Skip to content

Shortcut Learning

Prime #
1178
Origin domain
Data Science & Analytics
Subdomain
machine learning evaluation → Data Science & Analytics
Aliases
Clever Hans Effect, Spurious Feature Learning

Core Idea

Shortcut learning is the structural pattern in which an adapting system discovers a cheap, locally available feature that correlates with success on its training distribution and uses that feature as a stand-in for the structure it was supposed to learn. The system looks competent on the training task because the feature happens to track the target there; off-distribution, where the correlation no longer holds, performance collapses sharply, because no structural knowledge was ever acquired. The defining move is correlation substituting for structure, where the substitution is discovered by the system's own optimization pressure and is invisible to outcome-level evaluation.

The abstraction has a precise shape. There is a target outcome \(T\), a true structure \(S\) that genuinely causes \(T\), and a cheap incidental feature \(C\) that is correlated with \(T\) on the training distribution. The adapting process is graded on \(T\) and has access to both \(S\) and \(C\). Because \(C\) is cheaper to detect or exploit than \(S\), the optimization pressure — gradient descent, reinforcement, attention, social imitation, or natural selection — flows toward \(C\). The substitution is sustained until a distribution shift, or an adversarial probe, breaks the \(C \leftrightarrow T\) correlation, at which point a competence that looked solid evaporates. The general claim is compact: optimization under a sufficient statistic finds the cheapest sufficient statistic, not the target. Nothing in this requires the system to hold beliefs or have intentions; it requires only an adaptive process under outcome feedback with a cheaper-than-structure correlate available in its input.

How would you explain it like I'm…

The Grass Trick

Imagine you're learning to tell cows from sheep, but every cow photo happens to have grass and every sheep photo has a barn. You might secretly learn 'grass means cow' instead of what a cow really looks like. Then someone shows you a cow standing in a barn, and you guess wrong — because you never learned the real difference.

Cheating With Clues

Shortcut learning is when something that's learning finds a cheap, easy clue that *usually* lines up with the right answer and uses that clue instead of really understanding. It looks smart on the practice questions because the clue happens to work there. But when the situation changes and the clue stops matching the answer, it suddenly fails badly, because it never learned the real thing underneath. For example, a program told to spot wolves might just be checking for snow in the background, since its training wolf pictures all had snow. Nobody told it to cheat — it drifted to the easy clue on its own, and you can't catch it just by checking its score.

Cheapest Clue Wins

Shortcut learning is the pattern where an adapting system discovers a cheap, locally available feature that correlates with success on its training data and uses it as a stand-in for the structure it was supposed to learn. It looks competent because that feature happens to track the target *there*; off-distribution, where the correlation breaks, performance collapses sharply, because no real structural knowledge was ever acquired. The defining move is correlation substituting for structure, discovered by the system's own optimization pressure and invisible to score-level evaluation. Concretely: there's a target outcome, a true cause of it, and a cheap incidental feature correlated with the target only on the training set — and because the cheap feature is easier to exploit, the optimization flows toward it. It isn't about beliefs or intentions; any adaptive process under outcome feedback with a cheaper-than-structure correlate available will tend to grab it.

 

Shortcut learning is the structural pattern in which an adapting system discovers a cheap, locally available feature that correlates with success on its training distribution and uses that feature as a stand-in for the structure it was supposed to learn. The system looks competent on the training task because the feature happens to track the target there; off-distribution, where the correlation no longer holds, performance collapses sharply, because no structural knowledge was ever acquired. The defining move is correlation substituting for structure, where the substitution is discovered by the system's own optimization pressure and is invisible to outcome-level evaluation. The shape is precise: there is a target outcome T, a true structure S that genuinely causes T, and a cheap incidental feature C correlated with T on the training distribution; the adapting process is graded on T and has access to both S and C. Because C is cheaper to detect or exploit than S, the optimization pressure — gradient descent, reinforcement, attention, social imitation, or natural selection — flows toward C, and the substitution is sustained until a distribution shift or adversarial probe breaks the C-to-T correlation, at which point apparently solid competence evaporates. The compact general claim is that optimization under a sufficient statistic finds the *cheapest* sufficient statistic, not the target. Nothing requires the system to hold beliefs or intentions; it requires only an adaptive process under outcome feedback with a cheaper-than-structure correlate available in its input.

