intentionality, control, and foreknowledge (849), recurring across attribution contexts (845), decomposable into inspectable components (853), and distinguishable from its biased distortions (847). ↩
[4] Hart, H. L. A., & Honoré, T. (1985). Causation in the Law (2nd ed.). Oxford University Press. Comprehensive analysis of how courts use common-sense counterfactual reasoning to attribute causal responsibility; demonstrates that the same minimal-modification logic underlies legal, moral, and ordinary causal judgment. ↩
[5] Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press. Standard reference on the temporal credit-assignment problem: discounting and eligibility traces back-project credit for a delayed reward across the actions that produced it (850), the same backward propagation that, applied to incident review, resists stopping at the proximate actor (855). ↩
[6] Ross, L. (1977). The intuitive psychologist and his shortcomings: Distortions in the attribution process. In L. Berkowitz (Ed.), Advances in Experimental Social Psychology (Vol. 10, pp. 173–220). Academic Press. Seminal characterization of the fundamental attribution error — the systematic over-weighting of dispositional over situational causes — supporting its framing as a miscalibration of the control and counterfactual gates (854). ↩