An excerpt

Executive thesis

For most of modern institutional history, decisions were treated as legitimate because they came from authorized sources. That arrangement breaks once consequential decisions are produced by systems whose reasoning cannot be reconstructed.

Credit. Employment. Benefits. Insurance. Healthcare. Security. Administrative governance. The list is no longer hypothetical. Institutions already rely on machine-mediated determinations in each of those domains. Most cannot replay those determinations after the fact. The records of the outputs exist. The procedural ground does not.

This is not a technology problem. It is a procedural and constitutional one, sitting one institutional layer beneath the vocabulary currently used to describe it.

The manuscript proposes a replacement standard. Decision admissibility.

A consequential automated decision must become admissible in the same sense evidence becomes admissible inside a court.

The admissibility criteria

Reconstructable — on the inputs that produced it.

Reviewable — by an examiner outside the institution that issued it.

Replayable — on the model state, parameters, and pipeline that produced it.

Contestable — on procedural grounds, by the party it affects, after the fact.

The criteria are not separate questions. They are facets of the same question, which is whether the decision survives adversarial reconstruction. The current governance vocabulary — alignment and explainability and audit and transparency and safety — performs partial work while leaving that question untouched.

From Chapter Two

The Age of Assumed Truth

An excerpt from the opening of Chapter Two. The chapter — and the rest of the book — continues behind the wall.

§1.

In the spring of 1968, in a fourth-floor walk-up apartment on the Lower East Side of Manhattan, a twenty-year-old man named John Kelly received a letter from the New York City Department of Social Services. The letter informed him that his Home Relief benefits, the general-assistance program that paid his rent and bought his groceries, had been terminated. The reason given was administrative: he had moved from a city-assigned hotel into a friend's apartment without notifying his caseworker, and the move was treated as a violation of the program's residence requirements. Kelly went to the Mobilization for Youth Legal Services storefront on Avenue D, where attorneys had been representing welfare recipients in disputes with the city for several years. The attorney who took his case, Lee Albert, would spend the next two years arguing that the termination's specific mechanism, adverse action issued without prior notice or opportunity to be heard, violated the Fourteenth Amendment's Due Process Clause.

Two years later, on March 23, 1970, the Supreme Court of the United States decided Goldberg v. Kelly. The case involved Kelly and a group of fellow New York City residents whose public-assistance benefits had been terminated by state caseworkers without prior notice or opportunity to contest the termination. They had filed suit arguing that the terminations violated the Due Process Clause of the Fourteenth Amendment. The District Court had agreed. The state had appealed.

Justice Brennan, writing for the majority, affirmed. The opinion held that welfare benefits, while not property in the traditional sense, were a statutory entitlement, the product of specific legislative enactments that conferred specific benefits on specific classes of people meeting specific eligibility criteria. The entitlement, once conferred, could not be revoked by the state without the procedural protections the Due Process Clause required. Those protections, at a minimum, included timely notice of the proposed termination, opportunity to be heard before an impartial decision-maker, the right to present evidence and confront adverse witnesses, and a decision rendered on the record.

The opinion did not invent these requirements. Notice, hearing, impartial adjudication, and reasoned decision-making were, by 1970, substantively settled in American administrative law. What Goldberg did was extend them, explicitly, unambiguously, to the administration of welfare benefits, which had previously been treated as a matter of agency grace rather than legal entitlement.

Justice Brennan's opinion is a document of a particular moment in the American mind, the moment when the expansion of the administrative state had proceeded far enough that the Bill of Rights needed to be read as attaching to the new kinds of interests the state was now regulating. Welfare was not property in the common-law sense. It was property in the due-process sense, an expectation the government had created and on which recipients had organized their lives. The move Brennan made was not doctrinal invention. It was doctrinal recognition: the Constitution's procedural guarantees had to follow the state into the domains where the state was now acting. That move was made in 1970. It has to be made again, for the domain in which the state, and increasingly private actors standing in for the state, is now acting through automated decisions.

