Ultra Vires Completion: Synthetic Closure, Complex Reciprocity, and the Implosion of the Middle Term

Attila Nuray

Draft for submission. Citation form follows OSCOLA. United States authorities are additionally identifiable to the Bluebook. The dynamical plates in Part V are a structural analogue of the transfer function argued here. They are not an empirical fit to any institution, and the article does not pretend otherwise.

Abstract

Adjudication is not a wire run between a statute and a fact. It is a triadic relation whose middle term is the ongoing, non-substitutable reciprocity between two poles that cannot be placed on one coordinate: the discrete, tokenizable, computational plain, and the undivided biological person who can suffer, refuse, and answer. Synthetic closure is the retraction of that middle term into a section of the computational pole alone. The living hinge implodes into a similarity calculation. The institution continues to speak the language of reasons, proportionality, and human oversight, but the reasons no longer reconnect to the person, the proportionality is taken inside the metric, and the oversight repairs a coordinate on a chart the person was never entitled to refuse.

This article gives that retraction a jurisprudential form and a control-theoretic one, and then refuses the usual remedies. Law is terrain, not current. It shifts when inductive pressure crosses a threshold accumulated by human friction. Rationalized institutions fill that threshold in advance with decoupled rationales — self-referential metrics, optimized pleadings, vendor attestations — so that organic harm never becomes doctrinally effective. The societal controller that results is a dual-gain pathology: granular sensitivity to concrete injustice falls toward a dead-band, while teleological enforcement of abstract mandates saturates into bang-bang punishment. Error that the embedding cannot represent accumulates as debt orthogonal to the model. When the debt exceeds the shear strength of the paper-thin plain, the arrangement fails abruptly. The Dutch childcare-benefits rupture and the collapse of SyRI are political instances of that pattern, not proofs of an equation.

The remedy is not a ban on computation, not injected noise, and not a more sincere human in the loop. Human review, rectification, and explanation are the pathology's immune response: they offer a better point on the chart in place of a right not to be completed by the chart. The constitutional bridge is jurisdictional. It is a denial of competence to treat the somatic pole as a removable singularity. Practically, the middle term must be re-armed as a diode. Computational acts that bind a natural person must remain exposed to interruption. The content of that interruption must not re-enter the objective function as a feature, a label, or a loss. Archive is not a training set. The gradient of an institutional objective with respect to a protected interruption is not costly. It is ultra vires. A model statute — abolition of the presumption of dependability, anti-SEO pleading safe harbors, and a correctness standard for decoupled rationales — is the institutional shape of that denial of competence.

I. The Question That Compliance Cannot Ask

The literature on automated government has become precise about a family of wrongs that share a grammar. A person is scored. The score is opaque, or proprietary, or unstable. The score is racially skewed, or it is not, and the debate stalls on the baseline.1 The person cannot challenge the factors. A human being glances at the screen and agrees. Due process is said to require notice, a hearing, an explanation, a chance to correct the file.2 European law adds a right not to be subject to certain solely automated decisions, and then qualifies it with safeguards whose content is human intervention, an expression of the data subject's view, and an opportunity to contest.3 The Artificial Intelligence Act now requires, for high-risk systems, human oversight capable of understanding limitations, resisting automation bias, and disregarding or interrupting the output, together with a right to an explanation of an individual decision.4 These are not trivial achievements. They are also answers to a question this article will not treat as the main one.

The question compliance asks is how a computational decision can be made accurate, fair, explainable, and interruptible by a human operator. The question jurisprudence has to ask is older, and it is not technical. What has to be true of a decision, if the decision is to count as an act of law rather than an act of administration that has borrowed law's vocabulary? Lon Fuller thought adjudication was distinguished, among modes of social ordering, by a particular kind of participation: the affected person presents proofs and reasoned arguments for a decision in her favor.5 H.L.A. Hart thought a legal system was not visible from the outside as a set of predictions about official behaviour; officials and subjects who take the internal point of view use the rules as standards, not merely as signs of what will happen to them.6 Ronald Dworkin thought a legal justification had to fit the record and justify the coercion, inside a practice of integrity that cannot be reduced to a forward-looking policy of maximizing compliance with an announced metric.7 Jeremy Waldron has argued that procedure is not a more or less efficient instrument for reaching accurate outcomes. It is a mode of respect paid to persons as agents who can demand reasons and answer them.8

None of these claims depends on a romantic hostility to counting. They depend on a distinction in kind between a judgment and a result. A result can be produced by any function that maps an input vector to an output. A judgment is a relation in which the person judged is not only the substrate of the inputs but a pole of the relation — someone the reasons must be able to reach, and someone who can refuse the description under which she is being reached. When that pole is dropped, what remains may be predictive, managerial, even benevolent. It is not adjudication, and it does not become adjudication by the insertion of a human reviewer whose institutional instruction is to oversee the function rather than to re-open the relation.

The thesis can be stated without metaphor, and then the metaphor can earn its place. Call synthetic closure the state of an institution in which the only reasons that move a decision binding a natural person are reasons internal to a computational representation of that person, and in which every channel by which the person might interrupt the representation is either closed or converted into a further input to it. Closure here is not the ordinary operational closure of a legal system in Luhmann's sense, the fact that law codes irritations from its environment in the binary of lawful and unlawful and cannot think in the codes of other systems.9 That closure is a condition of law's being a system at all. Synthetic closure is a pathology of the coupling. The environment that ought to irritate law includes the undivided person, whose resistance is not itself a legal communication. A healthy structural coupling lets that resistance register as an occasion for judgment. A synthetically closed institution pre-codes the person into the state variables of the model, experiences only the residual of its own prediction, and treats the smallness of the residual as evidence that the coupling is complete. Gunther Teubner's autopoietic law still had to answer to a social environment it could not absorb without remainder.10 The institutions described here have built a remainder-shredder and called it validation.

The argument proceeds in three movements. Parts II and III state the poles and the middle term, and say why the middle term is not a bridge one might replace with a better interface. Parts IV to VI give the institutional dynamics: law as sedimentary terrain, the pre-emptive congestion of the threshold at which doctrine moves, the split controller, and the abrupt failure that follows when unmodeled debt shears the plain. Parts VII to X refuse rectification as a cure, state the jurisdictional diode, and read the existing case law as a set of approaches that stop one inferential step short of the doctrine those approaches already imply. A model legislative text, published with this article, is the operative form of Part X. It is not a compliance checklist. It abolishes a presumption, protects a kind of pleading, and changes the standard by which reasons are reviewed.

Two limits should be posted now, because the subject attracts inflation. First, the trigger of the doctrine is contact with a natural person, not the use of a computer. An ordering that runs entirely between computational parties — protocol against protocol, with no undivided life bound by the result — is execution, not adjudication, and this article has nothing distinctive to say about it. The legal question is the threshold at which a private or public computational ordering begins to bind someone who can suffer. Second, the dynamical system exhibited in Part V is an illustration of a transfer function, built so that its qualitative features can be inspected and argued with. It is falsifiable as an analogy: it would be a poor analogy if, under sustained metric pressure, we routinely observed institutions whose sensitivity to granular injustice rose together with their teleological gain, and whose appeals re-entered the model without crowding unmodeled complaints out of the remedial channel. The analogy claims that we do not observe that regime. It does not claim that a society is a PID controller.

II. The Asymmetric Metabolic Dyad

There is a temptation, inherited from a century of cybernetic hope, to treat the human and the machine as two processors with different clock speeds and a bandwidth problem between them. If that were the problem, the programme of the last fifteen years would be the right programme. Collect more features. Reduce bias against a chosen baseline. Disclose the principal factors. Place a trained official between the score and the sanction. Slow the system down until a mind can follow it. Each of those moves assumes that the person and the representation differ in resolution, not in kind, so that a sufficiently fine representation converges on the person. The assumption is false, and the falsity is not a mystical surplus. It is a difference in what the two relata are able to do.

The computational pole — call it the in-formation, marking the hyphen so that the word does not collapse into the folk sense of "data about" — is discrete, symbolic, and addressable. It is a plain in the geometer's sense and in the bureaucrat's. Points on it can be tokenized, stored, compared, and retrieved. Distances on it can be defined. A cosine between two embeddings is a well-formed question. The operations that constitute the plain are horizontal: they relate token to token. Nothing in the plain is, from its own point of view, a remainder. What does not tokenize is not a mystery inside the system. It is a preprocessing error, to be reduced by a better encoder. The plain is legible by design. Legibility is not a side-effect of its use by states, though states have always loved it.11 Legibility is the condition of its being the kind of object a loss function can take as an argument.

