Comparison

Prime #

None

Origin domain

Cognitive Science

Also from

Experimental Design & Statistics, Philosophy, Linguistics & Semiotics

Aliases

Comparative Evaluation, Comparison Operation

Core Idea

Comparison is the structural operation of placing two or more items — the comparands — under a shared frame, selecting one or more dimensions along which they will be co-considered, applying an alignment rule that makes them commensurable enough to relate, and reading off an output relation: same or different, greater or lesser, analogous or unmatched, ranked or unranked. The cognitive-science treatment by Holyoak and Thagard (1989) characterizes this as a constraint-satisfaction process running over structural, semantic, and pragmatic constraints — a formal model of the role-slots the operation requires.^[1] It turns isolated properties of individual items into relational information between them. The operation is what cognitive scientists, philosophers of method, and statisticians have all converged on as a foundational move of structured thought — Tversky's (1977) feature-matching account treats similarity itself as the output of a comparison process over weighted feature sets, parameterized by direction and salience.^[2]

Comparison can foreground likeness, contrast, ranking, analogy, equivalence, deviation, or fit, but the underlying operation is the same in each case: place the items in commensurable view and read off the relation. Its core commitments are two. First, comparison is relational — it generates information between items, not about a single item considered alone. Second, it is framed: comparability requires shared dimensions or criteria, even informal ones. Without a frame, there is no comparison, only juxtaposition; without dimensions, there is no determinate output, only suggestion.

A crucial consequence falls out of the framing condition: comparison names the operation, not the result. The catalog already contains result-shaped neighbors — contrast (the difference reading), analogy (the deep-mapping reading), commensurability (the precondition that a common metric exists), classification (assignment to categories). What none of those name is the underlying move itself: take two or more items, place them in a shared frame, and let a relation appear. Comparison is that move.

How would you explain it like I'm…

Looking at two things together

When you hold two apples next to each other to see which one is bigger, redder, or shinier, that's comparing. You can't really tell about one apple by itself — you need another to look at next to it. Comparing tells you how things are the same or different.

Putting things side by side

Comparison is what we do when we place two or more things side by side and look at them under the same idea — like size, color, speed, or fairness. You pick what to look at, line the things up so the question makes sense, and then read off whether they're the same, different, bigger, smaller, or alike in a pattern. Comparison gives you information that lives between the things, not inside any single one. Without a shared idea to compare on, it's just sitting next to each other.

Relating items under a shared frame

Comparison is the cognitive and methodological operation of placing two or more items under a shared frame, choosing dimensions along which to consider them, applying an alignment rule that makes them commensurable, and reading off a relation: same or different, greater or lesser, analogous, ranked, or unmatched. It turns isolated properties into relational information that lives between items. It is always framed: comparability requires shared dimensions, even informal ones, otherwise items are merely juxtaposed. Comparison names the operation itself, not its result. Specific result-shaped concepts — contrast, analogy, commensurability, classification — sit downstream of the same underlying move: place items in a shared frame and let a relation appear.

 

Comparison is the structural operation of placing two or more items — the comparands — under a shared frame, selecting dimensions for co-consideration, applying an alignment rule that renders them commensurable, and reading off an output relation: identity, difference, rank, analogy, equivalence, or deviation. Cognitive science models this as a constraint-satisfaction process running over structural, semantic, and pragmatic constraints; Tversky's feature-matching account treats similarity itself as the output of weighted feature comparison parameterized by direction and salience. Two commitments are core. First, comparison is relational: it generates information between items, not about an item considered alone. Second, it is framed: without shared dimensions there is no determinate output, only juxtaposition. Importantly, comparison names the operation, not the result. Contrast, analogy, classification, and commensurability sit downstream as result-shapes of this same underlying move.

Structural Signature

Comparison encodes a structural pattern: co-framing → dimension selection → alignment → relation read-off. The pattern separates an unstructured set of items from a structured relational claim about those items, and names the work required to get from one to the other; Gentner's (1983) structure-mapping theory formalizes the same separation in the analogy case, distinguishing the surface-level comparands from the higher-order relational structure that the alignment rule must preserve.^[3]

Recurring features:

• Two or more items placed under a shared frame
• Dimensions of comparison selected from a larger space
• Alignment rule that makes comparands commensurable
• Output relation (same/different, greater/lesser, analogous, ranked)
• Frame-relativity of the relation produced
• Operation that generates relational information, not properties
• Distinction between the comparing move and its readings

The structural insight is robust across substrates. A perception experiment measuring a just-noticeable difference, a controlled biology trial scoring two genotypes against a phenotype, a literary critic reading a simile, a benchmark suite scoring two language models, a metrologist comparing a sample against a calibration standard — each instantiates the same five-role machinery with different comparands, frames, dimensions, alignment rules, and output relations. Goodman's (1972) sharp critique of bare similarity claims — that similarity is empty without a respect-of-comparison — is the canonical statement of why the dimension-selection role is load-bearing and cannot be elided.^[4]

What It Is Not

Comparison is not a property of a single item. A claim like "this object is large" looks unary but is always implicitly comparative — large relative to an implicit reference class. Surfacing the hidden comparand turns covert comparison into explicit comparison, but the operation does not begin when made explicit; it was already running. Comparison is also not the output it produces. The output of comparison can be a similarity score, a difference, an ordering, an analogical mapping, or a categorical assignment, but the operation that yields any of these is one and the same.

