Peter Kahl

Around the end of the first millennium China was, on many measures, the most technically accomplished society in the world; sustained, mechanised growth began some six centuries later in Britain. This article does not try to explain that reversal — the Needham Question — but uses it to isolate a mechanism the economic-history literature sets aside. Evaluative institutions do not merely recognise a distribution of capability that exists independently of them; where recognition is strongly coupled across otherwise distinct reward domains, they help to produce it, by drawing developmental effort in advance toward whatever a criterion can already see and reward highly. Capability that departs from the dominant criterion is then under-supplied; and, as the criterion reshapes the population against which its own validity is later judged, it comes to manufacture the evidence of its own correctness. This endogenous self-sealing — not the gaming that Goodhart’s law describes, but a criterion’s growing control over the evidence that might defeat it — is the article’s central proposition. The risk arises as cross-domain coupling and dispositive authority redirect developmental investment; where independent routes of correction are weak, the resulting capability distribution increasingly validates the criterion that produced it, eroding its answerability to contrary evidence. China’s imperial examination is the historical case and Europe an existence proof that misrecognition can be made non-final; contemporary artificial-intelligence benchmarks and safety evaluations are examined for evidence of the same structure, the danger lying less in the exclusion of rival evaluation than in the concentration of consequential recognition weight in a few gates. Innovation, on this account, requires not merely good standards but institutions under which standards can lose.

指月 (Pointing at the Moon) is a cycle of six original fables in the classical Chinese tradition of the 寓言 (yùyán), written in traditional characters and presented in Chinese only, each paired with an ink-wash illustration and closing on a short aphorism (故曰). The six turn on a single figure: an ordinary instrument, made to point beyond itself, mistaken for the thing it was meant to serve. A ruler that measures conformance and cannot see new capability (尺); a steelyard weight — the same word, 權, as "authority" — that becomes dead stone the moment it is owned rather than held in trust; a map taken for the mountain (圖); a coin into which all things must convert, until the priceless is treated as worthless (兌); a seal that goes hollow when no one stands behind it to answer for what it certifies (印); and a fog (霧) — the remainder no instrument can capture — whose clearing leaves no one unmeasured, unpriced, or unaccounted. Together the six form one argument about the mediating layer between a person and the world — measure, authority, representation, exchange, recognition — and about what is lost when that layer is mistaken for reality, or extended until nothing is left outside it. The surface of each fable is complete; nothing is decoded, and the reader supplies the application. The concerns are contemporary — standards and measurement, institutional accountability, commensuration, legibility and surveillance, and the governance of technology and of artificial intelligence — cast in an old form.

We ordinarily treat a change of mind we have thought through as ours — as something we did and can answer for. This paper argues that whether a revision is ours to answer for is settled not by how completely the operative framing is integrated into our present reasoning but by the authority-history of that framing: how the construal which fixed what counted as a reason came to hold its authority for the agent. A framing may be fully integrated — endorsed, reasons-responsive, reflectively stable, available for revision — and this whole present signature can be installed by manipulation that bypasses the agent, a case intrinsically identical to one in which she appropriated the frame herself. The two should differ in authorship because authorship, in the sense relevant to attributing a reconstruction as the agent’s own achievement, must preserve the distinction between actual appropriation and the installed signature of appropriation — a distinction present integration cannot represent, since the two agents are integration-duplicates. This is a functional claim, not a demonstration that the concept must be historical on pain of contradiction: a synchronic concept of authorship is coherent but inadequate to the work the concept is meant to do. For that purpose revisional authorship does not supervene on the agent’s intrinsic present psychological organisation; it tracks the history by which the framing became answerably hers. The paper takes the responsibility literature’s notion of answerability upstream — from judgments and attitudes to the authority of the framing under which they are formed — and sorts a framing’s standing history into appropriated, available but unappropriated, and foreclosed. A consequence it embraces is that much of what we reason under, seamlessly and rationally, was never appropriated, and so is genuinely ours to reason under without being ours to answer for: absence of a defeater does not by itself confer authorship. Its paradigmatic defeaters are motivated capture — recast as stake-driven impairment of a frame’s accessibility to assessment — and manipulative bypass. The account is substrate-neutral, and in closing it constrains the debate over ‘agentic AI’: neither fluency within a frame nor runtime self-revision establishes a system’s authorship of it. The history that matters is not the history of the change of mind but the history of the frame’s authority.