The opinion was celebrated as a landmark. In the decades since, it has been cited in thousands of subsequent decisions across the full range of administrative adjudications, benefits terminations, licensing revocations, public-employment dismissals, educational disciplinary proceedings, prison-discipline cases. Its core principles remain controlling law.

The plaintiffs in Goldberg were residents of New York City whose public-assistance benefits had been terminated by state and city agencies without prior notice. The lead plaintiff, John Kelly, was twenty years old, a recipient of New York's Home Relief general-assistance program whose benefits had been cut off after he left a city-assigned hotel for a friend's apartment. His caseworker had treated the move as a violation of the program's residence rules and terminated his benefits without a hearing. The class the District Court certified included recipients of Aid to Families with Dependent Children whose benefits had been similarly cut off, many for reasons of caseworker error or administrative convenience, all without the opportunity to contest the termination before it took effect.

The suit was brought by Mobilization for Youth Legal Services, a Manhattan-based legal-aid organization whose work on behalf of welfare recipients was part of the broader War on Poverty legal-services movement. The attorneys, David Diamond, Lee Albert, and others whose names appear in the case record, argued that the terminations violated the Due Process Clause because benefits, once granted, were the foundation of the recipients' subsistence and could not be revoked by administrative fiat without the procedural protections that had long been required in other deprivations of property and liberty.

The District Court, Judge Wilfred Feinberg writing for a three-judge panel, agreed. The state of New York appealed to the Supreme Court. The state's argument was simple: welfare benefits were not property in the constitutional sense; they were public largesse, conferred at the state's discretion, and revocable on the same terms. The Due Process Clause, on the state's reading, did not reach them.

The Court split five to three, Justice Marshall recused himself, with Brennan writing for the majority. The opinion was twenty-five pages long. It traced the history of the due-process doctrine. It acknowledged the state's interest in efficient administration of public-benefits programs. It held, after working through the interest-balancing calculus the Court had been developing in the preceding decade, that the procedural protections required in the termination of welfare benefits included notice, a meaningful pre-termination hearing, the right to present evidence and cross-examine adverse witnesses, and a decision based on the hearing record. Justice Hugo Black dissented, arguing that the majority had unduly extended judicial review over administrative process. Justice Harlan joined a separate dissent. Justice Stewart concurred in part. The five-justice majority held.

…

The chapter goes on from here — the holding in Kelly, the procedural order it built, and why the model of assumed truth it secured is the one machine-mediated decisions have quietly broken. It continues in the full manuscript.

Unlock the full book Back to the book

From Chapter Four

Why the Old Fixes Fail

An excerpt from the opening of Chapter Four. The chapter — and the rest of the book — continues behind the wall.

I. The Six Substitutes

By 2024 a vocabulary had congealed around the problem described in Chapter 3. Policymakers, academics, corporate counsel, standards bodies, and journalists spoke a shared dialect. Its six load-bearing words were alignment, explainability, audit, regulation, litigation, and, beneath them all, a political settlement that treated AI governance as an extension of existing administrative practice.

Each word named something real. Each had adherents who did honest work. None of them, taken alone or together, closes the gap opened by the three breaks.

This chapter is not a survey. It is a demolition. Each of the six is shown insufficient in compressed form, not because the people working on it are wrong, but because the concept itself cannot bear the weight being placed on it. The six together form a vocabulary. The vocabulary cannot carry the load. What comes after them, the positive standard developed in Chapter 5, is not an improvement on the vocabulary. It is a different kind of answer to a different kind of question.