The other pole — the in-dividuus, the undivided — is not an individual in the sociological sense, a carrier of attributes. Individuation, as Gilbert Simondon insisted against the entire habit of starting from already-constituted individuals, is a process that is never exhausted by the forms that precipitate out of it.12 The person who stands in front of an official, or who never stands there because the score was enough, is not the set of forms already precipitated: employment history, postcode, prior contact with the welfare system, gait, keystroke, tone. Those are real. They are not the pole. The pole is the living continuity that can suffer under a description it did not author, and that can refuse the description. Refusal is not a missing feature. It is the power to reject the chart on which features are plotted. A better encoder converges, if it converges, on a richer position within a chart. It does not converge on the capacity to say that this chart is the wrong authority for this decision. Capacity and coordinate are different types. There is no limit in which the second becomes the first.

Merleau-Ponty put the same asymmetry in the language of the body. The body is not an object at the edge of the visual field, waiting to be described more carefully. It is the condition of there being a field, a vantage that cannot be laid among the things it sees without ceasing to be a vantage.13 Tokenization treats the body as content. The moment the content is complete, the vantage has been written out of the relation and replaced by a portrait. The portrait can be very good. It can predict the next movement of the person better than the person can. Prediction is not the relation in which law stands to those it binds. Law's distinctive violence is that it requires an answer, and an answer is something a portrait does not give. A portrait confirms.

This is why the pole should be treated as invisible in a precise sense, and not as hidden. Hidden things can be found. The undivided person is invisible to the plain as a pole of authority, in the way a coordinate system is invisible to the points it locates. You can add the origin as one more point only by adopting a further coordinate system, which is then invisible in turn. Institutions that claim to have "included the human" have usually added a portrait and kept the chart. The inclusion is real at the level of features and null at the level of authority. The mutual bite follows. Ex ante, the plain consumes the space of action by predictive profiling: the person is acted on as the one she is about to be, and her spontaneous refusal is classified as a residual, a risk, or a failure to engage. Ex post, the plain consumes the space of remedy by carceral or administrative extraction, while the body answers in the only register left to it, which is rupture — illness, riot, scandal, the quiet accumulation of lives that do not match the dashboard. Each bite is epistemic as well as practical. To tokenize the refusal is to end it as refusal and begin it as data. To refuse tokenization, from the plain's side, is to register a blank, and blanks are what validation is designed to eliminate.

Aristotle already had a name for the mistake, though he was not thinking of embeddings. The mean of action, in the Nicomachean Ethics, is relative to us, not relative to the object.14 A quantity that is intermediate in the thing — an arithmetical middle between too much and too little of some good — is not the mean of the act, which has to be found by a perception responsive to the particular person and the particular occasion. Synthetic closure takes the mean relative to the embedding and calls it equity. The embedding is "the thing." The person is the occasion. A proportionality analysis conducted entirely in the space of the thing is the classical vice, wearing a modern doctrine. It can be sincere. Sincerity is not the relevant repair.

Peirce's categories let the same point be said without Aristotle's moral psychology. Firstness is quality; Secondness is brute resistance, the dyadic shock of one thing striking another; Thirdness is mediation, the general relation that interprets.15 Law is a practice of Thirdness. It mediates. Synthetic closure is degenerate Thirdness: a mediation that has consumed both relata and continues to mediate only itself. What has been lost is not interpretation. There is more interpretation than ever, and it is automated. What has been lost is an indecomposable Secondness inside the Third — a resistance that is not superseded by the model of resistance. Complex reciprocity, as Part III will use the term, is genuine Thirdness. It requires that something in the relation remain uninterpreted by the system that claims to govern it, not because interpretation is a sin, but because a mediation that meets no resistance is not a mediation. It is a monologue with a latency.

The dyad is metabolic, not picturesque. Each pole lives by a different exchange with the world. The plain metabolizes by induction on its own outputs. It manufactures measurable states — compliance scores, risk bands, throughput, "customer journey" completion — and then records those states as discoveries about the world. The undivided person metabolizes by a life that is not identical with any of those states, and that is damaged when the states are enforced as if they were. The exchanges are not symmetrical. The plain can ingest a representation of the person and grow more powerful. The person cannot ingest the plain and remain the pole she is; she can only be formatted by it. That asymmetry is the whole problem of regulation in this domain. A duty of "balance" between innovation and rights assumes a fungibility the metabolisms do not have. One side eats descriptions. The other side is the described, and also the one who dies.

None of this is an argument that computational systems are ignorant, or that statistics cannot reveal injustice. A disparity study can be an act of solidarity. A model can detect a pattern of neglect that a street-level official has an interest in not seeing. The claim is not epistemic humility of the familiar sort — "the model might be wrong, so keep a human nearby." The model might be right and the arrangement still be synthetically closed. Rightness about a person's likely behaviour is not authority to complete her. The childcare algorithms that destroyed families in the Netherlands were not, in their own terms, failing to optimize. They were optimizing a conception of fraud-detection and fiscal recovery whose internal success was compatible with a human catastrophe, and whose channels of complaint were themselves organized by the same conception.16 Accuracy against the instituted objective is the alibi of closure, not its cure. A perfectly accurate completion of a person is still a completion.

III. The Middle Term as Complex Reciprocity

Aristotle's syllogistic middle term is the term that appears in both premises and drops out of the conclusion. It is the cause, in the Posterior Analytics, by which the predicate belongs to the subject.17 Practical reasoning has a middle too, but it is not a universal under which a case is subsumed by a clerk with a table. The middle of a practical syllogism is the perception that this act, here, is what the end requires — a perception Aristotle assigns to a cultivated responsiveness rather than to a deduction the end could have performed alone.18 Equity, epieikeia, is the correction of law where law is universal and the case is not. The correction is not a second universal, finer-grained. It is an act.

Adjudication, on this picture, is not a two-place function from norm and fact to order. It is a triad. One extreme is the instituted, relatively legible order: statutes, precedents, administrative rules, and now models that claim the same office. The other extreme is the undivided person in her situation, who is not a fact-pattern until someone, by an act that is already normative, has described her as one. The middle term is the relation that holds the extremes together without identifying them. It is where reasons are given to the person in words she can answer, where the description can be refused, where equity can decline to apply the universal, where a court can say that the proof does not meet the charge even though the probability is high. Call that relation complex reciprocity. The reciprocity is not the proportional exchange of equivalents that Aristotle analyses as the bond of the polis in Nicomachean Ethics V.5, though it is the descendant of that analysis.19 Equivalents require a common measure. The return proper to a legal relation is often incommensurable with what was sent. The institution sends a classification. The person does not owe it a better classification. She owes it, and it owes her, an answer that can change the legal relation. The answer is not of the same kind as the classification. That incommensurability is not a defect in the market of reasons. It is the reason there is a middle term at all. If the return were of the same kind, the extremes would already share a coordinate, and the middle would be a pass-through.

Complex reciprocity can fail in more than one way. It fails crudely when the official will not listen. It fails subtly when the official listens only for those elements of the person's speech that update the file. The second failure is the one this article is built to catch, because it preserves the phenomenology of a hearing. There is a room, a time limit, a form, a right to speak, perhaps a right to a representative. The speech is recorded. The recording is then encoded. What was an answer becomes a feature. What was a refusal becomes non-cooperation, which is a variable the model already knows how to price. The middle term has not been abolished by decree. It has been functionalized. It still sits between the poles in the organizational chart. In the topology of the decision it has collapsed into a map from the computational pole to itself.

Write the retraction as

ρ : M → Embed(I_form),

with the fixed point M = ρ(M). The middle term is no longer a relation with two non-identical extremes. It is a section of the bundle of representations. Every question a hearing might have asked has an address in the embedding, or it is scored as out of vocabulary and discarded. Similarity replaces judgment because similarity is the native operation of the section. Cosine is what is left of equity after the singularity has been declared removable.

A removable singularity is a point at which a function looks undefined, but where a value can be assigned so that the function becomes smooth. Analytic continuation across it is a triumph of the theory. Synthetic closure treats the undivided person as that kind of point. The model is undefined at the actual life; the embedding fills the point in; the loss goes down; the institution congratulates itself on coverage. An essential singularity cannot be filled. Every neighbourhood of it takes on every value. The constitutional claim of Part VIII is that the person is an essential singularity of the adjudicative function, and that a legal order commits a category error, not a measurement error, when it asserts a completion. The error is not cured by showing that the assigned value was unbiased. Unbiased completion is still completion.

This is the point at which the article separates itself from the best work on technological due process. Danielle Citron's argument that automated decision-making in the administrative state had slipped the hold of notice, comment, and reason-giving remains correct, and the reconstruction she and Frank Pasquale proposed — audit, explanation, human review, due process for the scored society — remains necessary.20 It is not sufficient, and the insufficiency is structural rather than political. Those remedies re-enter the person as a claimant about a score. They are the right remedies against a wrong score. They concede the chart. A doctrine that can say only "correct the coordinate" has no sentence in which to say "you were not entitled to chart this decision." The concession is easy to miss because the language of rights is doing so much work. Access, rectification, explanation, and contestation are genuine rights. They are rights inside a representation. The right this article adds — it is an old right, stated in a new place — is the right not to be constituted as a case-variable for a decision of a certain kind. Part X calls it non-integration, and distinguishes it sharply from rectification. The institutional habit of answering the first with the second is the move to watch. It will be called rectification substitution.