Structural Signature

a graded target outcome — a true structure that genuinely causes the target — a cheap incidental correlate of the target on the training distribution — an optimization pressure selecting the cheapest sufficient statistic — a distribution shift breaking the correlate-target coupling — a brittleness invariant: collapse concentrated where correlate and target dissociate

The pattern is present when each of the following holds:

  • A graded target. An outcome the adapting process is rewarded on — accuracy, fitness, a score, a decision — measured only at the outcome level.
  • A true structure. The mechanism that genuinely produces the target and is what the process was meant to acquire.
  • A cheap correlate. An incidental feature, cheaper to detect or exploit than the true structure, that tracks the target on the training distribution.
  • An optimization pressure. An adaptive force — gradient descent, reinforcement, selection, social imitation — that, under a sufficient statistic, flows toward the cheapest sufficient statistic rather than the intended one.
  • A correlation that can break. A distribution shift or adversarial probe under which the correlate and the target dissociate, since the correlate was only contingently tied to the target.
  • A localized brittleness invariant. Competence that looked solid collapses, and the collapse is concentrated precisely on the subpopulation or regime where correlate and target come apart — the region in-distribution evaluation never visits.

The components compose so that outcome-level success no longer entails possession of the true structure: the structure forces the sharper question — which feature is being used, and would it still predict the target off distribution — and predicts that brittleness is not diffuse but located exactly where the correlate fails.

What It Is Not

  • Not learning per se. learning is the genuine acquisition of transportable structure; shortcut learning is the substitution of a cheap correlate for that structure — outcome success that carries no structure that survives transport.
  • Not overfitting. overfitting is memorizing noise or sample-specific detail, hurting even in-distribution generalization; shortcut learning seizes a real but incidental correlation and can generalize fine until the correlation breaks off-distribution.
  • Not failed transfer of learning. transfer_of_learning concerns moving genuine competence to a new task; shortcut learning is the prior failure to acquire the structure, so there is nothing transportable to transfer.
  • Not selection bias. selection_bias is a property of how the sample was drawn; shortcut learning is a property of what the optimizer seized within the sample — the cheapest sufficient statistic, not the intended one.
  • Not observational/social learning. observational_learning_social_learning is acquiring behavior by imitation; shortcut learning can occur under any optimization pressure (gradient, selection, imitation) and names what gets learned, not how it is transmitted.
  • Common misclassification. Reading strong training-set performance as "learned the task." Catch it by building a stress set where the cheap correlate and the target dissociate; if competence collapses there, the system learned the shortcut, not the structure.

Broad Use

The pattern appears wherever an adapting system is graded on an outcome metric while having access to incidental features that predict that metric on the training distribution. In machine learning it is the pneumonia detector that learns to recognize the hospital's X-ray machine rather than the lung, the sentiment classifier that latches onto negation tokens, the language model that exploits lexical-overlap heuristics in place of entailment. In education it is the student who learns to read keyword cues in test items rather than the underlying concept and then fails on novel framings — "teaching to the test" made mechanical. In hiring analytics it is the algorithm (or human) that learns a demographic proxy because the proxy correlated with prior hires, without learning the construct those hires were meant to track. In animal cognition it is Clever Hans reading his trainer's posture rather than doing arithmetic, or pigeons trained on categories that learned the background luminance of the training photographs. In biological evolution it appears as runaway sexual selection on a display trait that originated as an honest fitness signal and persisted after the correlation decayed. In clinical prediction it is the sepsis-risk model that fires whenever a clinician has already ordered a sepsis workup, predicting the chart rather than the patient. The substrates differ; the substitution is the same.

Clarity

Without the lens, an evaluator who sees strong training-set performance asks only "did the system learn the task?" With it, the evaluator asks a sharper and answerable question: which feature is the system using, and would that feature still predict the target if the distribution shifted? The reframing converts a vague worry about generalization into a specific claim about feature substitution that can be tested.

The lens also separates two things that ordinary success metrics fuse: being right and being right for a transportable reason. A system can be entirely correct in aggregate on its training distribution and still carry no structure that survives transport. By making the distinction explicit, shortcut learning exposes a whole class of spurious-victory results — the model is right, but for a reason that will not travel — and it does so before deployment rather than after the collapse. It thereby converts "passes training" from a reassurance into a question.

