Observational Equivalence¶
Core Idea¶
Observational equivalence is a relation among distinct underlying candidates that cannot be distinguished by any observation admitted under a declared observation regime. The candidates may be scientific theories, statistical models, parameter values, programs, or processes. They need not share an internal mechanism, representation, interpretation, or ontology. What they share is the complete observable profile made available by the regime: every admissible probe returns the same prediction, probability distribution, value, trace, termination result, or other registered consequence. The prime therefore separates sameness for an observer under stated access conditions from sameness in all respects.[1][2][3]
The observation regime is constitutive, not background decoration. It determines which probes count, which outputs are visible, which differences count as distinguishable, and whether exact equality or a declared coarser criterion controls the result. Write the regime as R, and let O_r(x) be what probe r ∈ R reveals about candidate x. Then x and y are observationally equivalent relative to R when O_r(x) = O_r(y) for every admitted r. Enlarge the regime with a probe that yields unequal results and the equivalence collapses; restrict it and previously distinct candidates may merge into the same class.[1][3]
This gives the abstraction a sharp invariant: observational indistinguishability under a fixed regime despite possible internal difference. The invariant is stronger than a record of past non-detection. It is a universal claim over the observations the regime admits. It is also weaker than identity: two candidates can occupy the same observational class while differing profoundly behind the observation boundary. The structure is useful precisely because it makes both statements true at once—there is a justified level at which the candidates may be substituted, and there remain other levels at which they must not be conflated.
The prime also exposes the two ways an equivalence judgment can fail. A separating observation shows that two candidates assigned to one class yield different visible results. A regime error shows that the analyst silently omitted an observation that the governing inquiry should have admitted, used different observation rules for different candidates, or changed the equality criterion mid-comparison. The first revises the class from within the declared frame; the second reveals that the frame itself was misdeclared.
How would you explain it like I'm…
Can't-Tell-Them-Apart Boxes
Can't-Tell-Them-Apart Test
Indistinguishable Under Allowed Tests
Structural Signature¶
Observational equivalence has the structural form candidate plurality → shared observation regime → equality of complete observable profiles → licensed substitution within the regime. It begins with at least two possibly different underlying candidates. A common interface, experiment family, data-generating view, or semantic context exposes selected consequences while hiding others. Equality across every admitted exposure induces classes of candidates that the observer cannot separate. Because equality of profiles is reflexive, symmetric, and transitive, an exact fixed observation map partitions the candidate space into equivalence classes.[4][5]
Sig role-phrases:
- Carrier space: the set of theories, models, parameterizations, terms, processes, or other candidates being compared.
- Candidate alternatives: at least two members whose internal descriptions may differ.
- Observation regime: the declared set of admissible probes, contexts, variables, or tests and their access limits.
- Observation map: the rule sending each candidate to its complete profile of visible results under that regime.
- Indistinguishability invariant: equality of those profiles for every admitted observation.
- Separation criterion: the counterfactual rule that any admitted unequal result refutes the relation, while a probe added by a richer regime tests a newly refined relation.
The first five roles constitute the relation. With only one candidate, there is no relational equivalence to establish. Without a common regime, the output profiles are not comparable. Without universal equality over admitted observations, there is only partial agreement. Without room for underlying difference, the statement reduces to ordinary identity. The separation criterion is diagnostic rather than an additional required witness: equivalence can hold even when no actual separator has been found or supplied, but any unequal result from an admitted probe defeats it, and any probe outside the regime belongs to a separately declared refinement.
The same signature recurs literally in three unrelated disciplines. Philosophy of science compares rival theories by empirically testable implications. Econometrics compares parameters or structures by the distributions they induce over observable data. Programming-language and process semantics compare terms or processes by what valid contexts and observations reveal. Each supplies different carriers and probes, but each uses the same relation: multiple internals, one declared observation map, and equality of every visible consequence.[1][2][3]
A useful notation is x ~_R y, with the subscript retained whenever ambiguity is possible. It prevents a common error: treating equivalence under one measurement design, interface, or calculus as though it survived every richer observer. If R1 ⊆ R2, equivalence under richer R2 entails equivalence under poorer R1, but not conversely.