Michael Oakeshott, in a 1947 essay that has become unfashionable to cite and difficult to replace, distinguished technical knowledge, the kind captured by rules, manuals, and procedures, from practical knowledge, the kind captured only by doing. The characteristic error of modernity, on Oakeshott's view, is the rationalist's confusion: treating the manual as if it were the craft. Each of the six substitutes this chapter addresses commits a version of Oakeshott's error. Each captures some technical feature of accountability, a training procedure, an explanation method, a compliance check, a statutory rule, a litigation pathway, an administrative process, and mistakes the capture for the practice. The practice of accountability, like the practice of any genuine craft, requires an artifact produced under discipline, reviewed by a qualified practitioner, against a standard that can survive adversarial examination. No manual produces the artifact. The substitutes are manuals. The Decision Record is the craft.

The question the vocabulary tries to answer is: how do we make AI systems good? The question admissibility asks is: how do we make decisions reviewable? These are not the same question, and conflating them is the reason the last five years of effort have produced so much documentation and so little accountability.

II. Alignment

Alignment began as a technical research program and has become a civilizational aspiration. The program, in its narrowest form, asks how to train a large model so that its outputs track the intentions of its designers rather than the literal reward signal it was trained on. The aspiration asks how to ensure that increasingly capable AI systems act in accordance with human values.

The program is real work. Reinforcement learning from human feedback, constitutional AI methods, debate protocols, scalable oversight, these have improved the behavior of deployed models on measurable benchmarks. Models in 2026 refuse categories of requests they would have complied with in 2022. They cite their sources more often. They hedge more calibratedly. These are not nothing.

But alignment, even if it worked perfectly, would not solve admissibility. A perfectly aligned model that denied a benefits claim would still leave the affected party with no reviewable record of why. A perfectly aligned model that recommended a sentence would still leave the judge with no way to show that the recommendation was arrived at rather than generated. A perfectly aligned model that approved a loan would still leave the regulator with no substrate to audit against.

Alignment is about the internal quality of a model's reasoning relative to its designers' intent. Admissibility is about the external quality of a decision's record relative to the rights of people affected by it. These are orthogonal concerns. A model can be aligned and its decisions inadmissible. A model can be misaligned and its decisions, if properly recorded, reproducible, and reviewable, admissible enough to be challenged, which is what matters when you are the person being denied.

…

The chapter goes on from here — why each of the six substitutes — transparency, explainability, audit, and the rest — does partial work and leaves the real question untouched. It continues in the full manuscript.

Unlock the full book

From Chapter Five

What Is a Real Decision?

An excerpt from the opening of Chapter Five. The chapter — and the rest of the book — continues behind the wall.

§1.

The Loomis sentencing hearing took place in a small wood-paneled courtroom on the second floor of the La Crosse County Courthouse, in the early afternoon of August 12, 2013. The judge, Judge Scott Horne, had been on the La Crosse County Circuit Court bench for nine years. The defendant, Eric Loomis, was thirty-one years old. The state's prosecutor was a senior assistant district attorney named Rian Radtke. The defense was represented by a public defender named Michael Rosenberg. The pre-sentence report on Judge Horne's bench, beneath the open Wisconsin Statutes, included a single page that summarized Loomis's COMPAS risk score. The number was a "high", both for violence and for general recidivism.

Judge Horne, in his subsequent sentencing remarks, acknowledged the score. He did not, he said, rely on it exclusively. He explained that he had considered the pre-sentence report as a whole, the nature of Loomis's offense, the defendant's prior record, the circumstances of his life, and the broader interests of public safety and rehabilitation that Wisconsin's sentencing framework asked him to weigh. He imposed a sentence of six years in prison.

What the judge did not have, on the bench that afternoon, was access to the algorithm that had produced the score. He had a number. He referenced the number. The number was opaque to him as it was opaque to Loomis, to the public defender, to the prosecutor, to the appellate courts that would later review the sentence, and ultimately to the Supreme Court of the United States, which would deny certiorari in June 2017. The opacity is what this chapter is about. The number on the bench was the product of a process whose specific operations on Loomis's specific case nobody in the courtroom could reconstruct.