Nor is comparison reducible to measurement. Measurement assigns numerical values to a quantity by comparison against a unit or standard, so every measurement is comparison-shaped — but most comparison is not measurement. A qualitative judgment that two species share a body plan, an analogical mapping between a circuit and a hydraulic system, a literary critic's reading of a simile — these are comparisons that produce relational information without numerical assignment. Comparison is broader than measurement; measurement is comparison specialized to numerical output along a metric dimension, exactly the formal characterization the International Vocabulary of Metrology (JCGM 200:2012) gives by routing every measurement through a calibration chain anchored in a reference standard.^[5]

Comparison is not the same as contrast, though the two are often used interchangeably in ordinary speech. Contrast is comparison whose dimension selection and output relation are oriented toward difference: the operation runs, but the analyst foregrounds the gap rather than the overlap, a foregrounding that Markman and Gentner (1996) show empirically as the alignable-difference reading produced by the same structural-alignment machinery that yields commonalities.^[6] A contrast is therefore a kind of comparison — the difference-emphasizing kind — not a separate operation. The same comparands compared on the same dimensions can yield either a similarity reading or a contrast reading depending on the analyst's interest; the underlying machinery is identical. The DAG edge runs comparison → contrast as presupposes, flipping the historical reverse-subsumption in which contrast (the more familiar English word) was treated as the umbrella — a flip that Markman and Gentner's (1996) demonstration of difference judgments as a product of the similarity-comparison process directly supports.^[6]

Comparison is not judgment or evaluation. A comparison can feed into an evaluative judgment (this option is better than that one along criterion X), but the comparison itself produces a relational fact, not a normative ranking. Evaluation requires an additional layer in which one of the read-off relations is privileged as "better" by some external value standard. Plenty of comparisons produce no evaluation at all — comparing two artifacts on age, or two languages on phoneme inventory, generates relational information without imputing value.

Finally, comparison is not grouping or classification. Classification assigns items to categories; the result of classification is a partition of a set into kinds. Comparison places items in a relation; the result is a structured claim about how they stand to each other. Classification uses comparison as a subroutine (the item is matched against category prototypes or definitions), but adds the act of assignment, which comparison alone does not perform.

Broad Use

Cognition and perception: Recognizing similarity, difference, and order is among the most basic cognitive operations. Feature-detection circuits in early visual cortex literally implement comparison operations on adjacent stimuli; psychophysical methods quantify just-noticeable differences as the threshold at which a comparison output flips from "same" to "different." Pre-linguistic infants and non-human animals perform feature-matching and ordinal comparison without methodological scaffolding.

Scientific method and experimental design: Controlled comparison — Mill's methods, A/B tests, randomized controlled trials, difference-in-differences designs — is the structural core of causal inference. The experimenter holds the frame constant (matched conditions) and varies one dimension (the treatment), so any output difference is attributable to that dimension rather than to confounders. Campbell and Stanley's (1963) treatment of internal validity is, at its core, an enumeration of the failure modes of the alignment rule in controlled comparisons — selection, history, maturation, and the rest are all ways the frame fails to hold the comparands commensurable.^[7]

Physics, at the substrate extreme: CPT-symmetry tests in particle physics — measurements of the proton/antiproton mass ratio, comparisons of meson and antimeson decay rates — are pure controlled comparisons at a substrate where no human practice is in view. The frame is the experimental apparatus, the comparands are matter and antimatter species, the dimension is some observable (mass, lifetime, magnetic moment), the alignment is the symmetry hypothesis being tested, and the output relation is "indistinguishable to within experimental precision." That this is recognizably the same operation as a behavioral A/B test is strong substrate-independence evidence.

Biology: Controlled biological comparison — knockout-vs-wildtype crosses, treatment-vs-sham in clinical trials, common-garden experiments holding environment fixed across genotypes — sits between physics and the human-practice domains. Cohort studies and case-control epidemiology extend the same logic to populations where randomization is impossible, with attendant alignment-rule risks (confounding, selection bias) that are the central concerns of the field.

Literature and rhetoric: Simile, metaphor, juxtaposition, and parallel structure are comparison used for expressive, persuasive, or evocative effect rather than for inference. Juxtaposition is the pre-comparative setup that supplies comparands and a rough frame without itself performing the dimension selection or alignment — the curator places the two paintings side by side; the viewer runs the comparison. Markman and Gentner (1996) demonstrate experimentally that placing items side by side prompts the viewer to construct an alignment, with commonalities and alignable differences only emerging once the structural-mapping step is actually performed.^[6] Hofstadter and Sander's (2013) treatment of analogy as the core of cognition argues that even the most rhetorical-seeming comparisons run on the same dimension-selection and alignment machinery as scientific ones, with the output relation simply weighted toward expressive resonance rather than truth-preservation.^[8]

Evaluation and benchmarking: Comparing options against criteria — cost-benefit analysis, multi-criteria decision analysis, performance benchmark suites — is comparison whose output relation is then fed into a downstream evaluative or selection step. The benchmarking literature in computer science, with its careful attention to held-constant test inputs and varied system-under-test, is comparison's industrial-scale instantiation.