Modern institutions recognise, and recognition earned in one institution can be spent in another, so the fields form a single order rather than a set of separate competitions. This paper treats the machinery that couples them as an evaluative layer running two gates: an admissibility gate that settles what may count as a recognisable contribution, and a conversion gate that settles which recognised winnings travel to other games, and at what rate. The consequential power sits at these gates rather than in any single competition, and at present neither gate is answerable in any general way: no affected party can readily contest what a gate admits or converts, and no institution is charged with reopening a gate rule. The descriptive apparatus is drawn from existing work on positional scarcity, credential inflation, the conversion of capital, commensuration and reactivity, credited throughout, and the paper adds three things to it. It names the pathology of the admissibility gate ‘evaluative crowding’, the congestion of the category-space that turns capability away for want of an admissible form, and marks it off from positional crowding and from commensuration. It asks a constitutional question of both gates at once — by what authority admission and conversion are exercised, to whom they answer, who has standing to contest them, and which institution may reopen them. And it explains why the gates cannot be reopened from inside, through a second-order ‘evaluative capture’ in which incumbents govern not only the standard but what may count as evidence that the standard has failed. A method of detection follows, by which a governed subject may infer the structure from its symptoms while holding that inference to the same revisability it demands of the standard. The paper closes on five requirements — provenance, attribution, standing, revision, remedy — that state what an answerable evaluative order would need. The legal form of the answer is developed in a companion paper.

Three questions are routinely collapsed into one wherever institutional or automated decisions are governed: where a standard was constituted, who is answerable for it, and who is entitled to require that they answer. This paper holds them apart, and finds that they have different answers, terminate at different agents, and want different remedies. A claim recurring across this series — that responsibility sits where standards were constituted rather than where decisions are executed — gives the trace a direction and no destination. Since every structure has an antecedent, a principle saying only 'go upstream' either never halts or halts arbitrarily. The stopping condition offered here is historical rather than normative: of each agent one asks whether the standard, in the operative form it bore at the time of the conduct, pre-existed that agent's contribution. Where it did, the agent transmitted; where it did not, the trace stops. Standards have constitutive dimensions — content, authority, operation, jurisdiction — settled by different agents at different times, so a trace returns a profile rather than a constitutor. Someone settled what a benchmark measures; no one settled that the benchmark decides. Three kinds of vacancy are distinguished, each with its own remedy. An earlier claim of the author's is corrected, and the rule applied to an automated credit decision and to the Financial Services and Markets Bill [HL].

It is widely held that a directed-evaluation state possesses an advantage in the governance of innovation. This paper argues that the claim conflates the production of capability with its recognition, and that once the two are separated the governance question changes shape. Recognition criteria are altered constantly, and everywhere: by insolvency, by expiry, by the departure of those who wrote them, and by the replacement of those in whose practice they consisted. Each of these compels an alteration; none produces a reason. Sealing is therefore the closure of the reason-channel and not of the change-channel, and an evaluative layer may be altered daily and remain perfectly blind. Answerability is demoted accordingly. It is not the engine of correction but the condition of learning — the production of a statement, by a party competent to be heard, of the respect in which an extension failed to count them. Where the affected set is sufficiently closed, institutions can approximate the conjunction of compulsion and avowability; merger control supplies the strongest case examined here. Where the affected set is open they can be conjoined only through representation of the absent, and representation is neither guaranteed nor neutral: no procedure ensures that the interest a criterion excludes will be represented before the criterion forecloses its formation. The party injured by an innovation-recognition criterion’s failure to count a capability is characteristically a firm that does not yet exist. Innovation governance is on this account not the site of an impossibility but of the severest form of a representational asymmetry — the case in which the excluded interest is the one the exclusion prevents from forming. Both jurisdictions examined fix their innovation-recognition extensions by instruments that are sealed on the tests developed here: in each, criteria change without a stateable reason. Whether they differ in the rate at which drifted practice is consolidated intelligibly into declared criteria cannot yet be stated, that rate being defined only where criteria are written down and the two jurisdictions writing down different things. The visible difference lies in production rather than in evaluation.

