{"id":38,"date":"2026-02-15T18:38:00","date_gmt":"2026-02-15T18:38:00","guid":{"rendered":"https:\/\/yeralan.org\/wp\/?p=38"},"modified":"2026-02-17T10:53:38","modified_gmt":"2026-02-17T10:53:38","slug":"error-as-signal","status":"publish","type":"post","link":"https:\/\/yeralan.org\/wp\/2026\/02\/15\/error-as-signal\/","title":{"rendered":"Error as Signal"},"content":{"rendered":"\n

DOI: https:\/\/doi.org\/10.5281\/zenodo.18669061<\/a><\/p>\n\n\n\n

Prelude: The Offensiveness\nof Error<\/h2>\n

Error offends because it threatens status and predictability. In\nschooling and administration, it is treated as a moral weakness; in\nengineering, as a reliability risk; in public discourse, as a sign of\nuntrustworthiness. Yet the demand for errorlessness is often a category\nmistake. A calculator should not err at addition. A scientific community\nshould. A pilot should not confuse instruments. A research program\nshould \u201cconfuse itself\u201d regularly, because only surprise distinguishes\ndiscovery from repetition.<\/p>\n

The ordinary language of error lumps together fundamentally different\nphenomena: a flawed inference, a noisy sensor, a taboo violation, a\ndeviation from a social script, an exploratory move that fails, an\noutlier observation that eventually rewrites the theory. When we fail to\ndistinguish these, we overcorrect; and overcorrection is a quiet route\nto stagnation.<\/p>\n

\n

\u201cWithout deviation from the norm, progress is not\npossible.\u201d<\/em>
\n\u2014 Frank Zappa<\/p>\n<\/div>\n

This article is written for two audiences at once. The first is the\nengineer who knows, instinctively, that feedback requires an error term\nand that control without discrepancy is blind. The second is the\ninstitutional mind that wishes for frictionless consensus and therefore\ntreats nonconformity as malfunction. Both intuitions are partially\ncorrect. The work is to place them in the right ontology.<\/p>\n

Definitions and a Working\nTaxonomy<\/h2>\n

Error, Mistake, Blunder,\nand Deviation<\/h3>\n

We will use four terms with disciplined intent.<\/p>\n

    \n

  • Error<\/em><\/strong> is a discrepancy between a target\n(truth, specification, norm, or goal) and an outcome.<\/p><\/li>\n

  • Mistake<\/em><\/strong> is an error attributable to a\ndecision procedure (choice of model, plan, parameter, or rule).<\/p><\/li>\n

  • Blunder<\/em><\/strong> is a mistake with avoidable\nnegligence or gross mismatch between competence and act.<\/p><\/li>\n

  • Deviation<\/em><\/strong> is simply difference\u2014it may\nbe an error or it may be signal.<\/p><\/li>\n<\/ul>\n

    The critical claim is that the set of deviations is larger than the\nset of errors, and the set of errors is larger than the set of mistakes.\nSome deviations are the first symptom that the target was\nmisdescribed.<\/p>\n

    Five Types of \u201cError\u201d<\/h3>\n

    For analytical clarity, we classify common cases into five\nfamilies.<\/p>\n

      \n

    1. Logical error<\/em><\/strong>: invalid inference,\ncontradiction, or misuse of implication.<\/p><\/li>\n

    2. Empirical error<\/em><\/strong>: a claim about the\nworld that fails under evidence.<\/p><\/li>\n

    3. Measurement and instrument error<\/em><\/strong>:\nnoise, bias, drift, quantization, sampling artifacts.<\/p><\/li>\n

    4. Normative \u201cerror\u201d<\/em><\/strong>: deviation from a\nsocial convention, protocol, or expectation (not necessarily\nfalse).<\/p><\/li>\n

    5. Productive deviation<\/em><\/strong>: an anomaly that\nexposes model insufficiency, hidden variables, or new\nphenomena.<\/p><\/li>\n<\/ol>\n

      We will later show that \u201cproductive deviation\u201d is not a rhetorical\nflourish but a structural feature of learning systems: variation is the\nsubstrate of selection, and discrepancy is the substrate of control.<\/p>\n

