Congruence Bias¶

The reasoning bias of designing only tests that could confirm a favored hypothesis rather than tests that discriminate it from equally plausible rivals — a failure of test design, not evidence evaluation, that leaves a probe locally valid but globally non-diagnostic because its likelihood ratio sits near one.

Core Idea¶

Congruence bias is the tendency, when testing a hypothesis, to design only tests that could confirm it rather than tests that could discriminate between it and equally plausible rival hypotheses. The bias operates at the level of experimental strategy — not at the level of how evidence is evaluated once obtained — so even a scrupulously honest evaluator can be misled by it: if the test chosen was constitutionally incapable of returning a different result under the alternative hypothesis, a positive finding advances nothing. The classic demonstration is Wason's 2-4-6 task (1960): given three numbers that fit the unstated rule "any ascending sequence," subjects typically guess a more specific rule such as "even numbers increasing by 2" and then propose only confirming sequences (8-10-12, 14-16-18) rather than disconfirming probes such as 5-7-9 that would leave the true rule intact while falsifying the favored one. The mechanism is that hypothesis-confirmation is cognitively the most direct path: the reasoner simulates what the world would look like if H were true, then checks whether that prediction is satisfied. Generating the signature that a competing hypothesis H2 would produce, and designing a test whose outcome distinguishes P(data|H1) from P(data|H2), requires holding two hypotheses in working memory simultaneously and orienting toward a result one has not anticipated — a more effortful operation that is systematically underperformed. The result is a test that is locally valid (it did confirm H1) but globally non-diagnostic (it would have confirmed H1 equally well whether H1 or H2 were true), and the reasoner receives no useful information about which is the case.

Structural Signature¶

Sig role-phrases:

the favored hypothesis — a single hypothesis H1 the investigator holds and sets out to test
the unenumerated rivals — one or more equally plausible alternatives H2, H3 that also explain the data but are left unstated
the test-design stage — the choice of which probe to run (not the reading of its outcome), the level at which the bias operates
the easy confirmation path — simulating what the world looks like if H1 is true and checking that prediction, cognitively cheaper than holding two hypotheses and orienting toward an unanticipated result
the non-discriminating probe — a test whose expected outcome under H1 is indistinguishable from its outcome under a rival, so the likelihood ratio P(data|H1)/P(data|H2) sits near one
the locally-valid-but-empty pass — a clean confirmation of H1 that would have confirmed H1 equally whether H1 or H2 held, yielding no information about which is true
the across-the-body fingerprint — a run of confirmations with no probe that risked a different signature, detectable only by what was never asked
the discrimination remedy — enumerate the strongest rival and design a severe/strong-inference test that could fail (the 2-4-6 disconfirming probe), moving the likelihood ratio away from one

What It Is Not¶

Not confirmation bias. Confirmation bias is bias of evaluation — interpreting an already-diagnostic test to favor the hypothesis; congruence bias is bias of design — choosing a probe that could not have come out differently were a rival true. The two have structurally different remedies: blinding and pre-registration fix the first and do nothing for the second, which yields only to enumerating rivals and building tests that discriminate.
Not innocuous because the test was clean. A locally valid confirmation — honest execution, an unimpeachable positive result — can be globally non-diagnostic if the probe's likelihood ratio P(data|H1)/P(data|H2) sits near one. A test that passed but would have passed equally under a rival yields no information about which hypothesis is true, so a clean confirmation is reread as a warning sign, not as evidence.
Not a product of motivated or dishonest reasoning. The defect survives perfect scrupulousness: confirming is simply the cheaper cognitive path (simulate the world if H1 holds and check the prediction), while generating a rival's signature requires holding two hypotheses in working memory and orienting toward an unanticipated result. The bias misleads exactly the honest evaluator, which is why protesting motivated reading misses it.
Not fixed by more confirmations or by replication. A run of confirmations of a single hypothesis, with no probe that risked a different signature, is the fingerprint of the bias, not a strengthening of the case — and it is detectable only across the body of work, by what was never asked. Repeating a non-diagnostic test more carefully or more often cannot repair its inability to separate the rivals.
Not a property of non-human or automated systems. A confounded protocol or a poorly designed test suite is non-diagnostic but not biased — there is no reasoner whose cognition produces the failure, so invoking "congruence bias" for them is a category error. What travels to such systems is the corrective parent — diagnosticity / likelihood-ratio reasoning — not the named bias, which is a deficiency of a human inquirer.