A state that would recognise genuinely new capability must sometimes reopen the very standards by which it recognises. Its companion paper established that the coexistence of evaluative-layer generativity and reconstructable answerability is possible, conditional on the form answerability takes, and perishable even where the form is right, and it specified without building a device to hold the combination. This paper builds the device and defends it. It argues that the lawful reopening of a recognition standard is a distinct constitutional act with a grammar: a finite, rule-governed structure — declared mandate, bounded field, named responsibility, contestation of the standard, and revision duty — that makes a reopening well-formed rather than ill-formed. The grammar is shown to be an arrangement of materials English public law largely already possesses, grounded in vires and the principle of legality, the duty to give reasons, the anti-ouster principle, and sunset and review design; and it is tested against four live instruments — the Financial Conduct Authority regulatory sandbox and its legislated European counterpart, the Advanced Research and Invention Agency, and the Vaccine Taskforce — none of which holds all five features. The paper then anatomises capture, the reproduction of the grammar’s form without its licensing structure, in six modes each answered by a structural safeguard. Finally it supplies the rule the features cannot: a downside-allocation rule that fixes, in advance, which instrument bears a failed bet — a normative construction transposed from the accounting-officer and ministerial-direction practice in which the form already exists. The rule dissolves the defensive-rationality equilibrium the companion paper diagnosed, and converts the grammar from a licence into an institution. The paper closes on the kind of legal order that can hold answerable generativity: one whose forum cannot be closed.

A small number of media and platform institutions decide what a polity is able to consider, which claims are admissible, and which accounts carry the standing of a serious position. This paper asks on what terms that power can be held to account. It begins with media power in general but locates the strongest case in the dominant intermediary — the search, social or recommendation service that curates the visibility of others' expression at infrastructural scale — with the press and broadcasters entering mainly as contrast and limiting case. Its claim is deliberately modest. I do not argue that these institutions are organs of the state, nor that curating a feed is legislating or adjudicating; the influence such institutions exercise is not authority, and any argument that rests on equating the two collapses at the first objection. The narrower and more durable claim is that a handful of institutions concentrate three separable powers — selection, boundary-setting, and legitimation — over the conditions of public knowledge, and answer for none of them against any standard they must state or revise. I locate the constitutional interest in that concentration of evaluative discretion, and I distinguish it sharply from the ownership concentration that media-plurality law already addresses: the two can and do diverge. The remedy I defend is correspondingly narrow. It pairs a procedural duty of answerability — state the standard, explain the exercise, submit the standard to revision — with a thin substantive floor drawn not from any contested notion of fairness but from the narrowest part of the fiduciary duty of loyalty — the duty to disclose a material conflict of interest bearing on the exercise of that discretion. I own the cost of that choice: the floor reaches the interest served in the dark, not curation that is merely disclosed and self-interested, and not curation that is merely, in someone's view, unfair. It is calibrated to survive the free-expression constraints that cap it, including — decisively, and against a literature still arguing from Tornillo alone — the protection that Moody v NetChoice now extends to algorithmic curation, and the boundary that X Corp v Bonta draws between compelling a platform to characterise its editorial reasoning, which fails, and requiring it to disclose the interests that bear on it, which does not. Against the objection that any overseer will itself be captured, I claim only that reflexive answerability makes capture more costly and more visible, not that it makes capture impossible; that is the most a constitutional design can honestly promise.