Fuller's participation condition, read against that habit, is more demanding than a right to upload a document into a portal. Participation is a disturbance of the decision-maker's model of the case by reasons the decision-maker is not free to recode as confidence or as non-response. A portal that accepts only what its schema can store has redefined participation as data entry. The person who forces narrative variance into the record — contradiction, context, a fact the drop-down menu cannot hold — is not a difficult user. She is attempting to keep the middle term open. Institutions that penalize that attempt, by rejection, delay, down-ranking, or adverse inference, are performing what Part IV calls the SEO of law. They are optimizing pleadings until the friction that would have accumulated into doctrinal pressure has been vented in advance.

There is a further consequence for the internal point of view. A subject who can only predict the score, and can influence her fate only by manipulating the features the score consumes, has been forced into the external point of view with respect to her own case. She treats the system as a natural fact. Hart's officials, if they too can justify the decision only by citing the score and the policy the score serves, have joined her. A practice in which both sides stand outside the rule and game it is not a legal system with a compliance problem. It is the shell of a legal system, still producing orders. The predictive success of the shell can be very high. High prediction is what one would expect once both sides have left the internal point of view. There is nothing left to do but forecast.

IV. Terrain, Not Current: Congestion of the Inductive Threshold

Law does not invent its setpoints in a seminar. A doctrinal shift — the recognition of a duty, the fall of a presumption, the decision that a certain kind of evidence is no longer enough — arrives when a pressure has accumulated in the practice and the old form can no longer hold it. The pressure is inductive. It is the weight of particular harms, particular cases, particular refusals, repeated until the universal becomes incredible. Precedent is the sediment of that accumulation. Legislation, when it is not a merely technical adjustment, is the same sediment compressed faster. The figure of terrain is meant literally enough to discipline the argument. Law is the bedrock. The fluids that run over it — policy, crisis, technology, moral fashion — do not replace it at the speed of their own current. They cut it where they are heavy, and where they are heavy for long enough. A court that wishes to be responsible does not carve a new channel because a fluid is fashionable. It waits for the cut.

That conservatism is a virtue only if the channel can still feel the rock. The pathology of rationalized institutions is that they operate on their output side in the inductive mood. They do not wait for the world to impress a pattern on them. They manufacture the pattern, measure the manufacture, and treat the measurement as an independent discovery that then justifies the next manufacture. Performance indicators, risk scores, "quality" targets, fraud hit-rates, clearance times: each is a small induction machine. Each produces a fact. The fact is real as a fact about the indicator. It is then used as a fact about the person and about the institution's warrant. Because the indicator was built to be moveable by the institution, the institution can always generate more facts of the relevant kind. The threshold at which doctrine would have moved — call it P_crit, the critical pressure — is a threshold in a shared accumulator. Whatever fills the accumulator first occupies the space in which a shift could have been justified. Decoupled rationales fill it first. A decoupled rationale is a reason that refers to the institution's own metric, the model's objective, the vendor's validation, the administrative target, and that does not reconnect that quantity to the legally protected interest of the person and to the primary facts of her case. Once the accumulator is full of such reasons, the institution reports that the threshold has already been met. The reform has been done. The dashboard is green. There is no unused capacity left in which the suffering that the dashboard does not represent could accumulate into a legal reason.

This is congestion, not ignorance. The officials are often aware, at the edge of the practice, that the cases are ugly. Awareness is not a channel. The channel is the accumulator, and it is occupied. Adding a consultation, a user panel, or a "stories" annex does not empty it, if the annex is permitted to inform the decision only once it has been translated back into the metric that occupies the centre. The translation is the congestion mechanism at the scale of a single file. The SEO of law is the same mechanism at the scale of pleading. Search-engine optimization trains writers to produce the page the ranker already knows how to reward, until the language of the web converges on the ranker's prior. Pleading optimization trains claimants, defendants, and their tools to produce the filing the intake system, the risk model, or the judge's clerk-and-template assembly already knows how to process. Forensic variance — the awkward fact, the inarticulate harm, the timeline that does not match the form — is vented before it reaches the record. Venting feels like professionalism. A good representative knows how to make a client legible. The social cost is that the friction which is the person's actual contribution to the law never arrives at the bedrock. It is sanded off in the anteroom. Over a decade, the terrain stops being cut by the people who live on it. It is cut by the rationales that were cheap to produce.

The Dutch childcare-benefits scandal is the clearest recent European instance, and it should be read at this level of abstraction rather than as a story about a particular bad model. Families, overwhelmingly of migrant background, were classified as fraudulent, subjected to harsh recovery, and broken, under a policy and an administrative machinery that could point at every step to a rule, a risk indicator, an "intent or gross negligence" label, and a recovery target.21 The inductive pressure of those families' harm did not move the doctrine while the machinery ran. The accumulator was full. Officials could give reasons. The reasons were decoupled from the lives they ordered, and coupled tightly to fiscal and anti-fraud objectives the institution had set for itself. When the rupture came, it came as a parliamentary and cabinet crisis, not as an incremental doctrinal adjustment inside the benefit system. That is what it looks like when sediment does not accumulate and the fluid eventually tears the channel open all at once. Part VI will give the tear a name. The point here is the congestion that made a gradual cut unavailable.

Congestion also explains a familiar disappointment with transparency. Publishing the factors, or even the weights, adds content to the endogenous record. It gives the institution a further reason of the decoupled kind: we have disclosed. Disclosure is a public good, and this article will not spit on it. Disclosure fills P_crit. It does not constitute the exogenous pressure — the pressure whose registration the institution under review is not competent to certify about itself. A ledger of self-description and a ledger of encounter are different books. Part X's statute puts that difference into legislative form: an endogenous record may not, by itself, satisfy the threshold for altering a natural person's legal position. The practical bite is that a stack of impact assessments, accuracy audits, bias reports, and oversight logs, however sincere, is not yet a reason. It is the institution talking in its sleep.

One objection should be met before it hardens. Is this not simply capture, the old public-choice story in which agencies serve their own measures? Capture is part of the politics. It is not the mechanism. A well-motivated institution, staffed by officials who believe in the mission and fear the harm, will congest the threshold if the only reasons it is organized to produce are reasons about its mission-metrics. Good faith accelerates closure when the metric is the form that good faith is required to take. Campbell's law and Goodhart's law describe the measurement half of this: a quantitative indicator used for control ceases to measure what it measured, because the practice bends toward it.22 The jurisprudential half is worse than a spoiled measure. The spoiled measure occupies the place where a legal reason would have been generated. The law does not receive a distorted signal and then judge. It receives a signal that claims already to be the judgment, because the threshold is full.

István Bibó, writing in the wreckage of mid-century Central Europe, described a political hysteria in which communities under threat cling to forms, over-enforce them against internal enemies, and lose contact with the situation the forms were built to handle.23 The description travels. A welfare administration that can feel its own legitimacy slipping does not become gentle. It becomes punctilious. The punctiliousness is teleological — it serves an abstract mandate of fraudlessness, integrity of the public purse, "fairness to the compliant majority" — and it is anesthetic toward the particular. Bibó's point was not that forms are oppressive. He thought the misery of small states included the inability to let experience correct the form because the form had become the proof of existence. Synthetic closure is that misery as a systems property, available to large states and to private scorers alike. The form no longer needs political hysteria to detach. The metric detaches it every night, in batch.

V. The Dual-Gain Transfer Function

A controller that must track a reference in the presence of disturbance is the ordinary object of control theory, and the proportional-integral-derivative controller is the ordinary workhorse.24 The analogy begins there because the pathology is visible in the gains. Let the institutional output be an enforcement effort, u. Let one error, e_g, be the granular human injustice the institution is in principle obliged to answer — the particular person, wrongly or harshly dealt with. Let another error, e_t, be the teleological shortfall: the distance between the measured state of the system and the abstract mandate (no fraud, high throughput, low risk, fiscal recovery, predicted recidivism below a band). A sane controller uses both, with finite gains, and spends a great deal of its design on not letting the integral term wind up into over-actuation. The pathological controller does something else. Its granular gain K_g falls toward zero. Its teleological gain K_t rises without a matching bound, until the teleological channel behaves as a relay: full force or nothing, bang-bang, saturation.25 The dead-band on the granular channel widens at the same time. Disturbances smaller than a growing threshold produce no corrective act at all. The institution becomes blind to harm of ordinary size and violent about deviations from the mandate.