Analogy and case-based reasoning: Comparing a current situation to known prior cases to transfer insight is comparison whose alignment rule is structural mapping. Gentner's (1983) structure-mapping theory of analogy specifies the alignment rule precisely — preserve higher-order relational structure, drop surface features — and treats the analogical output as a transferable inferential pattern.^[3]

Measurement and metrology: Every measurement is a comparison against a unit or reference standard. The international metrology infrastructure — primary standards, calibration chains, traceability — is engineering devoted to making the alignment rule in measurement comparisons trustworthy at scale.

Clarity

A core function of "comparison" as a named prime is to surface the hidden machinery in claims that present as direct readings of the world. An unargued "these are similar," "this is better," or "X is unlike Y" is a comparison output whose comparands, frame, dimensions, and alignment rule have all been left implicit. Naming the operation forces the prior question: similar under what frame, along what dimensions, by what alignment rule? The same two organisms compared on body plan look very different from the same two compared on metabolic pathways; the same two policies compared on cost look very different from the same two compared on equity outcomes. Distinguishing the operation from its readings matters because the readings are not free-standing — they are functions of choices made earlier in the pipeline, and surfacing those choices is the analyst's leverage.

The clarity benefit extends to detecting covert comparisons. Claims that look unary — "this is large," "this is fast," "this is unfair" — are comparison outputs in disguise, with the reference class hidden. Naming comparison as a prime makes it natural to ask "compared to what?" and to refuse the claim until the reference is specified.

It also clarifies the structure of disagreement. When two analysts produce different comparison outputs for the same pair of items, the disagreement can be located at any of four sites: they may have chosen different dimensions, applied different alignment rules, used different frames, or simply read the relation differently. Without comparison vocabulary, these disagreements look like flat contradictions; with it, they become locatable and arguable.

Manages Complexity

Comparison decomposes an evaluative situation into five concrete roles: the comparands (the items being compared), the comparison frame (the shared context in which they are co-considered), the dimensions or criteria along which the comparison runs, the alignment rule that makes the comparands commensurable enough to relate, and the output relation — same/different, more/less, analogous, matched, incompatible, or ranked. Medin, Goldstone, and Gentner (1993) argue at length that any usable theory of similarity requires precisely this kind of role-decomposition — most centrally the explicit "respects" along which the comparison runs — without which similarity-talk collapses into the vacuity Goodman warned of.^[9]

Once those roles are named, an opaque "these are similar" or "this is better" becomes a structured claim with explicit machinery. The analyst can interrogate any single role without challenging the rest: were the comparands appropriately chosen, or was the comparison rigged by cherry-picking? Is the frame neutral, or does it pre-favor one side? Are the dimensions exhaustive, or has the comparison been narrowed to dimensions on which one comparand happens to win? Does the alignment rule force a false commensurability, treating non-equivalent items as equivalent under a metric that does not really apply to both? Auditing the alignment rule is the structural form of methodological critique in experimental design — selection bias, confounding, and equating failures are all alignment-rule failures in disguise — and the same audit transfers to cross-cultural and cross-domain comparison, as Sartori (1991) argues in his account of "cat-dog" miscomparisons: aggregates assembled under an alignment rule that papers over genuine non-equivalence, defeating the comparison from inside.^[10] Is the output relation correctly read, or has a same/different distinction been collapsed into a ranking that does not survive scrutiny?

The decomposition converts unargued comparative judgments into auditable ones. This is exactly the move that statistics education makes when it teaches students to distinguish a population, a sample, a treatment, a control, and an outcome measure — all of which are role-slots in a controlled comparison. It is also the move that critical literary analysis makes when it asks "tenor and vehicle?" of a simile. Different vocabularies, same decomposition.

The complexity-management payoff is large. Comparison-shaped problems abound — every claim of similarity, difference, ranking, fit, equivalence, analogy, or categorical match is comparison-shaped — and a single structural vocabulary lets the analyst transfer diagnostic moves across all of them. A trick learned for detecting a rigged frame in a corporate benchmark report is the same trick that detects a rigged frame in a legislative impact analysis.

Abstract Reasoning

Comparison supports the counterfactual move: if the frame, dimensions, or alignment rule were different, the output relation would change in this specifiable way. That move is what makes comparison the structural core of controlled experimentation — the experimenter holds the frame constant (matched conditions) and varies one dimension (the treatment), so any output difference is attributable to that dimension. The same abstract operation underwrites benchmarking (hold the task constant, vary the system), analogical reasoning (hold the structural mapping constant, vary the surface domain), and the literary uses (hold the juxtaposition constant, read the relation as expressive content).

A defining feature of comparison is its frame-relativity: the same two comparands compare differently along different dimensions, so any comparative claim is implicitly indexed to a frame. Surfacing that index — making the frame explicit — is the abstract reasoning leverage the prime provides. It enables the analyst to construct deliberate counterfactual reframings: "compare these two on cost, then again on equity, then again on durability — and see how the output relation moves." A reframing exercise that would be inarticulate in flat similarity-talk becomes routine in comparison vocabulary.

Comparison also enables a sharp distinction between robust and frame-dependent relations. A relation that survives reframing across many reasonable dimension choices is robust; one that collapses or inverts under minor reframing is frame-dependent. The robustness question is itself a meta-comparison — comparing the output of comparison under varied frames — and it underpins concepts like external validity in experimental design, generalization in machine learning evaluation, and structural soundness in analogical inference.