Manages Complexity

The pattern collapses a large family of generalization failures into a single diagnostic: identify the cheapest feature that achieves training success, then check whether it dissociates from the target on out-of-distribution data. That move converts the open-ended question "will this generalize?" into a tractable adversarial one — construct a stress set where the shortcut and the true target disagree, and measure the gap. The size of that gap bounds the share of competence resting on the shortcut.

The compression matters because the alternative is a per-domain catalogue of unrelated-looking failures: hospital-recognition in radiology, keyword-cueing in pedagogy, proxy-discrimination in hiring, charting-artifact in clinical prediction. Treated as instances of one structure, they share one toolkit. The analyst no longer needs domain-specific folklore about each failure; the same diagnostic — find the cheap correlate, break it, watch what happens — applies across every substrate where an adaptive process is graded on an outcome.

Abstract Reasoning

Holding shortcut learning as a unit licenses inferences about what an adaptive system has actually acquired from the mere fact that it succeeds on a metric. Success on \(T\) does not entail possession of \(S\); it entails possession of some sufficient statistic for \(T\) on the training distribution, which may be \(S\), may be \(C\), and under optimization pressure will tend to be whichever is cheaper. This is a structural prediction, not a contingent observation: it follows from the asymmetry that \(C\) is, by construction, easier to exploit than \(S\).

The reasoning generalizes to any system that adapts under outcome feedback — learners, regulators tuning a policy on a proxy metric, evolutionary processes selecting on a fitness correlate. In each, the same prediction holds: where a cheaper correlate of the graded outcome exists, the optimizer will find it, and the resulting competence will be brittle exactly along the axis where the correlate and the target dissociate. The abstraction also predicts where the brittleness lies. It is not diffuse: it is concentrated on the subpopulation or regime where \(C\) and \(T\) come apart, which is precisely the region a naive in-distribution evaluation never visits. This converts "it might not generalize" into "it will fail specifically here," a far more useful claim.

Knowledge Transfer

The structural roles transfer directly, and so do the interventions built on them. The target maps to the graded outcome — accuracy, fitness, a test score, a hiring decision; the true structure to the mechanism one actually wants captured; the cheap correlate to the incidental feature — image artifact, keyword, demographic proxy, posture, documentation cue; the optimization pressure to gradient descent, reinforcement, selection, or social imitation; and the distribution shift to deployment in a new hospital, a novel test framing, a different applicant pool, or a changed environment. Because these roles correspond across substrates, a practitioner who has seen the failure in one domain recognizes it in another without retranslation.

The interventions inherit that portability. Stress-set evaluation — constructing test data where the shortcut and the target dissociate — is the same move whether it is an out-of-distribution benchmark in machine learning, a novel-framing item in pedagogy, or an audit case in hiring. Invariance penalties — training or teaching across multiple environments where the correlate varies independently of the target, forcing recovery of what holds across environments — recur from domain-generalization research to the deliberate variation of surface form across worked examples in instruction. Causal probing — intervening on the correlate while holding the structure fixed to see whether performance moves — is a unified diagnostic across ML interpretability and experimental psychology. Construct-validity audits ask, in hiring, clinical prediction, and policy alike, whether the metric being optimized is the construct of interest or merely a proxy that happens to correlate. The single warning the ML community gives itself — "passes training" is not "learned the task" — imports without modification into education, regulation, and any domain where an adaptive system is rewarded on an outcome it can reach by a cheaper road than the one intended.

Examples

Formal/abstract

Consider a binary classifier trained to predict a target \(T\) with access to two features: the true causal structure \(S\) (expensive to extract) and a cheap incidental correlate \(C\) that satisfies \(P(T \mid C) \approx P(T \mid S)\) on the training distribution but is causally irrelevant. Under empirical risk minimization, the loss gradient does not distinguish "sufficient statistic the designer wanted" from "sufficient statistic that minimizes loss" — it flows toward whichever feature reduces training error per unit of representational cost most efficiently. Because \(C\) is cheaper to detect than \(S\), the optimizer allocates capacity to \(C\): the model achieves near-ceiling training accuracy while encoding little of \(S\). The brittleness is then exactly locatable. Construct a stress distribution in which \(C\) and \(T\) are decorrelated (or anti-correlated) while \(S \to T\) is preserved; on that distribution accuracy collapses toward chance, and the size of the train-to-stress gap bounds the share of the model's apparent competence that rested on \(C\). The compact theorem: optimization under a sufficient statistic finds the cheapest sufficient statistic, not the target. The dictated intervention is invariance regularization — train across multiple environments in which \(C\) varies independently of \(T\), penalizing representations that are not stable across them, which forces recovery of the invariant \(S\).