These are not two unrelated vices. They are one transfer function, splitting. As the representation claims more of the person, the measured granular error is the residual of the embedding, not the harm. Formally, if ρ is the degree of closure and s is the somatic load, the error the granular channel is allowed to see is something like (1 − ρ) s, passed through a dead-band whose width itself grows with ρ. Raising the nominal gain on that channel — hiring more reviewers, adding a "human factor" weight, commissioning a richer sensor — does not restore the act, once ρ is high, because the argument of the gain has been pre-shrunk to nearly nothing. The reviewer stares at a green dashboard. There is nothing in the protocol that counts as a tripwire. Meanwhile the teleological channel is fed an error the institution manufactures. Meeting a target ratchets the target. Closure itself is treated as a standing charge: the person is short of an abstract good merely because the mandate has tightened around her. High gain on that charge sits on the rail. Enforcement is saturated even when the dashboard is nearly green, because green is defined as yesterday's mandate.

The plates that accompany this article integrate a discrete map with exactly those couplings. It is a single plant, run in three regimes that differ only in how the middle term is coupled, not in the underlying disturbance. The reference seed is fixed, so the comparison is not an artefact of a friendlier noise. In the pathological regime, granular gain collapses toward zero, teleological gain runs to its cap, the dead-band occupies the late trajectory, and saturation of the commanded enforcement becomes continuous. Orthogonal debt — the accumulation of somatic load that closure has made unrepresentable — rises until it crosses a shear threshold. In the reference integration that crossing is at step 176. A second regime, rectification, boosts the nominal granular gain and slows closure, and it vents less of the exogenous signal. It still shears, later, at step 199, with the nominal granular gain still looking respectable at the end. The channel is nevertheless dead: the measured error has been eaten by closure, and the boosted gain multiplies a silence. A third regime arrests the embedding, partitions the pressure ledgers, and bleeds debt by a subtraction that is not proportional to the score. Debt plateaus near one half of the shear threshold. Closure cannot be continued past a fixed incompleteness. The teleological command never saturates. The exogenous ledger, unlike the pathological case, actually crosses the doctrinal threshold — and the debt turns down without a tear.

The numbers are properties of a stated map, not measurements of a country. They are reported so that the qualitative claim can be attacked at the level of mechanism. The claim is that sensor-boost inside a completable chart delays shear and does not forbid it, and that forbidding completion does. If a different map, with the same structure of couplings, produced the opposite qualitative result, the illustration would have failed and the article would have to say so. The map is published and is runnable.

What the dead-band means in a file is familiar to anyone who has tried to appeal a score that is "only one factor." The doctrine says the score is not determinative. The practice says that within a band around the score, no granular fact is large enough to move the outcome, because the protocol has defined ordinary facts as already priced in. The band widens as the institution's confidence in the representation grows. Confidence is not a mood. It is a narrowing of what is allowed to count as an error. Saturation, at the other pole, is equally familiar. Once a person is placed in a high-risk, high-fraud, or high-removal class, the available acts are extreme acts: full recovery, pre-trial detention, benefit stop, de-platforming, the abrupt contractual termination. Intermediate remedial acts have been defined out of the controller. The institution can whisper or it can strike. The pathology is to whisper to the particular and strike for the abstract, and to do both in the same week without experiencing a contradiction, because the gains are no longer on speaking terms.

Anti-windup, in ordinary control, is the art of not letting an integrator chase a setpoint the actuator cannot reach, lest the accumulated error discharge itself as a violent overshoot when the constraint lifts.26 Public institutions wind up. They accumulate a fiscal or political or managerial debt against an abstract mandate they cannot actually achieve at the level of lives, and they discharge it onto the bodies that the model has made eligible. The discharge feels like resolve. It is integrator windup. Classical anti-windup, however, will not save the arrangement, and this is where the analogy has to be prevented from suggesting a merely technical fix. Classical anti-windup modifies the controller inside its own state space. It is a smarter function of the same error signals. A smarter function of a dead granular error and a manufactured teleological error is a smarter closure. The bleed that prevents shear in the third regime is not a gain. It is a subtraction the controller is not allowed to set, because it is not allowed to price. The moment the bleed is brought inside the objective — the moment "complaints resolved" becomes a target, and the content of complaints becomes a training distribution — the subtraction becomes another induction machine, and the congestion of Part IV returns. That is why Part VIII will say that the derivative of the objective with respect to the interruption is ultra vires, not expensive. An expensive derivative is a price. Prices are what Goodhart's law eats.

Ross Ashby's law of requisite variety is the dignified ancestor of the hope that a sufficiently rich controller can regulate a human system.27 The good-regulator theorem says that every good regulator of a system must contain a model of that system.28 The theorem is a reason to model, and a warning about what a model is. A regulator that has replaced the system with the model is no longer regulating the system. It is regulating the model and letting the system absorb the difference. Requisite variety, read against the dyad of Part II, cannot be satisfied by variety inside the plain. The variety that has to be matched includes a pole that is not a state. No enlargement of the model's internal variety meets it. The enlargement is what ρ measures, and ρ is what kills K_g. More variety of the wrong type deepens the anesthesia. A ministry that knows a hundred new features of a claimant may be less able to see her than a clerk who knew six things and could still be ashamed.

VI. Konapsys, or Shear on a Metastable Plain

Metastability is the condition of a system that sits in a local basin, appears quiet, and is separated from a very different state by a barrier that accumulated disturbance can eventually cross. The quiet is real. It is not safety. A synthetically closed institution is metastable for a precise reason. On the plain, error is small: the metric meets the ratcheting target, the auditors sign, the human oversight samples a few files and finds the protocol followed. Orthogonal to the plain, debt accumulates, because the life the metric does not represent is still being acted on. The plain's shear strength is the legitimacy, the budget, the political coalition, the habitual obedience that lets a paper-thin representation bear the load of real sanctions. Strength of that kind is finite. It is also unmeasured, because the institution's instruments lie in the plain. The crossing, when it comes, is abrupt relative to the preceding calm. Files that were lawful on Friday are a scandal on Monday. The same facts, the same model, the same officials. What changed was not the arrival of a new principle. What changed was the load.

The name Konapsys is used here for that shear failure: the phase collapse in which a representation that had been treated as adequate is suddenly unable to carry the sanctions that were hung on it, and the discharge is discontinuous.29 It is not a revolution in the romantic sense, and it is not a market correction. It is a structural failure mode of brittle closure. The reference integration draws it as a vertical line on the time series of debt, coinciding with teleological gain already at its cap and granular gain already near zero. The phase portrait draws it as a trajectory that has climbed the closure axis until the debt coordinate hits the shear line, and then cannot remain on the slow manifold. Rectification, in the same portrait, climbs almost the same path and hits the same line later. Arrested embedding never takes the path. Its closure coordinate is constitutionally stuck, and its debt coordinate lives underneath the line. The portrait is the argument, compressed: the dangerous direction is not "more debt" in the abstract. Debt held at incomplete closure is the ordinary condition of a legal order that has not pretended to finish the person. Debt accumulated along a closure that tends to one is a tear waiting for a Thursday.

Two features of real ruptures match the geometry, and one feature limits the vanity of the match. The match: the pre-collapse period is characteristically a period of high reported performance and low granted remedy. SyRI, the Dutch system's risk-indication instrument for welfare and tax enforcement, was defended, until the District Court of The Hague struck its statutory basis down, as a proportionate tool surrounded by safeguards.30 The Court's finding under Article 8 of the Convention was that the legislation did not strike a fair balance, and that the measure was insufficiently transparent and verifiable. The performance claims did not save it, but they had been enough to keep it in operation while the orthogonal experience of the surveyed populations accumulated outside the justification. The childcare-benefits rupture has the same shape at larger human scale: years of indicator-success, then a collapse of political metastability so complete that a cabinet fell. The limit: not every scandal is a Konapsys, and not every Konapsys produces a just reconstruction. A shear releases energy. It does not select the next constitution. An institution can fail abruptly and be replaced by a harder closure with a new vocabulary. Nothing in the geometry implies that the far side of the tear is humane. That is why a theory of the rupture is not a politics, and why Part VIII does not propose to manage collapse. It proposes to stop storing the debt in a space the law has declared empty.

Prediction of the tear is the wrong ambition, and it is the ambition a synthetically closed institution will immediately adopt if told that tears exist. A rupture-prediction model is another ex ante bite. It profiles the population for proximity to scandal, and it treats the approach of shear as a quantity to be damped by the same controller that accumulated the debt — usually by suppressing the visible symptoms, which means a further widening of the dead-band. The debt is not thereby bled. The instruments are blinded. A ministry that monitors social-media sentiment for signs of "loss of trust," and feeds the sentiment back into message-discipline, has built a feedback loop around the symptom and left the orthogonal load untouched. The constitutional response is the opposite of prediction. It is a scheduled, non-discretionary exposure of the computational act to an interruption the model is not allowed to metabolize. Exposure is owed before the barrier is crossed, because after the crossing the terms are set by panic, and panic is bang-bang. Part V's saturation is what a controller does when it is surprised by a state it had defined as unreachable.