Scope of Application¶

Congruence bias lives in one domain — human hypothesis-testing cognition — and the contexts below are application settings of that single substrate (a reasoner choosing which test to run), not structurally distinct systems. The corrective parent it motivates (diagnosticity / likelihood-ratio reasoning, severe testing, strong inference) travels everywhere a test's worth is its capacity to come out differently under rival states of the world; but the bias requires a reasoner, so calling a confounded protocol "biased" is a category error and stays out of this map.

Cognitive psychology and judgment research — the canonical home, central to the literature on hypothesis testing and scientific reasoning (Wason's 2-4-6 task).
Scientific methodology — a frequent diagnosis of why "supported" theories later collapse: the original experiments could not have rejected the alternative, motivating strong inference and severe-testing discipline.
Diagnostic medicine — anchoring on a presumed diagnosis and ordering confirming rather than discriminating tests (Croskerry's diagnostic momentum and premature closure), countered by differential diagnosis.
Software debugging — exercising the suspected fault path while never testing the equally plausible alternative module, so a passing test confirms the wrong hypothesis.

Clarity¶

Naming congruence bias splits apart two failures of inquiry that intuition fuses under "confirmation bias." Without the label, a reasoner who has run an honest, clean test and gotten a positive result has no vocabulary for what went wrong — the evaluation was unbiased, the execution unimpeachable, yet the conclusion is unwarranted. The label localizes the defect upstream, in the choice of test rather than the reading of its outcome, and so distinguishes bias-of-evaluation (interpreting evidence to favor H) from bias-of-design (generating evidence via a probe that could not have come out differently were a rival true). That distinction is consequential because the two have structurally different remedies: blinding, pre-registration, and protest against motivated reading address the first and do nothing for the second, which yields only to explicit enumeration of rivals and tests built to discriminate among them.

The sharper question the concept licenses is diagnosticity rather than mere confirmation: not "did the result come out as my hypothesis predicts?" but "would this test have produced a different signature had a competing hypothesis been true?" A practitioner armed with the term stops counting confirmations and starts asking whether each test moves the likelihood ratio between H1 and H2 away from one — recognizing that a probe whose expected outcome is identical under both hypotheses is informationally empty no matter how decisively it "passes." This reframes a successful confirmation from evidence into a warning sign, and makes the strength of a test reside in its capacity to fail, not in the cleanliness with which it succeeds.

Manages Complexity¶

A reasoner faces an open-ended sprawl of ways an inquiry can go wrong — confounded protocols, missing controls, anchored diagnoses, untested fault paths, irreproducible findings — each looking like its own methodological flaw demanding its own fix. Congruence bias collapses a large slice of that sprawl to a single screening parameter applied to any proposed test: the likelihood ratio P(data|H1)/P(data|H2) it would generate. Instead of auditing each study's full design, the analyst asks one diagnostic question of every probe — "would this come out differently were a rival true?" — and reads the worth of the test off whether that ratio departs from one. A clean-but-confounded drug trial, a confirming 2-4-6 sequence, a diagnostic workup that anchors, and a debugging run that exercises only the suspected module all reduce to the same defect (ratio ≈ 1, locally valid, globally non-diagnostic) and admit the same family of remedies (enumerate the rival, design to discriminate). The move replaces a high-dimensional "is this inquiry sound?" with a one-parameter triage that predicts, before any data arrive, which tests can move belief and which cannot.

Abstract Reasoning¶

Congruence bias licenses a tight family of moves, all organized around the likelihood ratio rather than the bare outcome of a test. Diagnostic: confronted with a confirmed hypothesis, the reasoner infers backward to the test that produced it and asks whether that test could have come out otherwise under a rival — if a positive result was guaranteed whether H1 or H2 held, the confirmation is reread as informationally empty and the investigator's confidence as unearned. The surface signature (a clean test that passed) is taken to indicate a hidden defect (a probe with likelihood ratio near one), inverting the naive reading in which passing is good news. A second diagnostic reads the strategy of a body of work: a run of confirmations of a single hypothesis, with no probe that risked a different signature, is the fingerprint of congruence bias even when each individual study is methodologically spotless — the failure is detectable only by looking across the tests, at what was never asked.