In 2013, Eric Loomis was arrested in La Crosse, Wisconsin, for driving a car that had been used in a drive-by shooting. He pleaded guilty to two of the less serious charges. At his sentencing, the judge assigned him six years in prison. In the sentencing memorandum, the judge cited Loomis's score on a risk-assessment tool called COMPAS, Correctional Offender Management Profiling for Alternative Sanctions, which had rated him a high risk of violence and a high risk of recidivism. The judge did not rely on COMPAS alone. But the score was in the file, and the judge referenced it, and the six years were imposed.

Loomis appealed. He did not challenge the plea. He challenged the use of the score.

COMPAS was a proprietary product. Its algorithm was not disclosed to the court, to defense counsel, or to the defendant. The company that produced it, Northpointe (later Equivant), considered the methodology a trade secret. The questionnaire that produced Loomis's inputs was partially visible; the weights the algorithm applied to those inputs were not. The decisioning logic that had generated his specific score of high risk of violence was, in the judge's sentencing record, an opaque fact. The judge had the number. The judge did not have what the number was made of.

Loomis's argument was simple. He had been sentenced using a decisioning process he could not examine. He could not cross-examine the algorithm. He could not challenge the weights. He could not show that the score was inconsistent with his particular circumstances, because he did not know how the score had been calculated. The sentence had, in effect, been imposed partly by a tool whose reasoning no one in the courtroom could reconstruct.

The Wisconsin Supreme Court ruled against him in July 2016. The opinion, State v. Loomis, held that the use of the score had not violated his due-process rights because the judge had not relied on it exclusively. The court acknowledged that the algorithm was opaque. It acknowledged that the score's reasoning could not be examined. It imposed a prospective requirement, that future sentencing memoranda include a specific cautionary instruction about the score's limitations, and affirmed the sentence.

The United States Supreme Court denied certiorari in June 2017.

Loomis served his sentence. The tool remains in use, in various versions, across American courts.

…

The chapter goes on from here — what the Loomis court could not do, and what a decision worth the name would have required. It continues in the full manuscript.

Unlock the full book

From Chapter Seven

The Decision Record

An excerpt from the opening of Chapter Seven. The chapter — and the rest of the book — continues behind the wall.

In 1958, the U.S. Civil Aeronautics Administration mandated flight data recorders on commercial aircraft after a decade in which high-profile accidents, the 1956 Grand Canyon mid-air collision had killed 128 passengers, and its causes had been partially unreconstructible, had made clear that the absence of a contemporaneous record of cockpit activity was limiting the industry's ability to learn from failures. The mandate was resisted. Pilot unions argued that the devices were surveillance technology whose presence would undermine cockpit collaboration. Airlines argued that installation costs were disproportionate. Manufacturers argued that the weight penalty cut into already-tight operating margins. The mandate was implemented anyway. The subsequent sixty-eight years have vindicated the decision: every major aviation accident since 1958 has been investigated, in substantial part, through the FDR and cockpit voice recorder substrate. The Air France 447 disaster was reconstructed from recorders recovered from the ocean floor two years after the crash. The Boeing 737 MAX investigations traced specific MCAS behavior across fatal crashes because the recorders preserved what the cockpit crews had seen and done. The commercial-aviation fatal-accident rate has dropped by roughly two orders of magnitude since the mandate; many factors contributed to the drop, but the FDR is the specific condition of possibility for the improvement, without the substrate for identifying what went wrong, the work of preventing its recurrence could not begin.

The Decision Record, in structure and purpose, is the automated-decisioning domain's flight data recorder. The parallels are precise: a high-stakes outcome produced under time pressure across distributed actors, whose reconstruction at the moment of investigation is what makes accountability possible. The initial resistance, from operators who find the preservation requirement costly, from vendors who find the transparency uncomfortable, from regulators who have not yet grasped how radically the substrate changes their enforcement capacity, tracks the resistance to FDR adoption in the 1950s. The eventual universality is what the aviation pattern predicts for admissibility too.