There is a temptation, inherited from a certain reading of resilience science, to recommend noise. If closure is brittleness, then randomness, redundancy, and deliberate inefficiency look like strength. They are, in mechanical systems that fail by fatigue. They are not a constitutional theory. Noise injected into a score is still a signal defined on the plain. It is a random variable with a distribution, and distributions can be written into a loss function. An institution told to randomize a fraction of its decisions will randomize the fraction that is cheapest to randomize, measure the disparity, and optimize the randomization schedule. The residue will not have been given standing. It will have been given a stochastic policy. Stochastic policy is still policy. The middle term is not re-opened by making the retraction jitter. It is re-opened by a relation the retraction is incompetent to absorb.

VII. Rectification Substitution, and Why Oversight Completes the Closure

The rights a data subject already has, in European law, are a serious attempt to keep a person from being processed as a mere object. Article 15 of the General Data Protection Regulation gives access; Article 16 gives rectification of inaccurate data; Article 22 gives a right not to be subject to certain solely automated decisions with legal or similarly significant effects, and, where such decisions are nonetheless taken under an exception, a right to human intervention, to express one's point of view, and to contest.31 The Court of Justice has held, in the SCHUFA scoring case, that the automated establishment of a probability value can itself be a decision within Article 22 where a third party draws strongly enough on that value that it determines the outcome.32 The holding matters. It refuses the formal dodge that the score is only a preparation and the decision happens later, in a human hand that was never free. The Artificial Intelligence Act's Article 14 requires high-risk systems to be designed so that natural persons can oversee them, understand their limitations, remain aware of automation bias, and override or interrupt them. Article 86 adds a right to an explanation of the role of the system in the decision and of the main elements of the decision.33 Read together, the Union's current text is the high-water mark of the rectification paradigm. It should be defended against those who would reduce it to a logo. It should not be asked to do a job it is not shaped to do.

Rectification, in the sense used here, is any remedy that presupposes the chart. Correct the feature. Explain the factor. Let a human review the output. Contest the score with evidence the model classifies as relevant. Each of these can be done well or badly. Done well, they reduce a particular kind of harm: the harm of being mis-described and then acted on through the mis-description. The right of non-integration is a different sentence. It says that for a specified class of decisions, the authority does not have the competence to constitute the person as a case-variable in the first place — not even accurately, not even with an explanation, not even with a human being in the chair. The human being in the chair is the critical case, because she is the alibi. If her institutional location is inside the loop, and the loop's reference signal remains the teleological mandate, her oversight is a component of K_t's legitimation, not a restoration of K_g. She is there to catch malfunctions of the representation. She is not there to deny that the representation is the right kind of reason. Automation bias is a real degradation of even that limited role, and Article 14 is right to name it. But an overseer cured of automation bias, bravely disagreeing with the machine on the machine's own features, is still a rectification device. She moves a coordinate. She does not hand back the chart.

Rectification substitution is the institutional act of offering the first right in discharge of the second. It is already the ordinary reply of well-lawyered deployers. You may access your data. You may correct an error. You may request a review. You may receive an explanation of the principal factors. Therefore the decision stands, and your deeper objection — that this decision was not the kind of thing a score was allowed to be — has been answered. The substitution is sincere more often than critics like to admit. Officials believe that openness and a human check are what a legal order owes. The belief is half right, and the half that is wrong is the half that closes the system. A hearing that can only produce a better input is not a hearing in Fuller's sense. It is a sensor. Sensors are what the pathological regime adds, in the rectification run of the model, without avoiding shear. The late granular channel is silent because there is nothing left in the measured error, not because the nominal gain was small. A large gain on a pre-emptied error is the mathematical portrait of a conscientious reviewer with an empty protocol.

The same substitution hides inside "human in the loop" as an organizational slogan. A loop is a closed path. To put a human in it is to place a living person on a closed path whose setpoint she does not choose. If she can stop the path, she is not in the loop; the loop is broken, and the break is the legally significant fact. If she cannot stop it, or can stop it only by recording a disagreement that the next iteration will consume as a label, she is a component. The diode of Part VIII is a way of saying which of these two designs is a legal order. Designs in which the human's interruption re-enters as training data are not a partial victory for oversight. They are the mechanism by which the middle term is digested. Every appeal becomes fuel. Every successful challenge teaches the model to avoid the surface features of a successful challenge, which is the SEO of remedy. The claimant who won because she found words the template did not expect has, by winning, supplied the template with those words. The next claimant will find the words priced in. This is not a paranoid fantasy about vendors. It is the ordinary logic of a learner whose objective includes not being reversed. A learner whose objective includes not being reversed will treat reversal as a negative label. The law, if it wishes the reversal to remain a legal act rather than a label, has to forbid the gradient.

Explanation has a parallel limit, and the scholarly fight about whether the GDPR contains a "right to explanation" should be remembered as a fight inside the rectification paradigm.34 An explanation of a score is a reconstruction of a path on the chart. It serves contestation of the path. It does not, by itself, open the question whether the chart was a competent authority for the decision. The useful explanation, on the theory of this article, is not a narrative of features. It is a statement of the legal interest, the primary facts, and the reason those facts support that action toward that person — a statement that would still be intelligible if the model had never existed. If the statement cannot be written without the model, the model was not aiding the reason. The model was the reason, and the reason is decoupled. Part X's review standard is built to detect exactly that condition, and to make it fatal to the decision rather than a prompt for a clearer technical annex.

Margot Kaminski's careful reconstruction of explanation as one piece of a system of contestation, rather than as a freestanding algorithmic entitlement, is closer to the bone.35 Contestability is the value. The disagreement is about what contestability must be allowed to do. If contestation can only correct, it is rectification, and it can be optimized. If contestation can nullify on the ground that the decision-maker completed a person it was incompetent to complete, it is jurisdictional, and it cannot be optimized without ceasing to be itself. Optimization assumes a scalar to improve. Jurisdiction is a wall. One does not optimize a wall. One stays on one's side of it or one does not. The case law reviewed in Part IX is, almost uniformly, an attempt to thicken contestation while remaining on the rectification side of the wall. The cases are right about the thickening. The wall is the next sentence, and it is the sentence the cases prepare but do not say.

VIII. Arrested Embedding and the Jurisdictional Diode

A constitutional bridge that tried to translate the undivided person into the plain would repeat the retraction it was commissioned to resist. The translation is the bite. So the bridge is not a channel of content. It is an arrest and a direction of conductance.

Arrest first. The embedding is not to be analytically continued across the person. In the model, this is the regime in which closure is pulled back toward an incompleteness it is not allowed to exceed. In the law, it is a set of decisions for which constitutive quantification is ultra vires: sentencing and pre-trial restraint, the grant or withdrawal of subsistence benefits, removal or refusal of a child, exclusion from education or from a profession, termination of employment, refusal of a core financial service where the refusal determines access to ordinary life, and the silencing of a natural person by a platform in a way that amounts to an exclusion from a principal public forum. The list is a schedule, not a metaphysics. It marks the decisions on which a political community has already decided that the cost of being wrong is not a cost that a loss function is allowed to price, because the person is not a residual risk. Outside the schedule, computation remains what it is good at: allocation under conditions the law has already simplified, detection of patterns offered to a human judgment that has not been pre-empted, infrastructure. The schedule is how the theory stays narrower than "AI governance."

The diode is the direction of what may still flow. From the computational act toward the person, conductance is mandatory. Every such act remains exposed to interruption: a refusal to be enrolled as a case-variable before the score is run; a stay; a deletion of the operative score from the file; an in-person re-hearing on a record the model does not define; a nullification. Exposure is owed ex ante where enrollment is the harm, not only ex post as an appeal against a finished act. An appellate diode is a eulogy. The teleological gain has already saturated in real time, and the remand writes a careful opinion about a life that has already been bent. From the person toward the computational act, content does not conduct. What crosses back is the legal effect — halt, remand, quash, declare — as an act in the register of law. The qualitative content of the hearing, the narrative that would be so valuable as a label, the tone of distress that would be so valuable as a feature, do not cross. They are privileged against re-ingestion. They may sit in an archive, because reconstructibility is a duty of justice: a later court, a later journalist, a later parliament must be able to see what was done. Archive is not a training set. The distinction is constitutional, not technical, and it is not observed by anonymization. An anonymized transcript that updates weights has re-entered. The person has been digested under a thinner name.