Interventionist: the corrective moves are predictions, not just procedures. Enumerate the strongest rival H2 and design a probe whose expected outcome differs under H1 and H2 — the prediction being that such a probe can return a result the reasoner did not anticipate, and that its likelihood ratio departs from one; a test that could fail if H1 is wrong is precisely the one that can move belief if it passes. The 2-4-6 corrective is the template: propose a sequence the favored rule forbids (5-7-9 against "even numbers ascending by 2"), because only an outcome the favored hypothesis cannot accommodate carries discriminating weight. To shift the inquiry's trajectory, then, one changes not how the evidence is read but which test is run — adding a control arm, a discriminating probe, an explicitly enumerated rival — each a wager that the new test will separate hypotheses the old one fused.

Boundary-drawing: the concept fixes its own regime. It bites at the level of test design, not evidence evaluation, so it is the right diagnosis exactly when execution is honest and the result is positive yet the conclusion is unwarranted — and it is the wrong diagnosis where the defect lies in motivated reading of an already-diagnostic test, which is confirmation bias proper and yields to blinding and pre-registration rather than to rival enumeration. Locating a failure on the design side rather than the evaluation side tells the practitioner which remedy can possibly work: no amount of scrupulous re-reading repairs a probe that was constitutionally non-diagnostic, and no amount of rival-enumeration repairs a diagnostic probe that was read with a thumb on the scale. Predictive: before any data arrive, the reasoner can forecast which tests will be capable of moving belief — those whose anticipated signature differs across the live hypotheses — and which will not, triaging a planned program of work into informative and empty probes in advance of running any of them.

Knowledge Transfer¶

Within human hypothesis-testing cognition the bias transfers as mechanism, intact, with the caveat that its "domains" are application contexts of one substrate — a human reasoner choosing which test to run — not structurally distinct systems. With that understood, the diagnosticity screen (does this probe move the likelihood ratio P(data|H1)/P(data|H2) away from one?) and its corrective discipline (enumerate the strongest rival, design to discriminate) carry without translation across scientific methodology (a frequent diagnosis of why "supported" theories later collapse — the original experiments could not have rejected the alternative), diagnostic medicine (anchoring on a presumed diagnosis and ordering confirming rather than discriminating tests — Croskerry's diagnostic momentum and premature closure), and software debugging (exercising the suspected fault path while never testing the equally plausible alternative module). The vocabulary (bias-of-design versus bias-of-evaluation, diagnosticity, severe test, likelihood ratio), the diagnostic (a clean test that passed but could not have come out otherwise under a rival is the fingerprint, detectable only across the body of work by what was never asked), and the interventions (pre-register the rival; strong inference à la Platt; severe tests à la Popper/Mayo; differential diagnosis) all move freely, because the same effortful-to-hold-two-hypotheses cognitive limitation is at work in each.

Beyond the human reasoner the bias does not travel — and the reason is sharp. Congruence bias is, by definition, a deficiency of a reasoner: a poorly designed automated test suite or a confounded experimental protocol is non-diagnostic but not biased, because there is no cognition producing the failure. So invoking "congruence bias" for a non-human system is a category error, not merely a stretched analogy. What genuinely travels — and travels far — is the corrective parent: diagnostic discrimination / likelihood-ratio reasoning (and its normative siblings severe testing and strong inference), the substrate-independent pattern of arranging observations so that the conditional probability of the outcome differs between competing hypotheses. That parent recurs as a true co-instance across medical testing, radar and sonar detection, signal processing, forensics, and scientific experiment design — anywhere a test's worth is its capacity to come out differently under rival states of the world. The honest framing is therefore that congruence bias is the human failure mode whose existence motivates teaching the broader pattern: when the cross-domain lesson is needed, carry diagnosticity / severe testing, not "congruence bias," whose own cargo (the Wason 2-4-6 paradigm, the design-versus-evaluation split, the working-memory account of why confirmation is the easy path, the rival-enumeration remedies aimed at a reasoner) stays bound to human inquiry. It must also be kept distinct from confirmation bias proper, which is bias-of-evaluation (motivated reading of an already-diagnostic test, yielding to blinding and pre-registration) and is the wrong diagnosis where the defect is a non-diagnostic probe. So: as mechanism the bias stays inside human hypothesis testing; the diagnosticity/likelihood-ratio parent travels everywhere as the portable structure; carry that parent, not the named bias (see Structural Core vs. Domain Accent).