The earliest specification of such an artifact that still shapes contemporary practice was Luca Pacioli's 1494 treatise, which did not invent double-entry bookkeeping so much as crystallize the Venetian practice of the previous two centuries into a form that could be taught, reproduced, and audited. Pacioli's insight was that accountability required an artifact whose integrity properties were internal to its construction: if the books did not balance, the record itself told you so, and the error had to be found before any downstream claim could be trusted. The Decision Record stands in the same lineage. Its seven required fields are not a list of documentation recommendations. They are the internal integrity conditions of an artifact designed to reveal its own inadequacy before any downstream claim is made on it.

The history of double-entry bookkeeping is worth tracing briefly, because the parallel to admissibility is more precise than a passing invocation can show. The technique emerged in Italian commercial cities during the thirteenth and fourteenth centuries, Genoa, Venice, Florence, in response to specific operational problems that the merchants of those cities could not solve with older single-entry ledgers. A single-entry ledger records transactions as they occur; a double-entry ledger records each transaction as an equal and offsetting pair of entries in two separate accounts, such that the sum of all debits equals the sum of all credits at every point in time. The double-entry discipline does not prevent errors; it produces a property, the trial balance, that makes errors visible. If the trial balance does not balance, something is wrong, and the error must be found before any financial statement drawn from the ledgers can be trusted.

Pacioli's 1494 Summa de Arithmetica, Geometria, Proportioni et Proportionalità dedicated a section, Particularis de Computis et Scripturis, to double-entry. The treatise was not original research; it described practices Italian merchants had developed over the preceding two centuries. Its contribution was codification: Pacioli wrote the practices down in a form that could be taught in the emerging commercial schools, transmitted across languages and borders, and reproduced by practitioners who had not learned at the feet of a Venetian merchant. The printing revolution, then less than fifty years old, gave Pacioli's treatise reach the prior manuscript tradition could not have provided. By 1550, double-entry had spread to the merchant classes of Antwerp, Nuremberg, and London. By 1600, it was the default technique for commercial recordkeeping across Europe. By 1700, it had become the foundation of the joint-stock-company form whose subsequent evolution shaped modern capitalism.

…

The chapter goes on from here — what a record has to carry to be worth keeping — and what the flight recorder understood that most systems still do not. It continues in the full manuscript.

Unlock the full book

From Chapter Nine

The Legal Threshold

An excerpt from the opening of Chapter Nine. The chapter — and the rest of the book — continues behind the wall.

I. Daubert in a World of Models

In June 1993, the Supreme Court decided Daubert v. Merrell Dow Pharmaceuticals. The case was about Bendectin, an anti-nausea drug prescribed during pregnancy, and whether epidemiological testimony linking the drug to birth defects could reach a jury. The answer was a four-factor test for scientific evidence: testability, peer review, known error rate, and general acceptance. The test replaced the older Frye standard of "general acceptance in the scientific community" with a more demanding inquiry into methodological reliability.

The specific facts of the Daubert case are worth naming. Jason Daubert and Eric Schuller were children born with serious birth defects, limb reductions and other congenital malformations. Their parents, Joyce Daubert and her co-plaintiff, alleged that their mothers' use of Bendectin, an anti-nausea medication marketed by Merrell Dow for morning sickness, had caused the defects. Bendectin had been on the market since 1956 and had, by the time the Daubert family's case was filed in 1989, been the subject of extensive epidemiological research. The published research, reviewed by the FDA, had not found a statistically significant association between Bendectin and birth defects at the scale of the available studies. The plaintiffs' case rested on expert testimony from a group of scientists who, in their own re-analysis of the existing studies and through specific animal-toxicology research, argued that the published literature had underestimated the risk.