What It Is Not¶
Observational equivalence is not mere similarity. Two models whose predictions are close on average, two programs that usually return the same value, or two theories that agree on a familiar benchmark are not thereby observationally equivalent. Exact observational equivalence requires the declared equality criterion to hold across the entire admitted regime. A domain may define an approximate, probabilistic, or tolerance-indexed variant, but then the tolerance and aggregation rule belong in the regime and must not be hidden behind the unqualified term.
It is not the same as “no difference has been observed so far.” A finite record can fail to expose a distinction even when the regime contains a probe that would do so. Silence in the current evidence is an epistemic condition of the investigator; observational equivalence is a relational claim about the candidates and the full set of admitted observations. The former can result from weak data, low power, or an untried context. The latter survives every probe that the regime makes legitimate.[1][2][3]
It does not establish internal identity, explanatory identity, unrestricted semantic identity, or sameness of a causal structure that lies outside the declared observation regime. Two scientific theories may attach different entities or mechanisms to the same testable predictions. Two structural economic models may interpret their parameters differently while generating the same observable distribution. Two program terms may have different syntax or evaluation paths while no admitted context distinguishes their result. Observational equivalence authorizes substitution only for questions that factor through the declared observation map; it says nothing about questions that inspect the hidden structure directly.[1][2][3]
It is not a verdict that both candidates are true, adequate, desirable, or equally easy to use. An observer may be unable to distinguish them empirically yet prefer one for simplicity, computational cost, explanatory vocabulary, compatibility with background commitments, or other criteria outside the observation regime. Those criteria may be legitimate, but they are additional selection rules rather than observations already contained in the equivalence claim.
It is not underdetermination in its broadest sense. Evidence may underdetermine an explanation because the evidence is incomplete, because auxiliary assumptions vary, or because several accounts receive comparable but not identical support. Observational equivalence names the sharper case in which alternatives have the same observable implications throughout the declared regime. It supplies one rigorous route to underdetermination without exhausting the wider idea.
Nor is it confined to behavioral equivalence. Behavioral equivalence is a family of computational relations in which the admitted observations are behaviors—traces, interactions, termination, values, or responses in contexts. Such relations instantiate observational equivalence once the behavioral interface is fixed. The prime is broader because scientific predictions and observable data distributions can supply the observation profile even when “behavior” is not the natural description of the carrier.
Broad Use¶
Philosophy of science. Rival theories are observationally equivalent when all of their empirically testable predictions coincide under the relevant evidential regime. Theories may describe unobservable structure differently yet leave an empirical investigator with no admitted observation that selects between them. The abstraction locates the exact source of the tie: it is not merely that both have survived testing, but that their full testable consequence sets match within the stated scope. This permits careful discussion of theoretical difference without pretending that current empirical access can decide it.[1]
The regime qualifier is especially important here. “All empirically testable predictions” is never meaningful without assumptions about what counts as a test, the conditions under which predictions are derived, and the observational vocabulary used to compare results. A change in auxiliary setup, measurement capability, or the kinds of intervention regarded as admissible can refine the observation regime and split a former equivalence class. The prime therefore makes empirical equivalence a claim with an explicit boundary, not a timeless assertion that two theories can never differ.
Econometrics and statistical identification. Distinct parameter values or model structures are observationally equivalent when they induce the same probability distribution over the observable data. In that setting, the observation map sends an internal specification to an observable distribution. If two specifications have the same image, repeated samples from that same design can improve estimation of the common distribution but cannot identify which internal specification generated it. The problem is structural rather than a shortage that disappears merely by collecting more of the same kind of data.[2][6]
This use makes the relationship to identification concrete. The observational-equivalence classes are the fibers of the map from internal specification to observable distribution. Point identification requires the relevant fiber to contain one candidate. A fiber containing several candidates is an explicit witness of non-identifiability. A new instrument, restriction, intervention, or measured variable matters when it changes the map so that members of the old fiber acquire different observable profiles.