      Conformity as Error\nManufacture<\/h2>\n

      The most famous laboratory demonstration of socially induced\nmisperception is Asch\u2019s conformity paradigm, in which individuals\nconform to an incorrect majority judgment on an easy perceptual task\n(Asch 1951,\n1955)<\/span>. The immediate lesson is not that humans are stupid, but\nthat perception is not a private instrument; it is a socially\nconditioned output. Conformity is therefore a generator of error in the\nstrict sense: it increases discrepancy between judgment and reality.<\/p>\n

      This matters beyond psychology. In organizations, the majority\nopinion often becomes a proxy for truth. In scientific communities,\nreputational gradients can cause hypothesis lock-in. In bureaucracies,\nconsensus can function as a legitimacy machine that suppresses\ninconvenient observations. Conformity does not merely correlate with\nerror; it can produce<\/em> it by changing the cost function of\nreporting what one sees.<\/p>\n

      Minority influence and\nepistemic rescue<\/h3>\n

      If conformity manufactures error, dissent can manufacture correction.\nThe classical finding in minority influence research is not that\nminorities always win, but that consistent minorities can shift the\nprivate processing style of the majority toward more systematic\nevaluation (Moscovici\n1980)<\/span>. The point is structural: a minority position acts as a\nperturbation that prevents premature convergence.<\/p>\n

      The analogy to learning systems is tight. A group without dissent is\nlike a model trained only to minimize local loss: it converges quickly\nand confidently, and fails catastrophically when the environment\nshifts.<\/p>\n

      Serendipity: When\n\u201cWrong\u201d Opens the World<\/h2>\n

      Science and engineering histories contain a recurring motif: the\nproduct was not sought, the result was not predicted, the anomaly was\ninitially an error, and only later did it become a discovery. Accounts\ndiffer in detail, but the epistemic pattern is stable. Merton formalized\nthis as the serendipity pattern<\/em>: an unanticipated observation\nbecomes strategically fruitful because it reveals an underlying,\nunrecognized structure (Merton and Barber 2004)<\/span>.<\/p>\n

      In this sense, some \u201cerrors\u201d are a form of involuntary exploration. A\nsystem is probing the boundary of its model, and the boundary pushes\nback.<\/p>\n

      Contingency without\nsufficiency<\/h3>\n

      One must be careful. Most accidents are merely waste. Serendipity is\nnot a license to be sloppy; it is an argument for maintaining an\ninterpretive posture toward anomalies. The same observation can be\nthrown away as noise or cultivated as a signal. The difference lies in\ndisciplined curiosity: the willingness to ask, \u201cwhat assumption did this\nviolate?\u201d<\/p>\n

      Error\nas Control Variable in Cybernetics and Engineering<\/h2>\n

      In control theory, the error signal is not embarrassment; it is the\nfundamental variable that drives correction. Let \\(r(t)\\)<\/span> be a reference trajectory and \\(y(t)\\)<\/span> the measured output. The error is\n\\[e(t)=r(t)-y(t).\\]<\/span> If \\(e(t)\\equiv 0\\)<\/span> at all times, then either\nthe system is perfectly controlled or (more commonly) the measurement is\nlying, the reference is trivial, or the system is not interacting with\nan environment that can surprise it. In real systems, error is expected;\nthe question is whether the feedback loop transforms error into\nstability.<\/p>\n

      Wiener\u2019s cybernetics made this explicit: goal-directed behavior\nrequires feedback, and feedback requires discrepancy (Wiener 1948)<\/span>. Ashby\nsharpened the constraint: regulation requires variety sufficient to\nmatch disturbances\u2014the Law of Requisite Variety (Ashby 1956)<\/span>. The regulator that cannot\nexpress alternative actions cannot reduce error; the organization that\ncannot tolerate dissent cannot correct itself.<\/p>\n

      Boundary failure<\/h3>\n

      When organizations suppress error signals, they resemble unstable\ncontrollers that saturate or clip feedback. The resulting behavior is\nfamiliar: hidden drift, delayed recognition, and sudden collapse. Error\nsignals do not disappear when ignored. They migrate into unmodeled\nchannels.<\/p>\n