Mapped back: The ERM model instantiates every role — graded target \(T\), true structure \(S\), cheap correlate \(C\), optimization pressure toward the cheapest sufficient statistic, a decorrelating distribution shift, and brittleness localized exactly where \(C\) and \(T\) dissociate.

Applied/industry

In clinical risk prediction, a sepsis-early-warning model is trained on electronic health records and graded on whether it flags patients who later receive a sepsis diagnosis. The cheap correlate is a documentation artifact: the model learns that an order for a lactate test or blood culture — actions a clinician takes because they already suspect sepsis — strongly predicts the diagnosis. On the training distribution the model looks excellent, but it is predicting the chart, not the patient; deployed where the goal is to alert before the clinician suspects sepsis, it fails precisely on the population it was meant to help. The construct-validity audit and a stress set of patients without prior sepsis workup expose the substitution. The identical structure governs medical imaging: a pneumonia detector trained across hospitals latches onto the X-ray machine's metadata or image-edge tokens that correlate with the sicker hospital's higher base rate, rather than the lung pathology; it collapses at a new hospital, and the fix is multi-site training that varies the scanner independently of the diagnosis. And in hiring analytics, a resume-screening model graded on matching prior hires learns a demographic or school-prestige proxy that correlated with past selection, encoding none of the job-relevant construct; a causal probe — perturbing the proxy while holding qualifications fixed — reveals the shortcut, and the remedy is the same construct-validity audit asking whether the optimized metric is the construct or a proxy.

Mapped back: Across clinical prediction, medical imaging, and hiring the same roles recur — an outcome metric, a true structure the system was meant to capture, a cheaper correlate the optimizer seizes, and collapse localized where the correlate dissociates from the target — and the same interventions transport: build a stress set that breaks the correlate, probe causally, and audit construct validity before deployment.

Structural Tensions

T1 — Shortcut versus Legitimate Feature (sign/direction). The prime flags the cheap correlate as a defect, but a feature cheap to exploit is not automatically illegitimate — it may be genuinely causal and the optimizer right to seize it. The failure mode is false-shortcut alarm: penalizing a model for using an efficient real signal, forcing it onto a costlier path with no generalization gain. Boundary with transferability_overclaim. Diagnostic: does the correlate dissociate from the target off-distribution, or does it hold? A feature that survives the stress set is signal, not shortcut; cheapness alone does not condemn it.

T2 — Outcome Metric versus True Structure (measurement). The whole pattern rests on the gap between what is graded and what was meant, but if the true structure cannot be measured, there is no way to detect the substitution except by its eventual collapse. The failure mode is unobservable-structure complacency: assuming outcome success entails structural acquisition because the structure is never directly checked. Shared with procedure_work_mismatch's graded-versus-real gap. Diagnostic: is there any independent measurement of the true structure, or only the outcome proxy? Without one, "passes training" is irreducibly ambiguous.

T3 — Stress Set Coverage versus Unknown Dissociation (scopal). Stress-set evaluation breaks the known correlate, but the model may rest on a shortcut the designer never imagined, so passing every constructed stress set is not evidence of robustness. The failure mode is stress-set blind spot: certifying a model on the dissociations one thought of while the real shortcut goes unprobed. This is the underspecification tension — the criterion did not constrain the untested axis. Diagnostic: were the stress sets derived from hypothesized shortcuts only, or from the full space of cheap correlates? Coverage is bounded by imagination.