Examples¶

Canonical¶

In Peter Wason's 1960 rule-discovery task, subjects were told that the triple 2-4-6 conforms to an unstated rule and asked to find the rule by proposing further triples, receiving only "fits"/"does not fit" feedback and announcing the rule when confident. The true rule was simply "any three numbers in ascending order." Most subjects fixed on a narrower hypothesis — typically "even numbers increasing by two" — and then tested only triples their own rule generated (8-10-12, 20-22-24), each of which duly "fit." A run of such confirmations left them confident yet wrong, because those probes could not distinguish their rule from the broader truth. The move most failed to make was a disconfirming probe such as 5-7-9: a triple their favored rule forbids but the real rule permits, whose "fits" verdict would have exposed the narrower hypothesis as too specific.

Mapped back: "Even numbers ascending by two" is the favored hypothesis and the broader "any ascending sequence" is the unenumerated rival. Generating 8-10-12 at the test-design stage is the easy confirmation path — simulating what the world looks like if the favored rule holds — and because that triple "fits" under both rules it is the non-discriminating probe whose likelihood ratio sits at one, yielding the locally-valid-but-empty pass. Proposing 5-7-9 is the discrimination remedy: a probe the favored rule forbids, so its outcome carries discriminating weight.

Applied / In Practice¶

In diagnostic medicine the bias appears as what Pat Croskerry terms premature closure and diagnostic momentum: a clinician forms an early impression — say, that a patient's shortness of breath is an asthma exacerbation — and then orders tests and asks history questions that would confirm asthma while never running the probe that would separate it from a live rival such as heart failure or pulmonary embolism. A peak-flow reading consistent with asthma "confirms" the working diagnosis but is roughly as consistent with the alternatives, so it discriminates little. The taught corrective is the differential diagnosis: enumerate the plausible rivals up front and, for each, identify the finding or test whose result would differ across them — a D-dimer or CT angiogram that could come back positive for embolism — so the workup is built to separate hypotheses rather than to pile up confirmations of the first one.

Mapped back: The initial "asthma" impression is the favored hypothesis; heart failure and pulmonary embolism are the unenumerated rivals that diagnostic momentum leaves unstated. Ordering the confirming peak-flow at the test-design stage is the non-discriminating probe — a result nearly identical under all three hypotheses, so its likelihood ratio sits near one. The differential diagnosis is the discrimination remedy: it forces enumeration of the strongest rival and selects a probe (the CT angiogram) that could come out differently under it.

Structural Tensions¶

T1: Local validity versus global diagnosticity (a clean pass that carries no information). The unsettling feature of the bias is that a test can be impeccably designed and honestly executed at the level of its own conduct — no confound, no motivated reading — and still be worthless, because the likelihood ratio it generates sits near one. Local validity (it did confirm H1) and global diagnosticity (it would have confirmed H1 equally whether H1 or a rival were true) come apart, and the cleaner and more decisive the confirmation, the more falsely reassuring it is. The cleanliness that ordinarily signals a good result here masks the defect, so a successful confirmation must be reread as a warning sign rather than as evidence — the strength of a test resides in its capacity to fail, not in the polish with which it passes. Diagnostic: Would this clean, passing probe have produced a different signature had the strongest rival been true, or was its positive outcome guaranteed under both?