Programming-language semantics. Two terms are observationally equivalent relative to a calculus when every valid admitted context yields the same observable result for either term. The contexts are not illustrations; they constitute the observation regime. Terms can differ syntactically or internally while remaining interchangeable for all observations the calculus licenses. “Extensional equivalence” has been used for this semantics-local notion, but it is not a catalog-wide synonym: its established scope does not automatically cover the scientific and econometric instances of the prime.[3][7][8]
Across these uses, the abstraction is descriptive rather than prescriptive. It does not demand that investigators ignore internal differences. It says that any decision based solely on the declared observable profile must treat members of one class alike. An investigator may then enrich the regime, add non-observational criteria, or accept the class as the proper unit of reasoning.
Clarity¶
Observational equivalence clarifies four separate questions that are often compressed into “Can we tell these apart?” First, what are the candidates? The alternatives must be typed at the same level: two theories, two parameter values within a model class, two structural models, or two terms in a calculus. Comparing a theory to a single prediction, or a model to a dataset, mixes levels and cannot establish the relation.
Second, who or what observes, and through which interface? “Observable” does not mean visible to an unlimited knower. It means returned by the probes, variables, contexts, or tests admitted by the declared regime. Third, what counts as the same result? Scientific predictions may be compared as propositions or distributions; econometric candidates as distributions of measured variables; programs as values, termination behavior, or interaction traces. Fourth, how broad is the quantifier? Agreement on sampled cases differs from equality for every admitted case. Making all four explicit converts a loose resemblance claim into an auditable equivalence judgment.[2][3][9]
The prime also distinguishes uncertainty about membership from graded membership. If the analyst lacks enough evidence to show whether every admitted observation agrees, the correct state is equivalence unresolved, not “partly equivalent.” A domain may deliberately define approximate equivalence, but approximation then becomes a different relation with a tolerance, metric, and aggregation rule. Keeping exact relation, approximate variant, and evidential uncertainty separate prevents measurement limitations from altering the identity of the abstraction.
Clear use requires a negative condition as well as a positive one. The positive condition is equality of complete observable profiles. The negative condition is a separating witness: name an observation admitted by the regime on which the candidates differ. This makes disputes productive. If one party proposes a witness, the question becomes whether the probe is admissible and whether its outputs truly differ, rather than whether the candidates “feel like” the same account.
Manages Complexity¶
The prime manages complexity by replacing a potentially large set of internal candidates with a quotient organized by observable profiles. Instead of analyzing every parameterization, model, or term separately for a question that uses only admitted observations, the analyst can reason once about the class. Each class retains exactly the information the regime exposes and deliberately forgets distinctions the regime cannot use. This compression can transform a sprawling model-comparison problem into a smaller problem over empirically or semantically distinguishable possibilities.[4][5][6][10]
The compression is principled because its loss function is explicit. Hidden differences are not declared unreal; they are declared unavailable to the present observation map. A conclusion is safe at class level only when it depends solely on the observable profile. If it depends on causal mechanism, internal resource use, explanatory interpretation, or another feature outside the regime, the quotient has discarded information the conclusion needs. Observational equivalence therefore pairs compression with an obligation to state which downstream questions remain well-posed after compression.
It also localizes the source of a failed comparison. If two candidates remain in one class, there are only a few structural possibilities: the regime is too coarse for the intended distinction, the candidates genuinely make the same observable commitments within scope, or the equality criterion has erased a difference that matters. This is more actionable than a generic claim that “the data are inconclusive.” The analyst can inspect the carrier set, observation design, and equality rule separately.
Abstract Reasoning¶
The first reasoning move is partitioning. Define a common observation map and group candidates that have equal images. Exact observational equivalence then inherits the ordinary properties of equality: every candidate is equivalent to itself, the relation is symmetric, and candidates equivalent to the same profile are equivalent to each other. This supports substitution within any statement that depends only on the profile. It does not support substitution in statements about hidden structure.