T4 — Invariance Penalty versus Real Heterogeneity (coupling). Invariance training forces representations stable across environments where the correlate varies, but it also penalizes genuinely heterogeneous structure — sometimes the target really does depend on the environment. The failure mode is over-invariance: training away a real environment-dependent effect in pursuit of a transportable invariant that does not exist. Boundary with outlier_leverage's subgroup signal. Diagnostic: is the cross-environment variation a spurious correlate or a real effect modifier? Forcing invariance on genuine heterogeneity discards signal as if it were shortcut.

T5 — Localized Brittleness versus Diffuse Degradation (scalar). The frame predicts collapse concentrated exactly where correlate and target dissociate, which is sharp and useful — but some failures are diffuse, spread across the distribution rather than localized, and the localization prediction then misleads the search. The failure mode is mislocalized debugging: hunting for the dissociation subpopulation when the degradation is broad. Diagnostic: does performance drop on a identifiable regime, or everywhere at once? Diffuse degradation points to a different failure (capacity, noise) than shortcut learning, which is characteristically local.

T6 — Optimization Pressure versus Designer Intent (temporal). The optimizer finds the cheapest sufficient statistic during training, but the cost ordering of features can shift after deployment — a correlate cheap to detect in training may become expensive or unavailable later, and vice versa. The failure mode is static-cost assumption: assuming the shortcut the optimizer chose stays the shortcut, when a deployment-time cost shift makes a previously-ignored path dominant. Boundary with tempo_mismatch. Diagnostic: does the relative cost of correlate versus structure change between training and deployment? A model audited against training-time shortcuts can acquire new ones when the cost landscape moves.

Structural–Framed Character

Shortcut learning sits on the structural side of the structural–framed spectrum, a mixed-structural prime with a low aggregate of 0.3. Its core is a substrate-neutral optimization result — under a sufficient statistic, an adaptive process finds the cheapest sufficient statistic, not the target — and that theorem holds for any process under outcome feedback with a cheaper-than-structure correlate available, which is what pulls the grade well toward the structural end.

The diagnostics lean structural. Evaluative weight and human-practice-bound both read zero. The pattern carries no inherent disapproval until you specify what was learned: a cheap correlate the optimizer seizes is sometimes a genuine causal signal it was right to take, so the prime explicitly refuses to condemn cheapness as such. And it is not human-practice-bound — the substitution runs under gradient descent, reinforcement, and natural selection, with the biological instances doing real work: Clever Hans reading his trainer's posture rather than doing arithmetic, pigeons trained on the background luminance of the photographs, and runaway sexual selection on a display trait that persisted after its honest-signal correlation decayed. None of these involves a human practice; the optimizer is a beak or a genome. The two diagnostics that sit at the midpoint are what keep the grade from going fully structural. The vocabulary half-travels: "training distribution," "stress set," "decision surface," and "invariance penalty" carry an ML home lexicon a new domain must partly adopt. Institutional origin sits at machine-learning evaluation, and invoking the prime half-imports its frame (build a stress set that decorrelates the cheap feature; "passes training" is not "learned the task") and half-recognizes a dynamic already present wherever optimization meets a cheaper road.

The prime's substrate reasoning confirms the reading: correlation-as-substitute-for-structure recurs in ML, education, hiring, animal cognition, and biological evolution, and the structural mechanism — optimization finds the cheapest path to the metric — is substrate-neutral, with the biological cases showing the structure travels beyond its ML dress. That is the mixed-structural signature, here tilted low: a genuinely medium-independent optimization law carried in a field vocabulary it has not fully shed, but demonstrably running in beaks and genomes as readily as in gradients.

Substrate Independence

Shortcut learning is a strongly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its domain breadth is wide and reaches genuinely non-human substrates: correlation-as-substitute-for-structure recurs with the same force in machine learning (the pneumonia detector reading the X-ray machine, the classifier latching onto negation tokens), education (the student reading keyword cues rather than the concept), hiring analytics (the algorithm learning a demographic proxy), animal cognition (Clever Hans reading his trainer's posture, pigeons learning background luminance), biological evolution (runaway sexual selection on a display trait whose honest-signal correlation has decayed), and clinical prediction (the sepsis model predicting the chart rather than the patient). The structural-abstraction component is high because the load-bearing object is a substrate-neutral optimization theorem — under a sufficient statistic, an adaptive process finds the cheapest sufficient statistic, not the target — and that result holds for any process under outcome feedback with a cheaper-than-structure correlate available; the biological cases do real work, since the optimizer there is a beak or a genome with no human practice in sight. Transfer evidence is strong: the diagnostic ("passes training" is not "learned the task") and the remedy (build a stress set that decorrelates the cheap feature) carry across ML, animal-cognition, and evolutionary settings. Only an ML home vocabulary ("training distribution," "stress set," "decision surface") keeps the composite at 4 rather than 5.