      Fallibility as a\nCondition for Learning<\/h2>\n

      Bayesian\nupdating and the necessity of surprise<\/h3>\n

      A learning agent updates beliefs in proportion to prediction error.\nIn Bayesian terms, evidence modifies priors through likelihood; in\npredictive processing language, the system minimizes prediction error\nthrough model revision and action. If observations never contradict\npredictions, no update occurs. A perfectly \u201cright\u201d system is\nepistemically inert because it never receives differential\ninformation.<\/p>\n

      Exploration,\nexploitation, and productive failure<\/h3>\n

      In reinforcement learning, the exploration\u2013exploitation dilemma\nformalizes a deep truth: optimal long-run performance requires\nnon-optimal short-run actions. Exploration looks like error locally.\nGlobally, it is insurance against model misspecification and\nnonstationary environments (Sutton and Barto 2018)<\/span>. To forbid\nexploration is to demand that an agent behave as if it already knows the\nworld. That demand is logically incoherent.<\/p>\n

      Machine\nLearning and the Myth of Error-Free Output<\/h2>\n
      \n

      \u201cTo err is a cognitive invariant.\u201d<\/em>
      \n\u2014 Yeralan<\/p>\n<\/div>\n

      Public expectation often treats computational output as oracle. When\nan AI system makes a mistake, observers infer untrustworthiness. But\nmodern machine learning systems are, in important respects,\napproximation machines. They generalize by compressing; they predict by\ninterpolating; they err by design because the world is not fully\nobserved and the training distribution is finite.<\/p>\n

      Two distinctions matter.<\/p>\n

        \n

      • Training error vs.\u00a0generalization\nerror<\/em><\/strong>: a model can achieve low training error by\nmemorization and still fail in deployment.<\/p><\/li>\n

      • Calibration vs.\u00a0accuracy<\/em><\/strong>: a model may\nbe accurate on average yet systematically overconfident or\nunderconfident in its probabilities.<\/p><\/li>\n<\/ul>\n

        The \u201cerror-free AI\u201d ideal therefore invites the wrong kind of trust:\na trust in surface precision rather than in well-characterized limits.\nIn safety-critical contexts, what we want is not perfection but\nknown failure modes<\/em> and measured uncertainty<\/em>.<\/p>\n

        Adversarial fragility<\/h3>\n

        The existence of adversarial examples demonstrates that models can be\nconfident and wrong under tiny perturbations (Goodfellow, Shlens, and Szegedy\n2015)<\/span>. This is not a moral flaw. It is a geometrical fact about\nhigh-dimensional decision boundaries and training objectives. The remedy\nis not fantasy perfection, but robustness engineering and humility about\nepistemic reach.<\/p>\n

        The\nGenius Who Fumbles: Local Failure and Global Insight<\/h2>\n

        We often commit a social fallacy: we expect competence to be uniform\nacross domains. Yet cognitive specialization and resource constraints\nimply trade-offs. A person may have exceptional capacity for abstraction\nand weak capacity for mundane logistics; a research group may be\nbrilliant at invention and incompetent at documentation; an institution\nmay be excellent at credentialing and poor at truth-seeking.<\/p>\n

        The point is not to romanticize dysfunction. It is to reject a\nsimplistic inference: that a localized failure invalidates a broader\ncognitive contribution. Conversely, it is also to reject the inverse\nromantic myth: that brilliance excuses negligence. The rational position\nis structural: competence is multidimensional, and its failures are\ninformative about system design.<\/p>\n

        Normative Error\nand the Politics of Deviance<\/h2>\n

        Some \u201cerrors\u201d are not errors at all; they are violations of\nconvention. A student who challenges a professor may be \u201cwrong\u201d in tone\nand right in substance. A whistleblower violates protocol and restores\ntruth. A scientist who refuses to cite the fashionable paper may be\npunished socially while acting epistemically.<\/p>\n

        Here error language becomes a control technology. To label a\ndeviation as \u201cerror\u201d is to place it inside a moral economy: blame,\nshame, and correction. This is often useful. It is also often abused.\nInstitutions that conflate normative compliance with truth acquisition\ndrift toward what might be called epistemic authoritarianism: the map\nbecomes the enforcement of the map.<\/p>\n

        Synthesis: Error as Epistemic\nGate<\/h2>\n

        We can now state the central claim without metaphor.<\/p>\n

        \n

        A cognitive system capable of revision must be capable of error; a\nsocial system capable of truth must be capable of dissent; a control\nsystem capable of regulation must be capable of discrepancy.<\/p>\n<\/div>\n