Innovation failure in a rule-of-law state is commonly explained by deficits of input — investment, skill, infrastructure — or of throughput, the efficiency with which invention reaches adoption. This paper relocates the problem to an upstream evaluative function: the state’s capacity to recognise new capability before that capability has been made legible to inherited standards of credible expertise, viable enterprise, acceptable risk and public value. It defines this capacity as evaluative-layer generativity and gives it an ex-ante specification — observable in the institutional permission structure before any outcome — in three markers: an authorised departure from prevailing evaluative standards; a consequent lowering of the cost of illegibility, such that actors may act and be recognised before validation; and the traceability of the departure. The specification severs generativity from the success it sometimes produces, and so avoids defining it by hindsight. Three cases fix the relationship between generativity and answerability. Shenzhen shows generativity secured by an authorised departure that was traceable to the party-state but never reconstructably answerable to a public — generativity without the answerability a rule-of-law order requires. Britain shows the inverse: an abundance of proceduralised answerability that forecloses the reconstructable kind, an equilibrium in which officials rationally optimise for demonstrable defensibility and so recognise capability only after it has become legible. Hong Kong shows that the two properties can coexist — generativity under law — though only in certain economic recognition channels, and fragile once its answerability constraints erode. The resulting thesis is conditional: generativity trades off against answerability in its proceduralised form, not against answerability in its reconstructable form, and even the favourable combination is perishable, sustained only while its constraints are actively held. The paper’s principal contribution is to distinguish generative state capacity, located at the evaluative layer, from the administrative state that processes and the developmental state that builds. This paper is diagnostic: it locates the problem and stops at it. The companion paper, Answerable Generativity in Innovation Governance: The Constitutional Grammar of Lawful Standard-Reopening, is constructive, developing the institutional devices through which a rule-of-law state might reopen evaluative standards while keeping the reopening answerable.

The debate on artificial intelligence and the labour market has settled, within a year, on a set of distributive remedies: a basic income, a levy on AI use, and a sovereign fund holding public equity in the AI companies. These instruments recapture the value that automated systems produce and share it more widely. None of them governs the decisions those systems make. This paper argues that the gap they leave open is a gap in the rule of law, and that it has been mislocated. The objection is not that automated systems exercise discretion, which is ordinary and legitimate, nor that their discretion should be abolished, which is neither possible nor desirable. The objection is that automation removes the occasion on which discretionary power has always been checked: the point at which a consequential decision can be reconstructed and challenged before a forum. The distributive remedies do not restore that occasion, and the most ambitious of them, state equity in AI, works against it, because a state that holds a financial stake in the firms it must check acquires an interest in their expansion that its duty to check them then has to overcome. The paper sets the argument within the conception of the rule of law that treats discretion as a thing to be confined and reviewed, and locates the missing occasion within the procedural account of legality.

Debate on digital sovereignty has settled on a single axis: who owns and runs the systems on which critical services depend, and whether they can be kept running under pressure. The defeat of New Clause 13 to the Cyber Security and Resilience (Network and Information Systems) Bill on 16 June 2026, and with it a statutory Digital Sovereignty Strategy, was a defeat on that axis alone. This paper draws a second axis, ‘evaluative sovereignty’, the control of the evaluation a system performs rather than the asset it runs on, and argues that a sovereignty strategy confined to the first leaves the more consequential exposure untouched. That exposure is constitutional. The rule of law requires that a consequential exercise of public power over a person remain capable of reconstruction and challenge, and a sealed evaluative layer can withdraw a binding decision from effective review in practical effect, leaving the court's jurisdiction formally intact and the statute book untouched. The deficit is not hypothetical: it has been realised in the Dutch SyRI and childcare-benefit cases and in the United Kingdom's own automated welfare-risk systems. None of the instruments now in view reaches it, neither pre-deployment audit nor the asset-focused work of procurement and a strategy duty, and the sovereign-funding posture can work against it. What evaluative sovereignty requires is an ‘answerability occasion’: a defined point at which a consequential evaluation can be reconstructed and challenged before a forum the deploying authority can answer to and the affected person can invoke. The Bill’s Lords stages are the live opportunity to provide for it.

Debates on the governance of artificial intelligence assume that the constitutional difficulty is one of regulatory content — which models, which harms, which agency. This paper argues that frontier AI also alters the temporal conditions under which regulation can remain answerable. Anchored to the framework of evaluative sovereignty, it distinguishes procedural latency from deliberation and contends that parliamentary scrutiny rests on an inherited assumption — that material conditions hold still between legislative stages, sittings, and recesses — which compression falsifies. The effect is not merely slower legislation: compression removes the occasion on which discretionary power is checked, and transfers governance toward the answerability-poor fast lane of delegated and prerogative instruments. The claim is sharpened by a structural feature of United Kingdom practice — that during recess the power to recall Parliament depends on executive request — and by a reading of R (Miller) v The Prime Minister [2019] UKSC 41 for the effect-based principle it states rather than the deliberate prorogation it condemned. The paper rejects distributive remedies as a substitute for governance, sets out a continuity architecture that includes a reformed, answerability-generating emergency-powers regime, and cautions that accelerating Parliament’s seeing must never delegate its judging.