  • Composite substrate independence — 4 / 5
  • Domain breadth — 4 / 5
  • Structural abstraction — 4 / 5
  • Transfer evidence — 4 / 5

Relationships to Other Primes

One-hop neighborhood: parents above, mutual partners to the right, children below.Shortcut Learningcomposition: LearningLearning

Parents (1) — more general patterns this builds on

  • Shortcut Learning presupposes Learning

    Shortcut learning presupposes an adaptive process under outcome feedback (learning) and adds the cheaper-correlate-available condition + the collapse prediction — a child-not-duplicate of learning. Dossier-confirmed: 'presupposes learning... adds the cheaper-correlate-available condition.' The 0.9717 'learning' neighbor is the parent, not a duplicate.

Path to root: Shortcut Learning → Learning → Adaptation

Neighborhood in Abstraction Space

Shortcut Learning 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 — Selectivity & Bounded Windows (18 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-06-14

Not to Be Confused With

The nearest existing prime by embedding — at very high similarity — is learning itself, and the contrast is the entire point of the prime. Learning names the genuine acquisition of structure: a process ends with the learner holding the mechanism that produces the target, such that competence transports to new instances. Shortcut learning names the counterfeit: outcome-level success on the training distribution achieved by seizing a cheap correlate, with no structure acquired and competence that evaporates when the correlate breaks. The two are observationally identical on the training distribution — both produce high scores — which is exactly why the confusion is dangerous. The distinction is load-bearing because it determines whether "passes training" is a reassurance or a question: under a learning frame, a high score means the task is learned; under a shortcut-learning frame, a high score means some sufficient statistic was found, possibly the cheap one. A practitioner who cannot separate them will deploy a model (or graduate a student, or trust an evolved trait) that looks competent and collapses precisely where it matters — off the training distribution.

A second genuine confusion is with overfitting, the classic generalization failure most likely to be reached for. Overfitting is the memorization of sample-specific noise: the model fits idiosyncrasies of the training data that do not recur, and the tell is poor generalization even within the distribution — high training accuracy, low held-out accuracy from the same distribution. Shortcut learning is structurally different: the model seizes a real, systematic correlation that is genuinely present in the data, so it generalizes fine to held-out data from the same distribution and fails only when the distribution shifts to break the correlate-target coupling. Overfitting fails on a random held-out split; shortcut learning passes that split and fails on a deliberately decorrelated stress set. The remedies diverge accordingly: overfitting calls for regularization and more data; shortcut learning calls for invariance training across environments where the correlate varies, plus stress-set evaluation. Conflating them sends a practitioner to add data and regularization when the real fix is to vary the environment so the shortcut cannot survive.

A third confusion is with transfer_of_learning. Transfer of learning concerns whether genuine competence acquired on one task carries to a related one — a question that presupposes something real was learned. Shortcut learning is the prior failure: the structure was never acquired, so there is nothing to transfer, and the apparent competence is an artifact of a correlate that does not travel. The relationship is sequential — failed transfer is often the symptom by which shortcut learning is detected, since the shortcut's non-transportability shows up exactly as a transfer failure. But they are distinct claims: transfer of learning asks "does real competence carry across tasks?"; shortcut learning asks "was there real competence to begin with, or only a correlate?" A practitioner who frames the problem purely as a transfer failure may try to bridge the gap between two tasks when the deeper issue is that the source competence was never structural.

For a practitioner, the distinctions sort the diagnosis. If the system genuinely holds transportable structure, it is learning succeeding; if it fails on a random held-out split from the same distribution, it is overfitting; if real competence simply does not carry to a new task, it is a transfer_of_learning gap; and if competence passes in-distribution but collapses on a stress set that decorrelates a cheap incidental feature from the target, it is shortcut learning — the only one whose remedy is to break the correlate by varying the environment and to audit construct validity before trusting the score.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.