Knowledge Transfer

The five-role structure transfers intact across substrates. A biologist comparing two species on phenotype, a metrologist comparing a sample to a calibration standard, a literary critic reading a simile, an economist running a difference-in-differences study, a particle physicist running a CPT test, and a perception researcher measuring a just-noticeable difference are all instantiating the same operation with different comparands, frames, and dimensions; Lijphart's (1971) systematic comparison of the experimental, statistical, and comparative methods in political science is the canonical demonstration that the same role-slots underlie all three traditions, with the differences confined to how each fills the alignment rule.^[11]

The substrate-furthest cases are the strongest transfer evidence. In particle physics, comparison runs entirely without human social or institutional scaffolding — the apparatus implements the frame, nature supplies the comparands, and the output is a number with an uncertainty interval. In early-visual-cortex feature detection, comparison runs in milliseconds in non-human animals and pre-linguistic infants, again without any methodological apparatus. These cases rule out any suspicion that comparison is a specialty of formal science or human reasoning. It is a substrate-independent operation that happens to have particularly powerful institutional implementations in experimental method and benchmarking.

The pedagogical transfer is also clean. A practitioner who has internalized the five-role structure in one domain can use it to diagnose claims in another. An experimental designer can read a literary simile in comparison-vocabulary (what are the comparands, the frame, the implied dimensions of mapping?) and discover that the same machinery is running. A literary critic can read an A/B test in comparison-vocabulary and identify the same role-slots, with different content in each.

Examples

Formal/abstract

Particle physics — CPT symmetry test: Consider an experiment measuring the magnetic moment of the proton and the antiproton to test the CPT theorem. The comparands are the proton and the antiproton, prepared in a Penning trap. The comparison frame is the apparatus, which by construction holds the magnetic field, trap geometry, and measurement protocol identical for both species. The dimension is the cyclotron-to-Larmor frequency ratio, from which the magnetic moment is extracted. The alignment rule is the assumption that the trap behaves symmetrically with respect to charge sign (a non-trivial assumption requiring its own calibration). The output relation is "the magnetic moments are equal to within experimental precision," and Smorra et al.'s (2017) BASE-collaboration measurement of the antiproton magnetic moment as −2.7928473441(42) nuclear magnetons — agreeing with the proton at the parts-per-billion level — is the current canonical instance.^[12] Mapped back: Every role of the comparison operation is filled, and every role can be challenged independently — comparand preparation, frame symmetry, dimension choice, alignment-rule fidelity, output reading. The comparison operation is fully recognizable here, at a substrate maximally far from human practice.

Controlled biology — common-garden experiment: Two plant genotypes, A and B, are grown side by side in a common garden under matched soil, water, and light. The comparands are the two genotypes; the frame is the garden, which by design holds environmental factors constant; the dimension is biomass at harvest; the alignment rule is the randomized spatial placement that controls for within-garden microclimate variation; the output relation is "genotype A produces 17% more biomass than genotype B in this environment." The exact same five-role structure governs a Mill's-method causal inference, a difference-in-differences economic study, a benchmark run comparing two language models on a reasoning suite, and a literary simile ("my love is like a red, red rose" — comparands my-love and rose, frame poetic praise, dimensions vitality and freshness, alignment by metaphorical mapping, output relation positive resemblance). Mapped back: Comparison is the umbrella; controlled comparison, simile, juxtaposition, and benchmarking are its frame-and-dimension specializations. The same five role-slots run across maximally different content.

Applied/industry

A/B test on a checkout page: An e-commerce team compares two checkout-page designs on conversion rate. The comparands are the two page variants; the comparison frame is "users arriving at checkout during the same time window with traffic split randomly"; the dimension is conversion rate; the alignment rule is the randomization, which makes the user populations exchangeable, so any rate difference is attributable to the page rather than to the users; the output relation is "variant B converts at a higher rate." Notice that the same two pages compared on a different dimension — load time, accessibility score, brand consistency — could yield the opposite output relation. The comparison is well-defined only once the frame and dimensions are fixed. Mapped back: Industrial A/B testing is a controlled comparison whose alignment rule (randomization) makes the comparison machinery cheap and trustworthy at scale. The same operation that runs in a biology common-garden experiment runs in a checkout-conversion test; the only differences are the comparands, the dimensions, and the alignment-rule implementation.

LLM benchmark suite: A team comparing two language models on a reasoning benchmark sets the comparands (the two models), the frame (the benchmark task set, the prompting protocol, the decoding parameters), the dimensions (per-task accuracy, calibration, latency, cost), the alignment rule (identical prompts and grading rubrics applied to both), and the output relation (a per-dimension win/loss/tie record). Benchmark methodology debates — about test-set contamination, prompt sensitivity, grader reliability — are debates about each of these role-slots in turn. Mapped back: Modern AI evaluation is comparison at industrial scale, and its methodological pathologies (cherry-picked dimensions, contaminated frames, brittle alignment rules) are diagnosable in exactly the comparison vocabulary the prime supplies.