The district court excluded the plaintiffs' expert testimony under the Frye standard, the experts' conclusions were not generally accepted in the scientific community, so their testimony could not reach the jury. The Ninth Circuit affirmed. The Supreme Court granted certiorari to resolve the question of whether the Federal Rules of Evidence, passed in 1975, had displaced the older Frye standard. Justice Harry Blackmun wrote for the Court. The opinion held that the Federal Rules had displaced Frye, and that the new standard for admitting scientific expert testimony required the trial court to assess the reliability of the methodology rather than the consensus of the field. Blackmun's opinion listed four factors, testability, peer review, known error rate, general acceptance, as non-exhaustive considerations that trial courts could apply in the reliability inquiry. The case was remanded to the lower courts, which ultimately held that the plaintiffs' testimony failed the reliability inquiry even under the new standard. The Dauberts lost their case on the merits. The doctrine their case established became controlling for every subsequent scientific-evidence question in federal court.

The Frye standard the Court replaced had been established in 1923 in a case involving an early lie-detector technology. The United States Court of Appeals for the District of Columbia held that expert testimony based on novel scientific techniques was admissible only if the technique had gained "general acceptance" in the relevant scientific community. The standard was conservative by design, it protected the jury from novel scientific claims that had not been vetted through the peer-review process. It was also, by the 1980s, increasingly criticized as allowing admission of forensic methods that had gained community acceptance without having gained methodological rigor. The specific failures of forensic toolmark analysis, bitemark identification, and comparative bullet-lead analysis, each of which was admissible under Frye because each had gained community acceptance, and each of which was subsequently discredited through methodologically rigorous examination, illustrated the standard's weakness. Daubert shifted the inquiry from community acceptance to methodological reliability. The shift was partial, Daubert explicitly retained general acceptance as one of its four factors, but the direction was clear.

The Daubert doctrine has evolved through two subsequent Supreme Court cases. General Electric Co. v. Joiner (1997) clarified that trial courts' gatekeeping decisions under Daubert are reviewed for abuse of discretion rather than de novo. Kumho Tire Co. v. Carmichael (1999) extended the Daubert reliability inquiry to non-scientific expert testimony, engineering, clinical, and technical testimony, holding that the gatekeeping obligation applied to all expert evidence, not merely to scientific evidence in the narrow sense. The three cases, Daubert, Joiner, Kumho, together establish what courts now call the Daubert framework. Federal Rule of Evidence 702, amended in 2000 and substantially revised in 2023, codifies the framework.

…

The chapter goes on from here — how the law already decides what evidence is allowed to decide — and what that standard demands of a model. It continues in the full manuscript.

Unlock the full book

From Chapter Eleven

The Procurement Revolution

An excerpt from the opening of Chapter Eleven. The chapter — and the rest of the book — continues behind the wall.

I. How Standards Actually Propagate

Most accountability standards do not propagate through regulation. They propagate through procurement.

HIPAA did not change healthcare IT because the statute was powerful. It changed healthcare IT because every hospital system's vendor contracts began to require HIPAA-compliant data handling as a condition of sale, and vendors who could not demonstrate compliance lost their markets within three years of the rule taking effect. PCI-DSS was not legislated; it was a consortium standard that the major card brands imposed on merchants, who imposed it on acquirers, who imposed it on payment processors, who imposed it on the entire ecosystem within five years. SOC 2 was not a government mandate; it was a contract requirement that large enterprise buyers began inserting into their vendor agreements in 2013–2016, and the SaaS industry reorganized its operations around it in less than a decade. FIPS 140-2 certification became the default for federal cryptographic modules not because a regulator required it universally but because GSA schedules, FedRAMP authorizations, and DoD acquisition packages began requiring it as a line item that excluded non-compliant vendors from bidding.