Upward mobility under meritocratic conditions is analysed as a recognition order in which a single cardinal metric of credentials and ranks rations standing and access. The paper’s central mechanism is ‘signal decay’, the fall in the value of an attained position as the strategy for reaching it crowds, a result that holds even where credentials build real capacity and that is distinct from both positional congestion and rising inequality. A stylised assurance game shows the conforming order to be a stable equilibrium collectively dominated by a plural alternative the population cannot reach unilaterally, with its degradedness rising in the tightness of the gate. The order cannot correct itself, because its would-be correctors are the winners the metric selected, so the corrective never reaches the evaluation layer and the reforms it can pass deepen signal decay on the positional margin. When ascent fails, the outcome is reported as the verdict of the metric rather than owned as a judgement. The paper offers a discriminating prediction that separates its mechanism from inequality. It reads the order against the European Union’s prohibition on social scoring, and proposes an external corrective that restores answerable judgement while keeping the metric as a check.

The evaluation layer of financial markets, the benchmarking and manager-selection machinery that allocates capital on a commensurated scale, runs a recognition game, an order in which a single scale makes standing scarce and positional and the contest for standing deforms allocation. Its distinctive consequence in markets is that the layer’s own corrective is captured. The arbitrage that should discipline the layer’s misallocations is performed by agents the layer itself scores, so correction is a losing move under the same recognition order it would overturn, and the layer’s errors accumulate into fragility rather than being competed away. The paper develops this through three claims, each anchored outside the author’s prior work. The captured corrective extends the limits-to-arbitrage account by locating the bound on arbitrage inside the recognition order rather than outside it. The recognition game supplies the ex ante formation of the crowd whose collapse the crowded-strategy and endogenous-risk literatures already describe, a claim operationalised through active share and offered as a falsifiable prediction that fragility concentrates where commensuration is tightest at a given level of leverage. Algorithmic allocation carries the displacement of answerable judgement to its limit, where no judgement is formed to be reconstructed and the model-risk apparatus governs a process without recovering a reason. The remedy returns the metric to the standing of evidence and makes answerable judgement safe to exercise across the field, which is shown to be why the reform attempted under Dodd-Frank section 939A under-delivered. The paper is a framework with one prediction and a model sketch rather than an empirical test.

Metric governance does more than measure research; it governs recognition, and in doing so it reorders the relations through which knowledge is produced. This paper argues that the commensuration of academic standing onto a single scale makes recognition scarce, and that the scarcity produces two pathologies at once. Horizontally it adversarialises researchers, turning colleagues into rivals for a positional good. Vertically it produces ‘epistemic clientelism’, the moderation of independent judgement to remain legible to the patrons who control advancement. These compose one ‘recognition game’ rather than two separate complaints. Through ‘evaluative-layer capping’, the game rewards work that extends inherited standards and misrecognises the generative work that would reopen them, while the use of a metric as a verdict launders the answerability for the misrecognition. A simple assurance game shows why the degraded order is stable although few would choose it and no one need act in bad faith, with the better culture held out of reach by the risk of pursuing it alone. Because metrics were themselves a reform against unaccountable discretion, the corrective cannot be a return to discretion. The prescription that judgement should govern indicators already belongs to the responsible-metrics reforms; what they lack, and what the paper supplies as ‘epistemic governance’, is the coordinating device that makes the better equilibrium safe to reach, an arrangement that restores answerable, owned judgement while keeping the metric as a check rather than a verdict.