The second move is regime refinement. When R1 ⊆ R2, the richer regime cannot merge two classes distinguished by the poorer one, because it retains all old probes. It may leave classes unchanged or split them using new probes. Conversely, restricting observations can only preserve or merge existing classes. This monotonic refinement law supplies a clean account of how new instruments, measured variables, experimental interventions, or semantic contexts change what can be distinguished.[3][1]
The third move is separation. To disprove x ~_R y, it is enough to exhibit one admitted r with O_r(x) ≠ O_r(y). To establish the equivalence, by contrast, the analyst must justify equality across the entire admitted regime. This asymmetry explains why counterexamples are diagnostically powerful and why positive evidence from a handful of tests may be insufficient. In a finite formal regime, exhaustive comparison may be possible. In an open or infinite regime, a proof, theorem, or domain-valid reduction may be needed rather than enumeration.
The fourth move is quotient-relative inference. If a function f assigns the same value to every member of each observational class, then f is well-defined on the quotient. If it assigns different values to observationally equivalent members, it relies on hidden information and cannot be recovered from the declared observations alone. This provides a direct audit of a proposed conclusion: ask whether it is constant over the class.
The fifth move is duality of aims. Some inquiries seek to break equivalence because they want unique recovery or discriminating tests. Others preserve equivalence because substitution, abstraction, or implementation freedom is valuable. The structure itself is neutral. It identifies the available class; the practical goal determines whether to refine the observer, choose a representative, or deliberately maintain the boundary.
Knowledge Transfer¶
Observational equivalence transfers by mapping four roles rather than borrowing surface vocabulary: candidates, observation regime, complete observable profile, and separating witness. Once these are named, results from one domain improve reasoning in another without turning the domains into metaphors. Philosophy of science emphasizes that empirical equality does not erase theoretical difference. Econometrics makes the observation map concrete as a map to probability distributions and shows why more data from an unchanged design may not break a structural tie. Programming-language semantics makes the universal quantifier over contexts explicit and treats substitutability as relative to what contexts can observe.[1][2][3][11]
From semantics to empirical modeling comes the idea of an observation contract: before declaring equivalence, specify the interface through which comparison occurs. From econometrics to philosophy comes the discipline of distinguishing the distribution of observables from the internal parameterization that generates it. From philosophy to semantics comes the warning that indistinguishable consequences do not license claims of shared meaning or mechanism beyond the interface.
Transfer fails when the observation boundary is smuggled across domains without translation. “Observable” in a scientific experiment, a statistical dataset, and a programming calculus is not the same concrete thing. What transfers is the structural role, not a universal list of probes. A sound transfer therefore ends by restating admissibility and equality in the receiving domain’s own rules.
Examples¶
Formal/abstract¶
Consider a carrier set X = {a, b, c, d} and a regime with two admitted probes, p and q. Suppose the complete profiles are O(a) = (0,1), O(b) = (0,1), O(c) = (1,1), and O(d) = (1,0). Under this regime, a ~_R b, while c and d each form separate classes. The internal construction of a and b may differ, but no conclusion computed only from p and q can distinguish them. Add a third probe s with O_s(a) = 0 and O_s(b) = 1, and the old class splits. Remove q, and other classes might merge. The example makes regime relativity and refinement visible without any domain assumptions.[4][5]
Now map that form into program semantics. Let M and N be distinct terms in a fixed calculus. The regime is the set of valid contexts C[·] together with a declared observable result, such as the value returned when the composed term is valid. If every admitted context gives the same observable result for C[M] and C[N], the terms are observationally equivalent in that calculus. Their syntax and evaluation paths may differ. If a newly admitted context returns different results, the richer regime separates them. The equivalence is thus neither textual identity nor an informal claim that the programs “behave similarly”; it is equality of the complete contextual profile.
Mapped back: In both constructions, the carrier contains distinct candidates, the regime specifies every legitimate probe, the observation map assigns a complete visible profile, and profile equality induces a class. The class persists through internal variation but collapses under one admitted separating probe. All reasoning at class level is valid only for properties determined by the profile.