Alfred Chandler, the historian of American managerial capitalism, identified the procurement function as one of the quiet engines of industrial standardization. When large-scale vertically integrated firms, railroads, steel, oil, automobiles, became the dominant organizational form in the late 19th and early 20th centuries, their purchasing departments became the mechanism by which engineering specifications propagated across supplier ecosystems. A locomotive manufacturer did not invent its wheel standards alone; it specified them in contracts with its steel suppliers, which forced the suppliers' upstream metallurgy practices to converge. This procurement-led standardization is how the interchangeable-parts tradition of Eli Whitney scaled from armaments into industrial production broadly, and how it propagated across national and eventually international supply chains. Admissibility is about to repeat the pattern. The specifications are engineering; the propagation mechanism is commercial. Chandler's observation is still operative.

The HIPAA case deserves development, because it is the clearest American example of a standard propagating through procurement rather than regulation alone. The Health Insurance Portability and Accountability Act was enacted in 1996. Its privacy and security rules were, however, not immediately operational; they required administrative rulemaking through the Department of Health and Human Services. The Privacy Rule was finalized in 2000, with compliance required by 2003. The Security Rule was finalized in 2003, with compliance required by 2005. Enforcement of both rules, through HHS's Office for Civil Rights, began modestly, the first significant civil monetary penalty under HIPAA was not assessed until 2008, twelve years after the Act was passed.

What moved the healthcare-IT industry into compliance was not, primarily, the enforcement risk. It was the procurement requirement that cascaded through the hospital and health-system purchasing departments. By 2004, every major hospital system's IT procurement function required HIPAA compliance as a condition of any new electronic-health-record contract. By 2006, the requirement had extended to adjacent systems, laboratory-information systems, billing platforms, patient-portal products, pharmacy-management tools. By 2008, the healthcare-IT industry's product roadmaps were substantially structured around HIPAA compliance as a baseline expectation. Vendors that could not certify compliance lost access to the market. Vendors that could certified built their marketing materials and sales presentations around the certification. The standard propagated from the largest buyers through the ecosystem within a five-year window, driven by procurement rather than by the direct operation of the statute.

PCI-DSS followed a parallel pattern with different actors. The Payment Card Industry Data Security Standard was launched in 2004 by the major card brands, Visa, Mastercard, American Express, Discover, JCB, as a consortium standard. The card brands imposed the standard on their acquiring banks; the acquiring banks imposed it on their merchant customers; the merchants imposed it on their payment processors and technology vendors. The cascade reached its saturation point within about six years. By 2010, PCI-DSS compliance was a baseline requirement for any vendor handling credit-card data in any capacity. The standard was never legislated; the card brands' commercial authority, amplified through the procurement cascade, produced more comprehensive compliance than most legislative frameworks achieve.

…

The chapter goes on from here — how a standard actually spreads — not through law, but through what the buyers require. It continues in the full manuscript.

Unlock the full book

Lineage

The manuscript reads against a specific shelf, not against the contemporary technology-policy literature.

Lawrence Lessig's work on architecture as a regulatory modality is the closest single referent — the manuscript extends the argument that code precedes law into the procedural domain one layer further in: the architecture of review that determines whether the decisions produced inside an institution can be reconstructed at all. Laurence Tribe's work on constitutional reviewability, and the durability of procedural protection when state action becomes opaque, is the second. The constitutional pressure point the manuscript names — that automated state and institutional decisions become unstable when they cannot be examined adversarially — sits inside the tradition his work has carried.

The further references are partly historical, partly methodological: Hannah Arendt on the conditions under which public judgment is possible; Jürgen Habermas on procedural legitimacy as the load-bearing element of legitimate authority; Niklas Luhmann on second-order observation, and the discipline of asking how a system makes its own decisions visible to itself; Bernard Stiegler on what institutions inherit, store, and lose; Karl Polanyi on the social conditions under which institutional forms are forced to change; Carroll Quigley on the structural transitions between institutional periods. The lineage is named because the argument belongs inside that conversation.

End of excerpt

This is an excerpt from the editorial galley of Admissible Reality. The full book is forthcoming. If you would like to be notified when it is available, you can join the list from the book home.

Join the list Back to the book