T2: Design-locus versus evaluation-locus (why mislocating the failure wastes the fix). Congruence bias and confirmation bias present with the same surface — unwarranted confidence flowing from a positive test — but they live at different stages and demand opposite remedies. Congruence bias is a defect of which probe was run; confirmation bias proper is a defect of how an already-diagnostic outcome was read. Blinding, pre-registration, and protest against motivated reading repair the second and do nothing for the first, which yields only to enumerating rivals and building tests that discriminate. The tension is that a practitioner who diagnoses the wrong locus applies a remedy that cannot possibly work: no amount of scrupulous re-reading rescues a probe that was constitutionally non-diagnostic, and no amount of rival-enumeration rescues a diagnostic probe read with a thumb on the scale. Diagnostic: Is the defect in the choice of test that was run, or in the reading of a test that could in principle have come out otherwise?

T3: Cognitive economy versus discriminating cost (the bias is efficient, not irrational). The confirmation path is not a lapse of rigor; it is the cheaper computation. Simulating what the world looks like if H1 holds and checking the prediction is a single-hypothesis operation, while generating the signature a rival would produce and designing a probe that separates them requires holding two hypotheses in working memory at once and orienting toward a result one has not anticipated. The corrective discipline — always enumerate the strongest rival, always design to discriminate — imposes a real and recurring cognitive and resource cost, which is exactly why the bias is systematic rather than occasional. The pull is between the economy that makes confirmation the default and the expense of the discriminating test that alone moves belief. Diagnostic: Is the discriminating probe being skipped because it would be uninformative, or because constructing and running it is more effortful than the confirming one?

T4: Per-study spotlessness versus across-the-body fingerprint (a flaw invisible where you audit). The bias is undetectable at the level at which inquiry is usually reviewed. Each individual study can be methodologically flawless — clean protocol, honest execution, a genuine positive — so a study-by-study audit finds nothing wrong. The failure is legible only across the body of work, in the pattern of confirmations with no probe that ever risked a different signature, in what was never asked. This cuts both ways: the bias evades the per-study scrutiny that catches confounds and motivated reading, yet it is plainly visible to a strategy-level review that asks whether any test could have separated the favored hypothesis from its rival. Diagnostic: Does any single test here look flawed, or does the flaw appear only in the run of tests as the absence of any probe that could have disconfirmed the favorite?

T5: Autonomy versus reduction (a named human failure mode or the shadow of the diagnosticity parent). "Congruence bias" is a canonically studied cognitive phenomenon with its own cargo — the Wason 2-4-6 paradigm, the design-versus-evaluation split, the working-memory account of why confirmation is the easy path, the rival-enumeration remedies aimed squarely at a reasoner. But it is, by definition, a deficiency of a reasoner, and what travels cross-domain is not the bias but the corrective parent it motivates: diagnostic discrimination / likelihood-ratio reasoning, with its siblings severe testing and strong inference, the substrate-independent pattern of arranging observations so the outcome's probability differs between rival hypotheses. That parent recurs as a true co-instance in radar detection, forensics, and experimental design; the bias does not, and calling a confounded protocol or a poorly designed test suite "biased" is a category error, not a stretched analogy, because no cognition produces the failure. Diagnostic: Resolve toward diagnosticity / likelihood-ratio reasoning when carrying the lesson to any non-cognitive system; toward the named bias only when diagnosing a human reasoner's choice of which test to run.

Structural–Framed Character¶

Congruence bias sits at the framed-leaning end of the structural–framed spectrum, pulled off the pure pole by one genuine structural credential — it names a real, reliable, directional regularity of cognition rather than a social convention — but held on the framed side by its evaluative load, its binding to a cognitive substrate, and the fact that its portable content is really its corrective parent, not the bias itself. On evaluative weight it points framed: "bias" is a normatively charged label, a diagnosis of a defect in inquiry, and the whole entry is organized around when confidence is unwarranted and how to correct the failure — the word convicts a piece of reasoning rather than neutrally describing a mechanism the way feedback or diagnosticity itself does. On human-practice-bound it points framed for an instructive reason: the bias is by definition a deficiency of a reasoner, so it dissolves the instant the cognitive substrate is removed — a confounded protocol or an automated test suite is non-diagnostic but not biased, and calling it "congruence bias" is a category error, not a loose analogy; the phenomenon is pinned to human hypothesis-testing cognition even though it is not, unlike a rhetorical figure, constituted by any social institution. Institutional origin is mixed but leans framed: the underlying tendency is a fact about how minds under working-memory pressure choose tests, which no survey or agency invented — but the concept, with its Wason 2-4-6 paradigm, its design-versus-evaluation split, and its rival-enumeration remedies, is furniture of cognitive psychology and scientific-methodology discourse. On vocab travels it points framed: the operative vocabulary is keyed to a reasoner choosing a probe and does not float free of the cognitive substrate. And on import versus recognize it patterns as import/category-error when stretched: the bias does not recur as the same mechanism in radar or forensics — what recurs there is the parent, so any off-substrate use of "congruence bias" imports the label rather than recognizing the named phenomenon.