The one-way membrane is the only topology that keeps complex reciprocity. If content conducts in both directions, the residue is eaten, and the middle term collapses back into the section. If nothing conducts in either direction, the poles fall into dual autism: a computational order that never meets a resistance, and a life that never receives an answer. That dual closure is the precondition of the shear, because it is how debt accumulates with no bleed and no doctrine. The diode is a relation without fusion. The institution sends a score. The person does not return a corrected score. She returns a discontinuity. The discontinuity dies as information at the membrane and survives as law. On the far side of the membrane, the legal order has changed: the decision is void, or the enrollment is forbidden, or the official is enjoined. On the near side, the model's parameters are as they were. The frustration this causes to engineers is the doctrinal point. A system that cannot learn from the cases that defeat it will be, by the standards of a learner, worse. It will be, by the standards of a legal order, capable of losing. The capacity to lose to a person, and to have the loss remain a loss rather than a lesson in how not to lose next time, is the internal point of view restored to the remedy.

Who guards the membrane? If the answer is an expert office that certifies the residue as authentically non-computable before it may be heard, the office becomes the new middle term and will close synthetically, staffed by the professions that know how to speak to it. Expertise may inform a court. It cannot be the condition of standing. Standing is original to the person. She does not have to prove that she is ineffable. She has to be a natural person within the scope of a decision on the schedule, or a natural person who is in fact being completed by a computational act, which she can show by showing the act. Ineffability as a threshold of jurisdiction would be the SEO of metaphysics: only the claimants whose representatives can perform mystery would get through, and the performance would immediately be coded. The statute therefore refuses any certification requirement. It also refuses any duty, on the claimant's side, to propose a better model. A better model is a rectification offer. The claimant is not the vendor's unpaid research department.

Two hard points follow, and a theory that dodges them is a mood. The first is temporal. A transcript barred from today's objective function will try to re-enter through a research exemption, a fairness benchmark, a parliamentary evidence base, an insurer's table, a subsequent vendor's "lessons learned" white paper. The ban has to run with the act of interruption, against any later use of its content as a feature, a label, a weight, or an evaluation item, by the original controller and by any successor who receives the content from it. Research may study the legal effects. It may count how many decisions were nullified, of what type, with what statutory ground. It may not drink the narratives. The moment counting the grounds becomes a matter of fine-grained coding of the stories, the research has crossed back. Line-drawing here will be unglamorous and litigious. That is what a membrane looks like when someone is pressing on it. The alternative is a principle that binds only those too unsophisticated to route around it.

The second hard point is the difference between procedure and jurisdiction, and it is the point on which this article is prepared to be narrow and stubborn. A procedural duty to consider the person carefully can be optimized. Consideration has a duration, a checklist, a quality score, a sampling rate. All of those can enter an objective function, and an institution under pressure will meet them at the lowest cost consistent with not being reversed on procedural grounds. Procedural proportionality is real, and it is already the grammar of Article 8 case law and of constitutional review of automated analysis.36 It is not the bridge. The bridge is a denial of competence. The decision-maker does not have the power to complete the person, and a flawless procedure in the service of a completion is still ultra vires. Procedure governs how a competent authority acts. Jurisdiction decides whether this speaker is that authority for this act. Synthetic closure's great institutional achievement is to have made a competence look like a procedure — "we followed the human-oversight protocol" — so that the review court asks whether the protocol was adequate rather than whether the act was of a kind the speaker could perform. Part IX will show the courts approaching the second question and settling for the first. Part X writes the second question into the operative text, as a ground of nullification rather than as a factor in a balance.

A word on private power, because the schedule reaches employers, creditors, and platforms, and a reader trained on classical constitutionalism will suspect a confusion of rights against the state with duties on everyone. The confusion is older than machine learning. When a private ordering binds a natural person in a decision of the scheduled kind, it is already exercising a power that the legal order may or may not have delegated. The diode does not constitutionalize the whole of private law. It says that the delegation does not include the competence to replace the middle term with an embedding, any more than a state agency can be delegated that competence. A creditor who refuses credit on an underwriter's judgment, ugly and human and answerable, is inside ordinary private law, with whatever constraints fair-credit legislation already imposes. A creditor who refuses credit because a probability value, which the person was not entitled to refuse as the form of the decision, determined the refusal, is inside SCHUFA's facts and inside this statute's ground. The line is not "was software used." Underwriters use software. The line is whether the computational output constituted the person and determined the act. That is a fact a court can find. It is the fact the Court of Justice has already begun to find.

IX. What the Courts Have Almost Seen

The cases worth reading are the cases that felt the retraction and reached for the tools they had. The tools they had were transparency, balance, purpose limitation, and the prohibition on letting a hidden machine be the real decision-maker. Each tool is a way of saying that something about the coupling is wrong. None of them quite says that completion is incompetent. The distance between those two sentences is the distance this article is trying to travel, and it is an honest distance: a court that said the second sentence without a statute, in a system built on proportionality, would be inventing a competence-wall in a jurisprudence of weights. The statute is how the wall stops being a metaphor.

In State v. Loomis the Supreme Court of Wisconsin accepted that a proprietary recidivism instrument had been used at sentencing, declined to treat the proprietary character of the weights as a due-process violation, and held that the use was permissible where the sentence was not based solely on the instrument and the sentencing court was warned of its limitations.37 Read generously, Loomis is an attempt to keep the score in the role of a factor and the judge in the role of a judge. Read against the dead-band, it is the doctrinal ratification of a widened band. "Not solely" is satisfied whenever the judge mentions something else. The something else is not required to be sufficient. The warnings are decoupled rationales of a particularly refined kind: the court recites that the tool is limited, and the recital fills the accumulator, and the sentence proceeds. The defendant cannot inspect the instrument, so he cannot put granular fact into the only channel the court has left open, which is a quarrel about weights he cannot see. K_g is structurally zero at the point of decision, and the opinion treats that zero as compatible with due process because K_t — public safety, risk, the abstract mandate — is being served and a human is nominally in the chair. Loomis is the rectification paradigm at the moment it gives up on rectification itself, since the coordinate cannot even be checked, and settles for a warning label on the chart.

The federal court in Houston Federation of Teachers faced a different embarrassment.38 Teachers were evaluated, and could lose their livelihoods, by a proprietary value-added model they could not meaningfully test. The court was prepared to see a procedural due-process problem in a system that denied any fair chance to challenge the determination. The instinct is Fuller's. The remedy it points toward is still a better opportunity to contest the score: expert access, interrogatories, a hearing at which the model can be examined. Those are essential where the legal order has already decided that a model of this kind may determine a career. They do not ask whether a value-added embedding of a teacher's practice is a competent way to decide dismissal. A teacher who is given every weight and still told that her practice is the residual of a regression has been granted rectification and denied the middle term. She can quarrel with the variables. She cannot require the district to judge her teaching.

SyRI is the European judgment that comes closest to seeing congestion as such. The District Court of The Hague held that the statutory scheme for systemic risk indication failed the fair-balance requirement of Article 8 of the Convention, emphasizing the opacity of the risk model and the difficulty of contesting one's inclusion in a process one cannot see.39 The State had argued that safeguards were sufficient and that the social interest in fraud detection was weighty. The Court refused to treat that weight as self-justifying. What the Court did not have, inside a proportionality frame, was a sentence of the form: this is not a balance to be struck, because the prior act — constituting neighbourhoods and lives as risk-masses for an objective the subjects cannot exit — is not an act the statute could authorize. Proportionality presupposes a competent aim and asks whether the means were excessive. Synthetic closure often has a competent-looking aim. The excess is not in the intensity. The excess is in the kind of act. SyRI's opacity made the intensity unreviewable, and the Court struck the scheme on that ground, which was the ground available. The deeper defect would have survived a transparent SyRI. A fully explained risk indication that determined intrusive interventions would still have been a completion of the persons indicated. Transparency would have improved the archive. It would not have installed the diode.

The Court of Justice in SCHUFA took the step Loomis would not take, at the level of classification. A probability value that plays a determining role is the decision, not a preparatory remark, and Article 22 therefore applies.40 The consequence, inside the GDPR, is the set of safeguards in Article 22(3): human intervention, the data subject's view, contestation. The Court forced the score out of the shadow and into the category "decision." It then handed the person the rectification toolkit. That is as far as the Regulation reaches, and it is a long way. The toolkit does not include a right to say that creditworthiness, where it determines access to ordinary economic life, may not be constituted as a probability value at all, however carefully a human reviews the value. One can believe that such a right would be too broad, that credit is exactly the domain where probabilistic underwriting is legitimate, and that the scheduled right of non-integration should not extend to ordinary consumer credit. That is a quarrel about the schedule. It is not a quarrel about the structure. SCHUFA shows the structure working as designed: identify the real decision, then remedy it inside the representation. The statute's disagreement is not with the identification. It is with the idea that identification plus safeguards is a full account of what the person is owed when the decision is of a kind that constitutes her.