Structural Tensions

T1: Frame neutrality is rarely achievable, but frame partiality is rarely detectable. Comparison requires a shared frame; the frame is usually chosen by the analyst; an analyst with a stake in the output has every incentive to choose a frame that favors a preferred comparand. Detecting frame partiality from the outside requires constructing alternative frames and seeing whether the output relation moves — but alternative frames are themselves chosen by the critic, who may have opposing stakes. Frame neutrality is more an aspiration than an achievable state. The tension is unavoidable: every comparison is frame-relative, every frame is chosen, and every chooser has interests. The practical response is triangulation — reporting comparison outputs under multiple frames and being transparent about the choice space — but triangulation is costly and is itself comparison.

T2: Dimension selection trades exhaustiveness against tractability. A comparison along one dimension is tractable but narrow; comparison along many dimensions is exhaustive but produces a multi-dimensional output that may not collapse into a single relation. Multi-criteria decision analysis can produce Pareto-dominance verdicts ("A beats B on every dimension") but more often produces Pareto-incomparability ("A wins on cost, B wins on durability"), which is a comparison-output that does not yield a clean ranking. The tension between rich and decidable comparison is structural: adding dimensions increases fidelity but reduces the chance of a clean output, and the analyst must choose where on this curve to sit.

T3: Alignment rules can force false commensurability. Making comparands commensurable requires an alignment rule; aggressive alignment can paper over genuine non-equivalence. Comparing student test scores across school districts requires alignment rules (curriculum equivalence, test-form equating, demographic weighting) any of which can force commensurability where it does not really hold. The tension is that without an alignment rule there is no comparison, but with too strong an alignment rule the comparison conceals the differences it purports to measure. The "apples and oranges" complaint is precisely a complaint about alignment-rule overreach.

T4: Comparison can be morally weaponized by selective framing. Every comparison output is consumable as evaluation, and the framing of any comparison can be steered toward an evaluative conclusion. Comparing two policies on cost but not on equity, two groups on income but not on opportunity, two artists on technique but not on intent — each is a comparison whose framing pre-determines the evaluative reception. Comparison is morally neutral as an operation but morally loaded in deployment; the same prime that underwrites careful science underwrites tendentious rhetoric. Naming comparison as a prime helps surface this, but the same machinery is at work.

T5: Operation-versus-result conflation is structurally persistent. The same word "comparison" denotes both the operation and its output, and the ambiguity infects ordinary speech and even technical practice. "Run a comparison" denotes the operation; "the comparison shows X" denotes the output; "the comparison is unfair" can be either — unfair operation, or unfair output read off a fair operation. The tension is that the language does not distinguish the levels, so the analyst must keep them distinct in thought even when the vocabulary blurs them. Treating comparison as a prime — explicitly the operation, with separate names for each kind of output — is the principled response, but it does not eliminate the slippage.

T6: Comparison produces relational information, but stakeholders demand absolute claims. A comparison output is intrinsically relational — A is greater than B, A is similar to B — but downstream consumers often need an absolute claim ("A is good") or a context-free ranking ("A is best"). Converting relational outputs into absolute claims requires a further step (an external standard, an aggregation, a normative interpretation) not part of the comparison operation. Comparison is the cleanest tool for generating defensible relational claims and a poor tool for absolute ones, yet consumers persistently treat its outputs as absolute. Reviewers of benchmarks, ratings, and impact assessments routinely make this slippage.

Structural–Framed Character

Comparison sits at the structural end of the structural–framed spectrum: the operation of placing items under a shared frame and reading off a relation is statable abstractly across cognition, philosophy of method, statistics, and any substrate that supports relational reasoning. It is the bare move on which classification, contrast, ranking, analogy, and equivalence judgments all depend.

No domain vocabulary needs to come along; "place items in a common frame and read off the relation" is field-neutral. The prime carries no evaluative weight — comparing is descriptive of a relational operation, not normatively loaded. Institutional origin reads zero: no school, court, or convention is presupposed. Human-practice-bound also reads zero in its bare form, though a faint cognition-binding lingers from origin since most paradigm cases involve cognitive comparison; the formal operation, however, is exercised equally well by a statistical test, a learned similarity function, or a sorting algorithm. Import-vs-recognize is recognition: when a statistician runs a between-group comparison or a vision system computes feature-based similarity, the relational operation is already inherent in the structure of the problem, not imported from cognitive science. On the spectrum, the verdict is canonical-structural — a bare relational operation that nearly any substrate can support.

Substrate Independence

Comparison is about as substrate-independent as a prime can be — composite 5 / 5 on the substrate-independence scale. The operation is a single substrate-neutral move: place two or more comparands under a shared frame, select dimensions of co-consideration, apply an alignment rule that makes them commensurable, and read off a relational output. Every diagnostic lands at the ceiling. Domain breadth is maximal because the same co-framing operation recurs across cognitive perception, scientific method (controlled comparison, A/B testing), literary device (simile, juxtaposition), benchmarking, evaluation, and analogical reasoning. Structural abstraction is at the top because the pattern is purely relational: it turns isolated properties of individual items into relational information between them, with no commitment to any home vocabulary or medium. Transfer evidence is just as strong, since cognitive scientists, philosophers of method, and statisticians have all converged on the same shared-frame-plus-alignment-rule structure, and the Tversky feature-matching formulation, the controlled-trial structure, and the benchmarking apparatus all instantiate the same core. The verdict is that comparison is a paradigm structural prime, one of the catalog's canonical 5s, recognized rather than translated wherever items are brought under a common frame to extract relational information.