This article theorises a structural condition in contemporary governance: delegated discretion, defined as the persistence of decision-making authority in the absence of evaluative control. It arises where authority is formally distributed across an institutional chain while the evaluative framework governing decisions is not operable by any actor within that chain. Existing literatures in fiduciary law, administrative law, and principal–agent theory presuppose that such control is located somewhere within the authority structure. The administrative-law tradition on discretion makes the same presupposition: the programme of confining, structuring and checking discretion, in place of its elimination, assumes throughout an identifiable actor who holds the discretion and can be made answerable for the standards it applies. The article shows that this presupposition can fail without doctrinal recognition, such that governance arrangements may satisfy formal oversight requirements while lacking any actor capable of assessing the standards they apply. Three consequences follow. First, responsibility misalignment: fiduciary doctrine attributes responsibility on the assumption that authority entails evaluative control. Second, recognition indeterminacy: actors cannot determine whether such control exists within the chain. Third, doctrinal distortion: oversight and non-abdication operate on assumptions that are not satisfied under this condition. These claims are developed through fiduciary doctrine in Delaware and England, supported by a case study in AI-mediated corporate governance and illustrations from algorithmic administration and central banking. The article concludes that governance is mis-specified as a problem of discretion rather than evaluative control: even the discretion tradition's most developed response presupposes the very locus whose absence defines the condition.

Frontier AI governance is calibrated to catastrophe — to severe, capability-driven risks of the kind that threaten national security or the loss of human control. No framework of that kind has a trigger that fires when a lawful and non-catastrophic deployment leaves an institution unable to reconstruct and defend a decision it has made. This paper names that harm ‘answerability capture’, marks it off from explainability and from accountability, and argues that it forms across composed systems whose parts are each individually compliant. Its deeper source is a loss of evaluative control: where the criteria governing a decision fall outside the reach of every actor in the chain, even a complete record can yield no account, and the window for revising the decision closes faster than retrospective review can engage. Existing decision-level doctrines cannot reach it, and the trust now offered in its place is reputational confidence rather than a capacity that can be tested. The paper proposes an ‘answerability fuse’: a statutory escalation duty, held by a body outside the institution governed, that fires on three defined occasions and can compel disclosure and the preservation of evidence, with a rebuttable temporary restraint where those do not suffice. Current standards practice shows these controls are already feasible, so what the proposal adds is the condition on which they fire and the external hand that holds them.

This working paper is a feedback submission to Singapore’s Infocomm Media Development Authority on the Model AI Governance Framework for Agentic AI (version 1.5). It argues that the framework is operationally strong, but that its aspiration to secure ‘meaningful accountability’ requires three further refinements. First, the paper distinguishes allocation from answerability: assigning a named human, team, or role does not by itself ensure that anyone can reconstruct, justify, and bear responsibility for a specific agentic-AI action. Secondly, it argues that composition is a distinct source of risk in multi-agent systems, because a duty may be satisfied at every individual node yet fail at the level of the composed system. Thirdly, it argues that binding constraints on high-stakes or irreversible agent actions must sit at an objective layer outside the agent’s generative reach, rather than merely inside prompts, behavioural instructions, or model-context safeguards. The paper also clarifies the distinction between autonomy and agency, recommending that responsibility-language be reserved for humans and institutions, while capability-language be used for artificial agents. The submission therefore develops a governance account in which traceability, allocation, oversight, and identity management are necessary but insufficient unless linked to institutional answerability and structurally enforceable constraints.

Composability is not only a technical design choice; it is a relocation of market reliance. Reconstructing financial instruments from a common, reusable grammar of title and entitlement tokens moves the basis of reliance out of asset-specific silos and into shared, self-executing components. That relocation does not remove risk; it relocates it. Once instruments are assembled from shared components, the reliability of the whole ecosystem depends on the integrity of those components, the authority under which they self-execute, and the contestability of the mappings between token-state and legal or economic entitlement. This paper specifies the governance layer that the leading UK composability proposal expressly defers. It sets out four integrity conditions — data, system, epistemic and mandate integrity — for composable infrastructure; it argues that such infrastructure can ‘fail without crashing’, remaining available and immutable while the meaning relied upon has degraded; it proposes a ‘mandate envelope’ that extends, rather than replaces, the entitlement-token grammar, so that self-execution remains bounded by authority; and it sets out a minimum integrity-and-answerability framework, anchored in existing UK regulatory instruments, specifying the conditions under which participants, regulators, central counterparties, custodians, courts and investors may rely on a tokenised representation. The argument is offered as the governance companion to the composability proposal and as a contribution to the Call for Input issued by the Financial Conduct Authority and the Bank of England on 18 May 2026.