Applied/industry¶
Suppose an empirical team compares two structural economic models, (A) and (B). The models assign different meanings or values to hidden parameters, yet under the team’s study design they induce the same joint probability distribution for every measured variable. The team can estimate that common observable distribution more precisely by enlarging the sample, but an arbitrarily precise estimate of the same distribution still does not reveal whether (A) or (B) supplied the hidden structure. For the original design, the models occupy one observational-equivalence class.[2][6][3]
The team then proposes a new measured variable or instrument. This is not automatically a solution. It matters only if adding it changes the observation map so that (A) and (B) induce different enlarged distributions. If it does, the new design supplies a separating observation and splits the class. If both models still yield the same enlarged profile, the equivalence survives. The procedure distinguishes three outcomes often blurred in practice: improved precision about a shared observable profile, a genuine refinement that separates internal candidates, and a merely decorative addition that collects more information without discriminating the models.
A philosophy-of-science version has the same architecture. Two theories may use different unobservable mechanisms while yielding the same testable predictions under the available experiment family. Repeating those experiments can test the shared prediction and might refute both theories, but cannot select one over the other. A newly admissible experiment matters only if the theories predict different outcomes for it. The example shows why observational equivalence and falsifiability can coexist: one observation might defeat the shared profile while no observation in the current regime chooses between its rival generators.
Mapped back: The models or theories are the candidates; the study design or experiment family is the regime; the joint distribution or prediction set is the observable profile; equality of that complete profile is the invariant; and a newly admitted variable, instrument, or experiment is a collapse witness only when it produces unequal results. More of the same observation sharpens the common profile but does not by itself identify the hidden member.
Structural Tensions¶
T1: Explanatory difference versus observable sameness. Observational equivalence permits theories or models with different internal stories to occupy one empirical class. That is the abstraction’s point, but it creates pressure in both directions. Treating the candidates as wholly different ignores their warranted substitutability for observable questions; treating them as wholly identical erases differences in mechanism, interpretation, or ontology. The disciplined position is indexed: equivalent for claims that factor through the declared observation map, potentially different for claims that inspect what the map hides. Maintaining that index becomes harder when one candidate’s internal vocabulary is more intuitive or institutionally familiar.
T2: Regime choice versus neutrality. An equivalence class can look like an objective feature of the candidates even though its boundary depends on which observations were admitted. A deliberately narrow regime may manufacture sameness by excluding inconvenient probes; an unrealistically rich regime may demand observations unavailable to the actual inquiry. The analyst must therefore justify the observer as well as the candidates. Yet every justification depends on purpose: prediction, explanation, control, implementation substitution, or estimation can legitimately admit different observations. The tension is not resolved by choosing the largest imaginable regime, but by choosing and declaring the regime appropriate to the claim.
T3: Exact relation versus finite and noisy evidence. The abstraction is cleanest when observable profiles can be proved equal. Empirical work rarely supplies that vantage point: samples are finite, measurements are noisy, and many possible probes remain untried. Declaring exact equivalence from failure to reject a difference overstates the evidence; refusing to use any equivalence until omniscience is achieved makes the concept inert. Domains address the gap with proofs, identification analyses, equivalence tests, tolerances, or restricted regimes. Each remedy must say whether it establishes exact equality, an approximate variant, or only an unresolved judgment.
T4: Compression versus hidden obligations. Quotienting by observable profile can dramatically simplify inference, but the simplification deletes internal distinctions. That deletion is safe only for questions invariant across the class. A representative chosen for convenience may be cheaper to compute, easier to explain, or more compatible with surrounding theory; those advantages tempt users to project its hidden properties onto the entire class. The very success of compression can conceal its boundary. Every class-level conclusion therefore carries an obligation: show that changing the representative while preserving the observable profile would leave the conclusion unchanged.
T5: Stability versus refinement. Observational equivalence is stable under repetition of the same regime but potentially fragile under a richer one. This makes it simultaneously durable and provisional. Engineers or theorists need stable substitution rules to build on, while investigators seek new probes precisely to split classes. A relation can be rigorously established relative to today’s interface and still disappear after an extension. The tension is managed by versioning the regime and treating equivalence judgments as monotone only in the correct direction: richer observers refine or preserve classes; poorer observers preserve or merge them.