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

Relationships to Other Abstractions

Current abstraction Comparison Prime

Parents (1) — more general patterns this builds on

• Comparison is a decomposition of Self Checking Prime

  The comparator that flags divergence between the redundant representations.

Children (27) — more specific cases that build on this

• Simile Domain-specific is a kind of Comparison

  Simile is Comparison specialized to a marked, figurative, usually single-attribute rhetorical form that preserves the distinctness of tenor and vehicle.

• Analogy Prime is a kind of Comparison

  Analogy is a specialization of comparison in which the alignment rule is structural role-mapping rather than feature-matching.

• Juxtaposition Prime is a kind of Comparison

  Juxtaposition is a specialization of comparison in which proximity placement is the alignment rule and relational reading is the output.

▸ Show 24 more

Hierarchy path (1) — routes to 1 parentless root

• Comparison → Self Checking

Neighborhood in Abstraction Space

Comparison sits among the more crowded primes in the catalog (1^st 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 — Structure, Decomposition & Relational Mapping (39 primes)

Nearest neighbors

• Analogy — 0.81
• Interpretation — 0.79
• Juxtaposition — 0.78
• Classification — 0.76
• Bias — 0.76

Computed from structural-signature embeddings · 2026-07-26

Not to Be Confused With

Comparison must be distinguished from contrast, its closest neighbor in the catalog and the source of a long-standing reverse-subsumption confusion. Contrast is comparison whose dimension selection and output relation are oriented toward difference: the operation runs, but the analyst foregrounds the gap rather than the overlap or the ordering. Contrast is therefore a specific kind of comparison — the difference-emphasizing kind — not a separate operation. The same two comparands compared on the same dimensions can yield either a similarity reading, a difference reading, or an ordering reading depending on the analyst's interest; the underlying machinery is identical. The DAG edge runs comparison → contrast as presupposes (comparison must be in place for contrast to be the specific reading it is), and the historical confusion in which contrast was treated as the parent stems from contrast being a more familiar word in ordinary English than the more abstract "comparison." Once the operation/reading distinction is in view, the direction of subsumption flips and contrast settles into its proper place as a child.

Comparison must be distinguished from simile and analogy, also specific kinds of comparison rather than separate operations. A simile is comparison whose comparands are drawn from different domains, whose dimension is chosen for expressive resonance, and whose alignment rule is loose metaphorical mapping; the output relation is resemblance weighted toward affective rather than inferential payoff. An analogy is comparison whose alignment rule is structural mapping — preserve higher-order relational structure, drop surface features — and whose output supports inferential transfer from source to target domain. Simile and analogy share the comparison machinery but specialize it: simile foregrounds expressive output, analogy foregrounds inferential output. The DAG edges run comparison → simile and comparison → analogy as decompose. A frequent error is treating analogy as the umbrella and simile as its subtype; the cleaner reading is that comparison is the umbrella, with simile and analogy as siblings differing in alignment-rule structure.

Comparison must be distinguished from measurement, a related but distinct operation. Measurement is the assignment of numerical values to a quantity by comparison against a unit or reference standard, so every measurement is comparison-shaped — the comparand is the item being measured, the frame is the metrological context, the dimension is the quantity of interest, the alignment rule is the calibration chain, and the output is a number with units. But measurement is more specific than comparison: it requires that the dimension be a measurable quantity (continuous or counting), that the alignment rule yield a numerical output rather than a qualitative one, and that the comparand-to-standard relation be metrically meaningful. Most comparison is not measurement — a qualitative judgment that two body plans are similar, an analogical mapping between a circuit and a hydraulic system, a literary critic's reading of a simile produce relational information without numerical assignment. Measurement is comparison specialized to numerical-output-against-a-metrological-standard; comparison is the broader operation that includes measurement as one specialization and includes many non-numerical specializations as siblings.

Comparison must be distinguished from classification, the result of a sorting-by-comparison process. Classification assigns items to categories; comparison places items in a relation. Classification uses comparison as a subroutine — the item is matched against category prototypes — but adds the act of assignment, which comparison alone does not perform. A comparison yields "X is more like prototype P1 than P2"; the further step of assigning X to category C1 is the classification operation. Classification is a downstream consumer of comparison output. A biologist who compares two specimens on morphology produces a relational claim; one who classifies a specimen into a species is doing comparison-plus-assignment, with the comparison subroutine running first.

Comparison must finally be distinguished from juxtaposition, the bare placement of items side by side without yet performing the operation. Juxtaposition supplies the comparands and often the rough frame, but does not select dimensions, apply an alignment rule, or read off an output relation. Juxtaposition is pre-comparative: it sets up the operation but does not run it. A museum curator who juxtaposes two paintings invites the viewer to compare them but does not perform the comparison; the viewer supplies the dimensions, alignment, and output relation. The DAG edge runs comparison → juxtaposition as subsumption (juxtaposition is a degenerate or pre-completed comparison). Juxtaposition is not a withheld comparison; it is a not-yet-comparison, a setup awaiting the choices that would convert it into one.

Solution Archetypes

Solution archetypes in the catalog that build on this prime — directly (this prime is a source ingredient) or as a related prime.