Paradigm-seeking research organisations, frontier artificial intelligence laboratories prominent among them, increasingly concentrate two distinct forms of authority in a single operational leadership: control over research procedure, and practical control over what counts as sound knowledge. This article argues that fusing the two is a governance design error. It treats a research organisation as a distributed epistemic infrastructure and distinguishes routine epistemic maintenance — competent work within inherited standards — from generativity, the capacity to reconstruct governing evaluative standards when they fail. Generativity is exercised through preserved internal pathways of contestation; an organisation that lets those pathways close becomes, whatever its mission, an enforcement mechanism for standards it can no longer revise. The article then separates two authorities. ‘Procedural authority’ over how research is vetted, timed and released may and should be concentrated in identifiable, answerable hands. ‘Epistemic finality’ — standing to place a conclusion beyond reopening because of who holds it — should exist nowhere, though well-grounded judgement properly carries ‘epistemic weight’. Their fusion is termed ‘hollowed authorisation’: the forms of scrutiny retained around conclusions scrutiny may not reach. The article meets the objection that funding and deadline pressure make such reopenability unaffordable, distinguishing pressure that compresses an organisation’s research horizon from pressure that forecloses its contestation pathways. Meta’s 2025 artificial intelligence reorganisation is used as a worked illustration of how the pathology would present, not as proof of its occurrence or frequency; the Encode: AI for Science Fellowship and Bell Labs serve as briefer comparators marking, respectively, structurally conservative ambition and disciplined long-horizon generativity. The article concludes that research leadership is self-limiting: procedural authority, held in trust for the organisation’s generative capacity, is legitimate only insofar as it is not converted into epistemic finality, because that conversion turns the authority against the purpose for which it is held.

This paper develops Representational Sealing Theory (ReST), a theory of the limits encountered by any inquiry that turns to ask after its own preconditions. Its premise is a claim in the philosophy of representation: theoretical representation of a domain is conducted within a structured system of admissible distinction and inference, and that structure constitutes, rather than merely accompanies, what such representation can represent. From this premise the paper derives a ‘representational seal’: a domain dependent upon a prior generative condition cannot theoretically represent the intrinsic nature of that condition, since its only means of representation is a structure the condition does not share. What such a domain may establish is thin and exact: that it has a generative condition, and that the condition was adequate to it. The paper’s principal positive consequence concerns time. A generated domain must represent its own states as non-self-grounding, as referring beyond themselves to a condition it cannot represent. These features are identified as the materials of temporal order: not a single succession, and not the arrow of time, but the dependence-character through which generatedness appears from within. The account is distinguished from relational theories of time, in particular Rovelli’s. Its bounds are stated explicitly: the seal supplies the materials of temporal order, but not their assembly into a single succession, nor the orientation of such a succession. The account contributes nothing to the content of physics; it is a constraint on how a domain may theorise its own preconditions.

California’s AB 2013 requires developers of covered generative AI systems available to Californians to publish a high-level summary of training datasets. It does not require the corpus, model weights, source code, training recipes, document-level manifests, exact source lists, or technical methods. In xAI LLC v Bonta, xAI challenges this obligation on three constitutional grounds: First Amendment compelled speech, Fifth Amendment uncompensated taking of trade secret property, and Due Process vagueness. This paper argues that the brief’s three theories rest on doctrinal extensions the authorities cited do not support, and that the trade secret claim in particular rests on a category error. The paper’s principal original contribution is a three-layer analysis of the trade secret claim. It distinguishes the underlying knowledge commons from which training material is drawn, the categorical enclosure relationship AB 2013 actually requires disclosing, and the technical apparatus in which proprietary value most plausibly resides. AB 2013 reaches only the second layer; trade secret protection is doctrinally strongest in the third. Ruckelshaus v Monsanto, Philip Morris v Reilly, and California compilation trade secret doctrine do not extend protection from the third layer to the second. The First Amendment claim is best analyzed under Zauderer–Central Hudson–Stolfi as factual commercial disclosure, though current Supreme Court doctrine leaves the classification uncertain. The vagueness claim is xAI’s strongest argument, but the proportionate remedy is narrowing construction or safe-harbor guidance, not facial invalidation. The paper’s broader theoretical contribution extends the capture of answerability framework from administrative to constitutional tier. xAI’s three theories, in combination, would foreclose the legal authority from which future AI training-data accountability could be built. AB 2013 demands a webpage. xAI v Bonta asks whether even that documentary floor may be required.