The one structural-looking feature — and the corrective the whole entry points toward — is diagnostic discrimination / likelihood-ratio reasoning: arranging observations so that the probability of the outcome differs between rival hypotheses, so a test earns its worth by its capacity to come out differently under a live alternative (severe testing, strong inference). That skeleton is genuinely substrate-independent and recurs as mechanism across medical testing, radar and sonar detection, forensics, and experimental design, which is what tempts a structural reading. But it does not pull congruence bias off the framed side, because that portable structure is precisely the parent whose absence congruence bias diagnoses, not what makes "congruence bias" itself travel: the cross-domain reach belongs to diagnosticity/likelihood-ratio reasoning (and, one rung up, to the general bias pattern of reliable directional distortion), while congruence bias's distinctive content — the working-memory account of why confirmation is the cheap path, the design-not-evaluation locus, the reasoner-directed rival-enumeration remedies — is exactly the cognition-bound accent that stays home. Its character: a normatively charged, cognition-substrate-bound failure mode that is real enough to occur observer-free in a reasoning mind yet structural only as the shadow of the diagnosticity parent it fails to instantiate, whose own furniture does not travel past human hypothesis testing.

Structural Core vs. Domain Accent¶

This section decides why congruence bias is a domain-specific abstraction and not a prime, and it carries the case for its domain-specificity — so it is worth being exact about what could lift and what stays home. Its skeleton is genuinely doubled: congruence bias instantiates the general bias pattern (a reliable directional distortion of inference), while the corrective it points toward is diagnostic discrimination / likelihood-ratio reasoning — the parent whose very absence the bias names.

What is skeletal (could lift toward a cross-domain prime). Two abstract structures survive stripping the cognition. First, the general shape of a bias: a systematic, directional departure of a process's output from its target — here, hypothesis-testing that leans consistently toward confirmation rather than scattering as noise. Second, and doing most of the analytic work, the corrective diagnosticity / likelihood-ratio structure the whole entry organizes itself around: arrange observations so that the probability of the outcome differs between rival hypotheses, so a test earns its worth by its capacity to come out differently under a live alternative. That second skeleton is genuinely substrate-independent — it recurs, as true co-instances not metaphors, in medical testing, radar and sonar detection, signal processing, forensics, and experimental design, wherever a probe's value is its power to separate rival states of the world (its normative siblings being severe testing and strong inference). Both are portable; but they are the cores congruence bias shares, not what makes it congruence bias.

What is domain-bound. Almost everything that makes the entry congruence bias in particular is cognition-and-methodology furniture, and none of it survives extraction. The concept is by definition a deficiency of a reasoner: the Wason 2-4-6 paradigm; the working-memory account of why confirmation is the cheaper cognitive path (simulate the world if H1 holds and check, versus holding two hypotheses at once and orienting toward an unanticipated result); the design-versus-evaluation split that separates it from confirmation bias proper; and the reasoner-directed remedies (enumerate the strongest rival, run the disconfirming probe, differential diagnosis, strong inference). The decisive test the entry itself supplies: remove the reasoner and the failure is not congruence bias but a category error — a confounded protocol or a poorly designed automated test suite is non-diagnostic but not biased, because there is no cognition producing the leaning. The bias is constituted by the very mind the prime bar would ask it to shed; what remains without that mind is only the diagnosticity parent.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose cross-domain transfer is recognition of the same mechanism, not analogy. Congruence bias's transfer is bimodal. Within human hypothesis-testing cognition it travels intact — the diagnosticity screen and its corrective discipline carry without translation across scientific methodology, diagnostic medicine (premature closure, diagnostic momentum), and software debugging, because the same effortful-to-hold-two-hypotheses limitation is at work in each. Beyond the human reasoner the bias does not travel at all, and not merely by weak analogy: applying "congruence bias" to a non-cognitive system is a category error, since no cognition produces the failure there. What genuinely reaches those systems is the corrective parent — diagnosticity / likelihood-ratio reasoning — recurring as a true co-instance in radar, forensics, and experiment design; and one rung up, the general bias pattern carries the reliable-directional-distortion reading. So when the cross-domain lesson is needed it is already carried, in more general form, by the parents congruence bias relates to: the portable content is diagnosticity (the structure whose absence the bias diagnoses) and bias (the genus it instantiates), while congruence bias's own cargo — the 2-4-6 task, the working-memory story, the design-locus, the reasoner-aimed remedies — stays bound to human inquiry. The cross-domain reach belongs to the parents; "congruence bias," as named, is the human failure mode whose existence motivates teaching the broader pattern, not the broader pattern itself.