Built directly on this prime (11)

• Abductive Explanation Selection: Turn a surprising observation into a ranked, provisional best explanation, while keeping rivals, uncertainty, and revision triggers visible.
  ▸ Mechanisms (8)
• Context-Bounded Meaning Recovery: Make interpretation accountable by explicitly binding a reading to a substrate, a context, a framework, evidence marks, and a boundary around plausible alternatives.
  ▸ Mechanisms (14)
• Dimensioned Comparison Framing: Make comparison legitimate by aligning the items, dimensions, scales, context, and relation-readout rule before drawing conclusions.
  ▸ Mechanisms (8)
• Directed Asymmetry Mapping and Calibration: When two sides of a relation are not interchangeable, make the direction and dimensions of imbalance explicit before choosing symmetric treatment, side-specific treatment, compensation, or containment.
  ▸ Mechanisms (12)
• Emergent Similarity Partitioning: Find provisional groups by similarity when labels are not given, then validate and interpret the partition before using it.
  ▸ Mechanisms (10)
• Empirical Cluster Discovery: Discover provisional groups in unlabeled observations by making representation, similarity, validation, interpretation, and downstream use explicit.
  ▸ Mechanisms (9)
• Realized-Possible Outcome Gap Mapping: Compare what a process actually produced with what it could credibly have produced, then treat the gap as the main diagnostic object.
  ▸ Mechanisms (9)
• Reference-Baseline Deviation Flagging: Make departure meaningful by declaring the reference, calculating the observed-minus-expected difference, and recording the deviation as a fact with scope, direction, magnitude, and context.
  ▸ Mechanisms (10)
• Regret-Signal Calibration: Use regret as a calibrated counterfactual signal: compare the actual outcome with a credible better forgone alternative, then route the signal to learning, reversal, repair, or closure.
  ▸ Mechanisms (10)
• Salience-Significance Decoupling: Separate what got attention from what deserves weight.
  ▸ Mechanisms (12)
• Structured Comparative Case Design: Select comparable cases with an explicit contrast logic, align what is measured and when, and use cross-case differences plus within-case evidence to test causal explanations.
  ▸ Mechanisms (16)

Also a related prime in 15 archetypes

• Bidirectional Conceptual Translation: Translate concepts between frameworks by mapping meaning, use, assumptions, and consequences while making gaps and losses explicit.
• Counterfactual Proximity Signal Calibration: Calibrate how much an almost-happened better or worse outcome should teach, motivate, warn, or matter.
• Durable Identifier Binding: Create a durable handle for a referent, bind it in an authoritative record, and maintain enough lookup, lifecycle, and audit rules that later references can rely on the handle without re-describing the entity.
• Emic-Etic Dual-Account Interpretation: Preserve insider and outsider descriptions as separately governed accounts, then use their mismatch as evidence instead of forcing premature translation into one frame.
• Feasible-Alternative Comparator Calibration: Judge real options against reachable alternatives, not against perfection.
• Funnel Attrition Localization: Represent an ordered process as denominator-preserving stages, measure where the population is lost, and prioritize the stage whose repair most improves final yield.
• Holonic Autonomy Nesting: Design nested units as autonomous local wholes and dependent parts at the same time, with explicit boundaries, interfaces, escalation paths, and cross-level invariants.
• Interleaved Discrimination Practice: Mix related practice targets in a deliberate sequence so the learner must choose, recall, classify, or perform under discrimination pressure, improving durable retention and transfer beyond blocked fluency.
• Nearest-Exemplar Response Reuse: Use the closest remembered or stored case as the model for the present response, while making similarity, adaptation, confidence, and exception boundaries explicit.
• Perspective Depth Projection Design: Fix the observer and projection relation, construct depth through convergence, scale, foreshortening, overlap, and atmosphere, and disclose where the chosen viewpoint distorts or hides spatial truth.

▸ Show 5 more

Notes

This is the long-orphaned umbrella the project flagged in R9 when comparative_method and experimental_design had no clean parent, because comparison and controlled_comparison were both absent from the catalog. ChatGPT Pro's R16 pass independently surfaced the same gap with the same slug, an unusual convergent signal that lent weight to the case for promoting comparison from candidate to accepted prime. Once comparison is in place, contrast becomes a presupposes child (R17a retype, formerly subsumption — contrast is comparison oriented toward difference, not a sibling). simile → comparison and analogy → comparison are decompose edges (each specifies the alignment rule and output-relation type). The experimental_design → comparison decompose edge captures R9's structural-core observation: controlled comparison is the inferential heart of experimental design. value_commensuration → comparison is presupposes, because value commensuration requires that a comparison along a value dimension be performable.

Comparison's substrate-furthest cases — CPT-symmetry tests in particle physics, controlled experimental comparison in biology — are deliberately spotlighted in the Broad Use and Examples sections because they answer the most common skeptical move against comparison as a prime ("isn't this just a human-cognitive thing?"). The particle-physics case in particular runs the comparison machinery entirely without human social or institutional scaffolding, which is the cleanest available evidence for substrate independence.

The frame-relativity feature is the single most under-appreciated aspect of the prime. In ordinary speech, comparison outputs are routinely treated as frame-free properties of the comparands ("these two are similar," "this one is better"), when they are in fact joint outputs of the comparands, the frame, the dimensions, and the alignment rule. Naming the frame as a load-bearing role-slot — and treating "compared to what, along what dimensions, under what alignment?" as a routine prompt — is the practical reasoning leverage the prime supplies.