T6: Discrimination versus preservation. In identification and theory testing, a non-singleton class often appears as a problem to solve. In semantics and modular reasoning, the same structure can be valuable because it licenses implementation replacement without observable disruption. An intervention that breaks equivalence may therefore be progress for inference but a regression for abstraction or compatibility. The relation itself carries no preferred direction. The investigator must state whether the goal is to find a separating observation, reason safely over the quotient, or protect indistinguishability against observers outside a declared interface.
Structural–Framed Character¶
Observational equivalence sits at the structural end of the structural–framed spectrum: it names equality of complete observable profiles for distinct candidates under a declared regime. Internal mechanisms may differ, yet no admitted probe separates the candidates; one separating result splits the class.
Its vocabulary of candidates, observation maps, profiles, and separating probes states neutral relational roles rather than importing a philosophy-specific lexicon. The relation carries no built-in verdict, and its philosophy-of-science history supplies neither an institutional nor a normative referent. It can be defined without human practices: two econometric models may induce the same data distribution, just as two program terms may return the same result in every admitted context. Applying it recognizes profile equality already present under the chosen interface rather than imposing an interpretive perspective. On every diagnostic, it reads structural.
Substrate Independence¶
Observational equivalence is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its signature—distinct candidates mapped through one declared observation regime to equal complete profiles—is stated in fully relational terms, earning maximal structural abstraction. The same structure appears literally when scientific theories share all testable implications, econometric models induce the same observable distribution, and program terms agree in every admitted context, giving it strong domain breadth and transfer evidence. It remains below the universal tier because its demonstrated uses cluster in formal, inferential, and semantic fields rather than spanning physical, biological, cognitive, and social substrates alike.
- Composite substrate independence — 4 / 5
- Domain breadth — 4 / 5
- Structural abstraction — 5 / 5
- Transfer evidence — 4 / 5
Relationships to Other Abstractions¶
Current abstraction Observational Equivalence Prime
Parents (1) — more general patterns this builds on
-
Observational Equivalence is a kind of Equivalence Relation Prime
Observational Equivalence is the Equivalence Relation induced by equality of complete observable profiles under one declared observation regime.Equivalence Relation supplies the carrier, reflexive-symmetric-transitive relation, equivalence classes, and quotient. Observational Equivalence specializes the criterion to equality of complete observable profiles under one declared observation regime and adds the observation map, regime relativity, and substitution limited to profile-invariant questions. Removing those additions leaves the complete parent intact, while every unqualified instance of the child retains all parent axioms.
Hierarchy path (1) — routes to 1 parentless root
- Observational Equivalence → Equivalence Relation
Neighborhood in Abstraction Space¶
Observational Equivalence sits among the more crowded primes in the catalog (11th 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 — Unclustered & Miscellaneous (481 primes)
Nearest neighbors
- Proof of impossibility — 0.76
- Argument from incredulity — 0.75
- Idea — 0.75
- Appeal to novelty — 0.75
- Causal reasoning — 0.75
Computed from structural-signature embeddings · 2026-10-08
Distinction from Neighbors¶
Falsifiability is a close catalog neighbor, but the two primes ask different questions about observations. Falsifiability concerns a single claim or theory and asks whether it excludes at least one possible observation whose occurrence would refute it. Observational equivalence concerns two or more candidates and asks whether their complete profiles are equal under a declared observation regime. A pair of theories can be observationally equivalent and strongly falsifiable: both forbid the same risky outcome, so that outcome could refute both, while no admitted outcome selects one over the other. Conversely, two non-equivalent candidates may be difficult to falsify if neither risks a clear forbidden outcome. Falsifiability is about possible defeat; observational equivalence is about possible discrimination.
Identifiability is tightly connected but not synonymous. Identifiability is a uniqueness property of a target relative to a model and observation channel: can the internal unknown be uniquely recovered from what the channel exposes? Observational equivalence is the relation that groups internal alternatives with the same exposed profile. A class with more than one member witnesses non-identifiability, but the roles differ. Equivalence describes the fibers of the observation map; identifiability asks whether the relevant fiber is a singleton. The distinction matters operationally because one can study, compare, or quotient observational classes without having selected a particular internal target for recovery.