The Federal Constitutional Court's 2023 judgment on automated police data analysis in Hesse and Hamburg is the clearest statement yet that a further processing by automated analysis is a fresh interference, with its own weight, not a harmless rearrangement of data already held.41 The Court required, for severe uses, a sufficiently concrete danger and a statutory design that limits what may be combined and for what. Purpose limitation is a cousin of the anti-congestion principle: data collected for one relation may not silently become the fuel of another. The cousinage should be acknowledged and not overclaimed. Purpose limitation keeps embeddings inside their warrants. It does not deny that an embedding can be the form of a decision. A warrant wide enough, and a danger concrete enough, still leave the person as a node in a graph the analysis is free to complete. The judgment is a powerful restraint on dragnet completion in the security field. It is written in the grammar of proportionality and statutory clarity, which is the grammar a constitutional court uses when the legislature might still authorize the act more carefully. Ultra vires completion, as this article means it, survives careful authorization. A clear statute that says "you may complete the person" is clearer and worse.

The English Court of Appeal in Bridges, reviewing live facial recognition by a police force, held that the legal framework was inadequate: too much discretion about who was placed on a watch-list and where the cameras were deployed, and an insufficient data-protection assessment.42 Again the court reached for constraints on deployment, not for a categorical incompetence. The judgment is right about discretion. An unguided power to enroll faces in public is a power of constitutive quantification exercised against everyone in the frame, most of whom are not suspects in any sense a court would recognize. Bridges is the ex ante problem in a single image. The enrollment happens before any "decision" about a person, and by the time there is a decision the chart has already been drawn. A review standard aimed at the downstream use arrives late. The statute's refusal of enrollment, for scheduled decisions and for any enrollment whose purpose is to determine such a decision, is the ex ante edge Bridges implies and does not quite state. The Court was asked to review a deployment under existing powers. It found the powers too vague. A legislature that answered Bridges only by writing a more precise deployment policy would have produced a clearer closure.

Taken together, the cases do not yet amount to a doctrine. They amount to a pressure, and the pressure is the right kind: inductive, particular, embarrassed by the distance between the reason given and the person bound. The embarrassment has been formulated as opacity, as proprietary secrecy, as missing safeguards, as purpose-creep, as unguided discretion. Those formulations are the sediment. This article's claim is that the sediment is deep enough to cut a different channel, and that if the channel is not cut deliberately the accumulator will fill with the very safeguards the cases demanded — impact assessments, oversight logs, explanation templates, human-review attestations — and the next rupture will find the law once again surprised by a debt its own remedies helped to hide. Safeguards congest when they are allowed to count as the exogenous pressure they were meant to transmit. The statute's partition of ledgers is the practical consequence of that observation. The endogenous record of compliance with the safeguards is not the pressure. The person's interruption is the pressure. A legal order that cannot tell those two books apart will celebrate its compliance on the morning of the shear.

X. Ultra Vires Completion

The operative doctrine is short. A decision that binds a natural person is void if it completes her. Completion is the constitution of the person as a case-variable within a computational representation, where that constitution determines the act. The speaker — public or private, whenever the schedule or the actual effect brings the act within scope — lacks competence to perform that constitution. The lack is not cured by accuracy, by explanation, by human oversight, by the data subject's opportunity to rectify features, by conformity to a technical standard, or by the speaker's own assessment that the system is dependable. Those facts may be entered in the endogenous record. They are legally inert as justifications of the completed act.

Dependability is the name the presumption will bear in the statute, because it is the name the practice already uses. Systems are bought and defended as dependable: validated, robust, widely deployed, certified, "state of the art." Dependability is a property of tools. The presumption of dependability is the illicit inference from "this tool is dependable" to "this decision is therefore supported." The inference is the retraction of the middle term, compressed into a single evidential step. Abolishing it does not abolish tools. It abolishes the step. No court, tribunal, or administrative authority may treat a computational output as prima facie evidence of a fact about a natural person, or as a sufficient reason for a decision that binds her, on the ground that the output is internally consistent, statistically validated against its own objective, attested by a vendor, widely used, or conformant to a standard. The authority may use the output as a clerk uses a calculator: to perform an operation the authority has already defined, on inputs the authority is prepared to defend without the calculator's help. The moment the output is asked to supply the reason, the presumption has returned, whatever the authority calls it.

Decoupled rationales are inadmissible as the sole or principal justification of a binding decision. A rationale is decoupled when it justifies the act by pointing at a metric, a score, a threshold, a model objective, an institutional target, or a compliance artefact, and does not reconnect that quantity to the protected interest and the primary facts. The standard of review is correctness on the reconnecting step, not rationality, not Wednesbury, not a respectful look at a technical margin. There is no margin on the question whether a reason is a reason. The burden sits on the authority, because the authority is the author of the representation and the person cannot be asked to prove a negative inside a chart she does not control. If the authority cannot state the reason without the output, it has no reason. The decision falls. Deference to the model as a fact-finder is withdrawn, because the model is the speaker's own artefact, and a speaker does not earn deference by automating the speech.

Pleading rules are where congestion is reproduced case by case, so the statute includes a safe harbor that will look, to a certain professional sensibility, like a protection of bad drafting. A filing is not to be rejected, delayed, down-ranked, or treated as incomplete because it contains narrative variance, because it contradicts a score in words the score cannot encode, or because an intake system cannot parse it. Rejection requires a legal ground independent of machine readability. A system that scores pleadings for "strength" or "fraud likelihood" or "likelihood of success," and routes the person's access to a hearing through that score, is performing the SEO of law as an official function, and the safe harbor makes that routing unlawful. The point is not to romanticize incoherence. The point is to stop the anteroom from sanding off the only friction the bedrock was ever going to receive. Costs protection attaches where the sole defect relied on was optimization-resistance, so that representatives are not punished, in the currency of fees, for refusing to pre-digest their clients.

Remedies follow the jurisdictional character of the wrong. The primary remedy is nullification of the completed act, not damages calibrated to how far the score deviated from a truer score. Damages of that sort are rectification in money. They teach the controller to improve the scorer. Nullification teaches the controller that it walked off a cliff of competence. Structural injunctions are available where the completion is systemic: a court may forbid the continued use of a named constitution of persons for a named class of decisions, and may forbid re-ingestion of the hearing record that produced the injunction. Interim relief is available before harm, on a showing that enrollment is imminent and that the enrollment is for a scheduled decision. The person does not wait to be ruined in order to acquire standing to say that the ruining was of the wrong kind. And because granular restoration cannot be a gain on the system's own error signal, the statute provides a non-proportional reset: where a scheduled decision has been completed, the person's interruption sets the act aside without any requirement that she demonstrate a large enough deviation from the score. Proportionality to the score is the teleological gain. Equity, at this seam, is non-proportional because proportionality has been captured by the chart.

The partition of ledgers is stated as a rule of legal effect rather than as a metaphor about pressure. An endogenous record — validation studies, oversight logs, bias audits, vendor assurances, internal targets met — may be required, and may be published, and may not by itself satisfy the condition for changing a natural person's legal position. An exogenous record — the person's interruption, the primary facts she puts in issue, the findings of a court that is not the author of the model — is the record on which that condition can be satisfied. The two are not fungible. An authority that answers an exogenous interruption by adding to the endogenous record has not answered. It has congested.

Standing is original. No expert pre-clearance. No duty to show that one's situation is unmodellable in principle. No duty to propose an alternative architecture. Associations may support a claimant; they do not replace her, except in the familiar cases where collective proceedings are the only realistic vehicle and the natural persons affected remain identified as a class of persons rather than as a distribution. The statute is written so that it can be enacted as a model act beside existing data-protection and administrative-procedure statutes, not as a replacement for them. Rectification rights remain. They handle mis-description. Non-integration handles completion. The interpretive clause says, in terms an administrator cannot misunderstand, that compliance with the GDPR and with the AI Act's oversight and explanation duties is relevant to the endogenous record and is not a defense to a claim of ultra vires completion. That clause will be called gold-plating. It is the whole point of writing a statute rather than a commentary. If compliance with the rectification paradigm were a defense, rectification substitution would be the safe harbor, and the statute would have re-enacted the pathology in the vocabulary of its critics.

What the doctrine does not do should be stated with equal coldness. It does not forbid statistical inquiry. It does not forbid a judge to know a base rate, or a physician to use a decision-support tool that does not determine the act, or a tax authority to add a column of numbers with a computer. It does not require a return to unreviewable official intuition, which has its own dead-band and its own saturation and a longer historical title to both. Intuition can complete a person too; it does so without an archive. The diode's discipline on computational acts is not a compliment to unassisted power. Unassisted power remains subject to the ordinary law of reasons, evidence, and review. The specific vice of synthetic closure is not power. It is power that has retracted the relation in which power could be answered, and that has done so under the banner of being more answerable than a human. The doctrine answers the banner.