Relationships to Other Abstractions¶

Current abstraction Congruence Bias Domain-specific

Parents (2) — more general patterns this builds on

Congruence Bias is a kind of Bias Prime

Congruence Bias is a specific reliable directional distortion: human test selection leans toward probes predicted by a favored hypothesis rather than probes that separate it from live rivals.
Congruence Bias presupposes Evidence Prime

Congruence Bias presupposes an evidential trace-to-hypothesis relation because a probe is congruent or non-diagnostic only by comparing how likely its trace is under the favored hypothesis and its rivals.

Not to Be Confused With¶

Confirmation bias. The bias of evaluation — interpreting an already-diagnostic test to favour the hypothesis one holds, a thumb on the scale when reading evidence. Congruence bias is the bias of design — choosing a probe that could not have come out differently were a rival true, so the defect is present before any evidence is read. Their remedies are opposite and non-substitutable: blinding and pre-registration fix confirmation bias and do nothing for congruence bias, which yields only to enumerating rivals and building discriminating tests. Tell: Is the defect in how the outcome was read (confirmation bias) or in which probe was run (congruence bias)?
Diagnosticity / likelihood-ratio reasoning (the corrective parent). The substrate-independent structure of arranging observations so the outcome's probability differs between rival hypotheses — a test earning its worth by its capacity to come out differently under a live alternative (its normative siblings being severe testing and strong inference). This is the parent whose absence congruence bias diagnoses; it travels everywhere (radar, forensics, experiment design) as a true co-instance, while the bias does not. Tell: Are you naming the positive discipline of building tests that separate hypotheses (diagnosticity — carry this cross-domain), or the human failure to do so (congruence bias)?
The general bias genus. The umbrella pattern of a reliable, directional distortion of a process's output from its target. Congruence bias is one instance of it, keyed to confirmation-seeking test design; other instances (base-rate neglect, anchoring) distort other stages. Super-type versus instance. Tell: Is the claim about directional distortion in general (the genus), or specifically about designing only confirmable rather than discriminating tests (congruence bias)?
Premature closure / diagnostic momentum. Croskerry's clinical labels for anchoring on an early diagnosis and ordering confirming rather than discriminating workups. These are the same mechanism under a medical name — congruence bias as it appears in diagnostic reasoning — not a distinct bias; the differential-diagnosis corrective is the domain's version of rival-enumeration. Tell: Are these separate phenomena, or the identical design-stage failure named within medicine (they are the latter — the same bias, domain-localised)?
A non-diagnostic-but-unbiased test (confounded protocol, poor test suite). A confounded experiment or a badly designed automated test suite is non-diagnostic — its likelihood ratio sits near one — but not biased, because no reasoner's cognition produces the failure. Calling such a system "congruence-biased" is a category error, not a stretched analogy; what applies to it is the diagnosticity parent, not the named bias. Tell: Is there a human inquirer whose test-choice cognition produced the non-diagnosticity (congruence bias), or is the probe simply constitutionally non-discriminating with no reasoner behind it (a non-diagnostic protocol)?

Neighborhood in Abstraction Space¶

Congruence Bias sits in a sparse region of the domain-specific corpus (61^st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (309 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12