        Popper\u2019s emphasis on falsifiability can be read as an\ninstitutionalization of error: a demand that theories expose themselves\nto refutation (Popper\n1959)<\/span>. Kuhn\u2019s account of scientific change emphasizes the role of\nanomaly: persistent error in prediction becomes the seed of paradigm\ntransition (Kuhn\n1962)<\/span>. These are philosophical statements, but they align with\nthe engineering account: discrepancy is the driver of adaptation.<\/p>\n

        The practical moral is austere. We must distinguish error from\ndeviation, mistake from blunder, noise from anomaly. And we must\ncultivate systems that do not merely punish error, but interpret it.<\/p>\nConclusion:\nAgainst the Fantasy of Frictionless Cognition<\/h2>\n

        The fantasy of error-free cognition is attractive for the same reason\nutopias are attractive: it promises comfort. But comfort is not an\nepistemic virtue. Where uncertainty is real, error is inevitable; where\nlearning is real, error is necessary; where coordination is real,\ndissent is vital.<\/p>\n

        This is not an invitation to carelessness. It is an insistence on\nproper goals. In engineering, reduce error to preserve function. In\ninquiry, preserve error to preserve discovery. In governance, separate\ncompliance from truth. In AI, demand calibration and transparency rather\nthan oracle theater.<\/p>\n

        The higher aim is not perfection. It is corrigibility.<\/p>\n

        \n
        \nAsch, Solomon E. 1951. \u201cEffects of Group Pressure Upon the\nModification and Distortion of Judgments.\u201d<\/span> Edited by Harold\nGuetzkow, 177\u201390.\n<\/div>\n
        \n\u2014\u2014\u2014. 1955. \u201cOpinions and Social Pressure.\u201d<\/span> Scientific\nAmerican<\/em> 193 (5): 31\u201335. https:\/\/doi.org\/10.1038\/scientificamerican1155-31<\/a>.\n<\/div>\n
        \nAshby, W. Ross. 1956. An Introduction to Cybernetics<\/em>. London:\nChapman & Hall.\n<\/div>\n
        \nGoodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. 2015.\n\u201cExplaining and Harnessing Adversarial Examples.\u201d<\/span>\nInternational Conference on Learning Representations (ICLR)<\/em>. https:\/\/arxiv.org\/abs\/1412.6572<\/a>.\n<\/div>\n
        \nKuhn, Thomas S. 1962. The Structure of Scientific Revolutions<\/em>.\nChicago: University of Chicago Press.\n<\/div>\n
        \nMerton, Robert K., and Elinor Barber. 2004. The Travels and\nAdventures of Serendipity: A Study in Sociological Semantics and the\nSociology of Science<\/em>. Princeton, NJ: Princeton University Press.\n<\/div>\n
        \nMoscovici, Serge. 1980. Toward a Theory of Conversion Behavior<\/em>.\nEdited by Leonard Berkowitz. Vol. 13. Academic Press.\n<\/div>\n
        \nPopper, Karl R. 1959. The Logic of Scientific Discovery<\/em>.\nLondon: Routledge.\n<\/div>\n
        \nSutton, Richard S., and Andrew G. Barto. 2018. Reinforcement\nLearning: An Introduction<\/em>. 2nd ed. Cambridge, MA: MIT Press.\n<\/div>\n
        \nWiener, Norbert. 1948. Cybernetics: Or Control and Communication in\nthe Animal and the Machine<\/em>. Cambridge, MA: MIT Press.\n<\/div>\n<\/div>\n\n\n\n

        <\/p>\n","protected":false},"excerpt":{"rendered":"

        On Mistake, Blunders, Nonconformity, and Cognitive Expansion<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-38","post","type-post","status-publish","format-standard","hentry","category-essay"],"_links":{"self":[{"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/posts\/38","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/comments?post=38"}],"version-history":[{"count":8,"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/posts\/38\/revisions"}],"predecessor-version":[{"id":86,"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/posts\/38\/revisions\/86"}],"wp:attachment":[{"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/media?parent=38"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/categories?post=38"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/yeralan.org\/wp\/wp-json\/wp\/v2\/tags?post=38"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}