A frequent confusion to watch for in curation is the operation/result conflation. New candidate primes presented as "kinds of relation" (sameness, difference, ranking, fit, equivalence) are almost always comparison outputs rather than separate operations; the right test is whether the candidate can be reached by specifying the frame, dimension set, alignment rule, and output-relation reading on the comparison operation — if so, it is a child, not a sibling.

References

[1] Keith J. Holyoak and Paul Thagard. "Analogical Mapping by Constraint Satisfaction". Cognitive Science, 13(3) (1989), 295–355. Theory of analogical mapping driven by interacting structural, semantic, and pragmatic constraints (the ACME model) — supplies a formal model of the role-slots (comparands, alignment rule, output relation) the prime names; supports the constraint-satisfaction characterization. ↩

[2] Amos Tversky. "Features of Similarity". Psychological Review, 84(4) (1977), 327–352. Feature-contrast model: similarity is the output of a weighted set-theoretic matching of common and distinctive features, parameterized by direction (asymmetry) and salience — directly supports the claim that similarity is the output of a comparison process over weighted feature sets. ↩

[3] Dedre Gentner. "Structure-Mapping: A Theoretical Framework for Analogy". Cognitive Science, 7(2) (1983), 155–170. Structure-mapping theory: the alignment rule preserves higher-order relational structure and drops surface features — supports both the Structural-Signature separation of comparands from relational structure and the analogy-alignment-rule claim. ↩

[4] Nelson Goodman. "Seven Strictures on Similarity". In Problems and Projects (pp. 437–446). Indianapolis: Bobbs-Merrill, 1972. Argues similarity is empty without an explicit respect-of-comparison — supports the claim that dimension-selection is a load-bearing, irreducible role in comparison. (Minor: pages are 437–446, not 437–447.) ↩

[5] Joint Committee for Guides in Metrology. International Vocabulary of Metrology — Basic and General Concepts and Associated Terms (VIM) (3^rd ed., JCGM 200:2012). Sèvres: BIPM, 2012. Characterizes measurement as comparison of a measurand to a reference quantity via a calibration chain anchored in a primary standard — supports the claim that measurement is comparison specialized to numerical output along a metric dimension. ↩

[6] Arthur B. Markman and Dedre Gentner. "Commonalities and Differences in Similarity Comparisons". Memory & Cognition, 24(2) (1996), 235–249. Shows that difference judgments (alignable differences / contrast) are produced by the same structural-alignment process that yields commonalities, and that side-by-side placement prompts the viewer to construct an alignment — supports the operation/reading distinction and the comparison → contrast direction. ↩

[7] Donald T. Campbell and Julian C. Stanley. Experimental and Quasi-Experimental Designs for Research. Chicago: Rand McNally, 1963 (also issued as a Houghton Mifflin booklet). Canonical enumeration of internal-validity threats (history, maturation, testing, instrumentation, regression, selection, mortality, interaction) — supports the claim that these are the failure modes of the alignment rule in controlled comparisons. ↩

[8] Douglas Hofstadter and Emmanuel Sander. Surfaces and Essences: Analogy as the Fuel and Fire of Thinking. New York: Basic Books, 2013. Argues that the same cross-domain mapping structure underlies word use, scientific conceptualization, and everyday categorization — supports the claim that even rhetorical comparisons run on the same dimension-selection and alignment machinery as scientific ones. ↩

[9] Douglas L. Medin, Robert L. Goldstone, and Dedre Gentner. "Respects for Similarity". Psychological Review, 100(2) (1993), 254–278. Argues any usable similarity construct requires explicit "respects" (dimensions of comparison) supplied by the comparison process itself — the canonical articulation of dimension-selection as a structural role in comparison. ↩

[10] Giovanni Sartori. "Comparing and Miscomparing". Journal of Theoretical Politics, 3(3) (1991), 243–257. Identifies "cat-dog" miscomparisons — aggregates assembled under an alignment rule that conceals genuine non-equivalence — as the central failure mode, with alignment-rule audit as the discipline — supports the Manages-Complexity alignment-rule-audit claim. ↩

[11] Arend Lijphart. "Comparative Politics and the Comparative Method". American Political Science Review, 65(3) (1971), 682–693. Systematic comparison of the experimental, statistical, comparative, and case-study methods, showing the same role-slots underlie all of them, differing in how each implements the alignment rule — supports the Knowledge-Transfer claim about shared role-slots across traditions. ↩

[12] Christian Smorra, S. Sellner, M. J. Borchert, J. A. Harrington, T. Higuchi, H. Nagahama, T. Tanaka, A. Mooser, G. Schneider, M. Bohman, K. Blaum, Y. Matsuda, C. Ospelkaus, W. Quint, J. Walz, Y. Yamazaki, and S. Ulmer. "A parts-per-billion measurement of the antiproton magnetic moment". Nature, 550(7676) (2017), 371–374. BASE-collaboration (CERN) measurement reporting the antiproton magnetic moment as −2.7928473441(42) nuclear magnetons, agreeing with the proton at the parts-per-billion level — supports the Formal example's CPT-symmetry comparison at a substrate maximally far from human practice. ↩