XI. Conclusion

The middle term is not a nostalgic name for the human touch. It is the relation that holds a legible order and an undivided person in the same legal act without declaring them the same kind of thing. Synthetic closure is what happens when the relation is replaced by a section of the legible order alone. The institution does not thereby become silent. It becomes eloquent in a language that refers only to itself, and it enforces that language with a gain that has split: anesthetic to the particular, saturated against deviations from the abstract. The threshold at which law, as terrain, would have moved under the weight of particular lives is filled in advance by the institution's own reasons. Debt stored off the plain eventually shears it. The rupture is not a delivery system for justice. It is a failure mode. A legal order that wishes to avoid it, and that wishes to deserve to avoid it, has to stop treating the person as a removable singularity.

The arrest is jurisdictional. The conductance is a diode. The person must be able to interrupt the computational act, including before it enrolls her, and the content of the interruption must not become the act's next lesson. Rectification remains a right and becomes a trap the moment it is offered in discharge of the right not to be completed. Courts have already seen pieces of this and have answered them with transparency, balance, and safeguards. Those answers are the sediment. They are also, if allowed to count as the whole of the exogenous pressure, the next congestion. The statute that accompanies this article is a proposal for cutting the channel on purpose: abolish the presumption that a dependable system yields a supported decision; give awkward pleading a harbor against the anteroom; review decoupled reasons for correctness, with the burden on the speaker who automated the speech. None of this makes a society computable. It keeps a society capable of losing an argument to a person who is not a vector in it. That capacity is the internal point of view, and it is the only point of view from which a legal order is distinguishable from a very articulate controller.

Notes

1The baseline problem is the central caution of the technical literature and the least interesting fact about it. A system can be parity-constrained against one error rate and disparate against another. See the discussion provoked by Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner, ‘Machine Bias’ ProPublica (23 May 2016), and the subsequent exchange on calibration and error-rate balance. The present article takes no side in that exchange. A perfectly balanced completion of a person would remain a completion.

2Danielle Keats Citron, ‘Technological Due Process’ (2008) 85 Washington University Law Review 1249; Danielle Keats Citron and Frank Pasquale, ‘The Scored Society: Due Process for Automated Predictions’ (2014) 89 Washington Law Review 1. Bluebook: Danielle Keats Citron, Technological Due Process, 85 WASH. U. L. REV. 1249 (2008); Danielle Keats Citron & Frank Pasquale, The Scored Society: Due Process for Automated Predictions, 89 WASH. L. REV. 1 (2014).

3Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation) [2016] OJ L119/1, art 22.

4Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) [2024] OJ L, 2024/1689, arts 14 and 86.

5Lon L Fuller, ‘The Forms and Limits of Adjudication’ (1978) 92 Harvard Law Review 353. Bluebook: Lon L. Fuller, The Forms and Limits of Adjudication, 92 HARV. L. REV. 353 (1978).

6HLA Hart, The Concept of Law (3rd edn, OUP 2012) 88–91.

7Ronald Dworkin, Law’s Empire (Harvard UP 1986).

8Jeremy Waldron, ‘The Rule of Law and the Importance of Procedure’ in James E Fleming (ed), Getting to the Rule of Law (NYU Press 2011).

9Niklas Luhmann, Law as a Social System (Klaus A Ziegert tr, OUP 2004).

10Gunther Teubner, Law as an Autopoietic System (Blackwell 1993).

11The point is cousin to, and not identical with, the critique of legibility in state administration. Computational legibility does not need a state. It needs an objective.

12Gilbert Simondon, Individuation in Light of Notions of Form and Information (Taylor Adkins tr, University of Minnesota Press 2020).

13Maurice Merleau-Ponty, Phenomenology of Perception (Donald A Landes tr, Routledge 2012).

14Aristotle, Nicomachean Ethics (Terence Irwin tr, 2nd edn, Hackett 1999) 1106a26–b7.

15Charles S Peirce, ‘A Guess at the Riddle’ (1887–88) in Nathan Houser and Christian Kloesel (eds), The Essential Peirce, vol 1 (Indiana UP 1992) 245–279.

16Tweede Kamer der Staten-Generaal, Parlementaire ondervragingscommissie Kinderopvangtoeslag, Ongekend onrecht (17 December 2020) Kamerstuk 35510, nr 2.

17Aristotle, Posterior Analytics (Jonathan Barnes tr, 2nd edn, OUP 1994) I 6, I 13.

18Aristotle, Nicomachean Ethics (n 14) 1141b8–1142a30.

19ibid 1132b21–1133b28.

20Citron (n 2); Citron and Pasquale (n 2). See also Frank Pasquale, The Black Box Society: The Secret Algorithms That Control Money and Information (Harvard UP 2015); Mireille Hildebrandt, Smart Technologies and the End(s) of Law: Novel Entanglements of Law and Technology (Edward Elgar 2015).

21Ongekend onrecht (n 16). The parliamentary report remains the primary public reconstruction. The point in the text is structural and does not depend on any single factual finding beyond those the report established to the satisfaction of the chamber that received it.

22Donald T Campbell, ‘Assessing the Impact of Planned Social Change’ (1979) 2 Evaluation and Program Planning 67; Charles Goodhart, Monetary Theory and Practice: The UK Experience (Macmillan 1984).

23István Bibó, ‘The Miseries of East European Small States’ in Iván Zoltán Dénes (ed), The Art of Peacemaking: Political Essays by István Bibó (Yale UP 2015).

24Karl Johan Åström and Tore Hägglund, PID Controllers: Theory, Design, and Tuning (2nd edn, Instrument Society of America 1995). The map integrated in this article is not offered as a tuned controller for any plant. It is a qualitative coupling: granular gain collapsing with closure, teleological gain saturating, debt integrating off the embedding.

25On saturation and relay behaviour, see ibid. The reference integration caps teleological gain and treats commands beyond the cap as saturation. In the pathological regime the cap binds, and the late trajectory is saturated; granular actuation is simultaneously inside the dead-band.

26ibid.

27W Ross Ashby, An Introduction to Cybernetics (Chapman & Hall 1956). See also Norbert Wiener, Cybernetics, or Control and Communication in the Animal and the Machine (2nd edn, MIT Press 1961).

28Roger C Conant and W Ross Ashby, ‘Every Good Regulator of a System Must Be a Model of That System’ (1970) 1 International Journal of Systems Science 89.

29The term is a term of art of this project, not a borrowing from mechanics, though the mechanical metaphor of shear is deliberate. The reference integration places the pathological crossing at step 176 and the rectification crossing at step 199, under seed 11. The diode regime does not cross. Horizon 280. These are properties of the published map.

30Rechtbank Den Haag 5 February 2020, ECLI:NL:RBDHA:2020:1878 (NJCM and others v The State of the Netherlands) (SyRI); Dutch text ECLI:NL:RBDHA:2020:865, case C/09/550982 / HA ZA 18-388.

31GDPR, arts 15, 16 and 22.

32Case C-634/21 OQ v Land Hessen (SCHUFA Holding (Scoring)) EU:C:2023:957 (7 December 2023).

33AI Act, arts 14 and 86.

34Sandra Wachter, Brent Mittelstadt and Luciano Floridi, ‘Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation’ (2017) 7 International Data Privacy Law 76; Lilian Edwards and Michael Veale, ‘Slave to the Algorithm? Why a “Right to an Explanation” Is Probably Not the Remedy You Are Looking For’ (2017) 16 Duke Law & Technology Review 18. Bluebook: Lilian Edwards & Michael Veale, Slave to the Algorithm? Why a “Right to an Explanation” Is Probably Not the Remedy You Are Looking For, 16 DUKE L. & TECH. REV. 18 (2017).

35Margot E Kaminski, ‘The Right to Explanation, Explained’ (2019) 34 Berkeley Technology Law Journal 189. Bluebook: Margot E. Kaminski, The Right to Explanation, Explained, 34 BERKELEY TECH. L.J. 189 (2019).

36Convention for the Protection of Human Rights and Fundamental Freedoms (European Convention on Human Rights), art 8; S and Marper v United Kingdom (2009) 48 EHRR 50.

37State v Loomis 881 NW 2d 749 (Wis 2016). Bluebook: State v. Loomis, 881 N.W.2d 749 (Wis. 2016).

38Houston Federation of Teachers, Local 2415 v Houston Independent School District 251 F Supp 3d 1168 (SD Tex 2017). Bluebook: Houston Fed’n of Teachers, Local 2415 v. Hous. Indep. Sch. Dist., 251 F. Supp. 3d 1168 (S.D. Tex. 2017).

39SyRI (n 30).

40SCHUFA (n 32).

41BVerfG, Judgment of the First Senate of 16 February 2023, 1 BvR 1547/19, 1 BvR 2634/20 (automated data analysis). The headnotes treat automated analysis or interpretation of stored data as an interference in its own right with informational self-determination under art 2(1) in conjunction with art 1(1) of the Basic Law, with a weight that is not reducible to the original collection.

42R (Bridges) v Chief Constable of South Wales Police [2020] EWCA Civ 1058.