Observability asks whether a system’s internal state can be reconstructed from its external outputs, commonly with the model or transition rules taken as given. Observational equivalence instead compares whole candidate systems, models, parameters, or terms by their visible consequences. Poor observability can create observationally equivalent states, but observational equivalence is not restricted to state reconstruction and does not assume a known model. In control language, observability asks whether hidden states map uniquely to output histories; observational equivalence names the equality relation among anything that maps to the same history. One is a recoverability property, the other the induced relation.
Comparison supplies the broader operation: place candidates in a shared frame, choose dimensions, align them, and read off a relation. Observational equivalence is one highly constrained output of comparison. The frame is an observation regime, the dimensions are every admitted observable consequence, and the output is equality of the complete profiles. An ordinary comparison may reveal ranking, contrast, partial similarity, or difference along selected dimensions. Observational equivalence requires equality across the whole declared interface. Comparison can occur with one observed difference; observational equivalence collapses as soon as that difference is admitted.
Two further conceptual boundaries are useful even though they need not be DAG neighbors. Broad underdetermination includes many ways evidence can fail to decide among accounts; observational equivalence is its exact profile-equality case. Behavioral equivalence is a domain family in which behavior defines the observation regime; it is a specialization, not the catalog-wide prime. Keeping these boundaries separate prevents the prime from absorbing either a broader epistemic condition or a narrower semantic implementation.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.
Notes¶
“Extensional equivalence” is established terminology for a semantics-local family of observational relations. It should not be widened into a catalog-wide alias for this prime without evidence that the term also names the scientific and econometric instances rather than only their shared structural analogue.
References¶
[1] Kyle Stanford. "Underdetermination of Scientific Theory". The Stanford Encyclopedia of Philosophy, Summer 2023 Edition, 2023. Supports empirical equivalence relative to an evidential vocabulary, observational capabilities, and auxiliary assumptions. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h
[2] Thomas J. Rothenberg. "Identification in Parametric Models". Econometrica 39(3), 1971. Defines observationally equivalent structures by exact equality of their observable probability distributions. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h
[3] Dan R. Ghica, Koko Muroya, and Todd Waugh Ambridge. "A robust graph-based approach to observational equivalence". Logical Methods in Computer Science 21(2), 2025. Supports contextual equivalence over admitted program contexts and refinement by richer observers. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j
[4] David I. Spivak. Category Theory for Scientists. Massachusetts Institute of Technology OpenCourseWare, 2013. Supports kernel equivalence, fibers, partitions, and quotients for an exact fixed map. registry ↩a ↩b ↩c
[5] Lean mathematical library. "Mathlib.Data.Setoid.Basic". Official Mathlib documentation. Formally corroborates that quotienting by a function’s kernel yields an object equivalent to its range. registry ↩a ↩b ↩c
[6] Raffaella Giacomini and Toru Kitagawa. "Robust Bayesian Inference for Set-Identified Models". Econometrica 89(4), 2021. Supports identified sets of observationally equivalent structural parameters and the limit of sample learning within a fixed observable law. registry ↩a ↩b ↩c
[7] James Hiram Morris Jr. Lambda-Calculus Models of Programming Languages. PhD thesis, Massachusetts Institute of Technology, submitted 1968. Supports the historical semantics-local lineage of extensional program equivalence; some later bibliographies use 1969. registry ↩
[8] Flavien Breuvart, Giulio Manzonetto, Andrew Polonsky, and Domenico Ruoppolo. "New Results on Morris’s Observational Theory: The Benefits of Separating the Inseparable". LIPIcs — FSCD 2016, vol. 52, article 15, 2016. Corroborates Morris’s extensional theory of contextual equivalence as semantics-local terminology. registry ↩
[9] Unverified encyclopedia synthesis. The cited domain sources support components of the four-part audit, but no validated external source establishes this exact checklist as a named or standard method. ↩
[10] Partially supported encyclopedia synthesis. Kernel quotients and observable images are externally supported; describing the construction as principled complexity compression with a general preservation guarantee is not directly source-verified. ↩
[11] Unverified encyclopedia synthesis. The four roles can be reconstructed from the cited domain examples, but the instruction to transfer the abstraction by mapping exactly those roles is not an externally established result. ↩