When an institution controls not only the decision affecting a person, but also the criteria, evidence, forum and remedies through which that decision can be challenged, accountability becomes circular. This paper calls that condition the capture of answerability. Building on Schedler, Bovens, Pettit and Mashaw, it argues that accountability mechanisms may be formally present while substantively hollow where the epistemic and procedural preconditions of challenge are themselves institutionally controlled. The paper introduces a two-tier taxonomy: structural captures of decision, criteria, evidence and forum constitute the core condition; amplifying captures of remedy, exit, narrative and liability indicate severity and durability. The self-sealing architecture is most acute where criteria capture and forum capture coincide: the affected person cannot challenge the evaluative standard without a competent forum, while the forum cannot test the challenge without access to the evaluative standard. The taxonomy is developed through a doctrinal synthesis of public law, data protection, contract and regulatory materials, showing both where existing law already supports answerability and where AI-mediated decision-making exposes structural gaps. The framework is developed through AI-mediated decision-making generally and UK university AI-assisted academic-integrity processes specifically. The paper anticipates and addresses ten predictable objections, distinguishing answerability from accountability, transparency, explainability, and existing doctrine. It proposes an Answerability Capture Audit, presented as a doctrinal synthesis and practical governance instrument, alongside four differentiated reform proposals directed at the ICO, the Office for Students, the OIA, and the Law Commission.

This paper advances a social-epistemological account of research metrics as epistemic infrastructure: socio-technical systems whose outputs function as authoritative signals within institutional practices of evaluation. Citation counts, download statistics, h-index scores, and rankings are not merely descriptive indicators but inputs into decisions about promotion, funding, and reputational standing. Their epistemic status therefore depends not only on accuracy in a narrow sense, but on the conditions under which they are produced, validated, and contested. The paper introduces the concept of reliability-for-reliance as a normative standard for such systems. An output meets this standard where it is sufficiently accurate, auditable, and contestable to justify its use in consequential decisions without independent re-verification. Failures of epistemic infrastructure need not involve deception or breakdown; they may arise through ordinary system defects, aggregation practices, or institutional reliance on outputs that lack adequate provenance and contestability. These are cases of silent epistemic degradation, in which signals appear plausible while no longer warranting trust. Drawing on the structural lessons of the Post Office Horizon scandal as a case of institutional over-reliance on system-generated outputs, the paper argues that similar, though less severe, epistemic risks are present in the governance of research metrics. The core problem is not only that systems can be wrong, but that institutions may lack the procedural conditions required to recognise and correct error. This generates asymmetries of knowledge and power that bear comparison to forms of epistemic injustice. The paper situates these concerns at the intersection of social epistemology, philosophy of technology, and information ethics. It argues that epistemic infrastructure requires institutionalised forms of auditability, provenance, and contestability, and that current practices in research evaluation fall short of these requirements. The analysis is developed alongside a policy framework grounded in data protection and governance principles, but its primary contribution is philosophical: to characterise the epistemic conditions under which system-generated signals can legitimately function as knowledge in institutional decision-making.

This article examines Musk v OpenAI (Musk v Altman) as a potential test case for applying charitable-trust doctrine to artificial intelligence governance. The dispute is not merely a founder conflict, commercial rivalry, or disagreement about corporate form. At its strongest, the case asks whether property, rights, governance powers, or other legally recognised interests accumulated under sufficiently definite public-benefit AI commitments may later be governed by standards of safety, access, release, licensing, monetisation, auditability, and public benefit inconsistent with that purpose. OpenAI’s founding matrix — nonprofit orientation, non-private-gain language, safety commitments, public-benefit representations, and openness ‘when applicable’ — may be relevant to construing whether a restricted charitable purpose existed, though the commitments may instead be found aspirational, adaptive, discretionary, or too indefinite. If a restriction is proved, breach requires more than hybridisation, private capital, or closed models; it requires inconsistent control over relevant assets, rights, governance powers, or evaluative standards. Musk’s strongest standing theory is not as founder or competitor, but as an alleged settlor, restricted-purpose donor, or purpose-connected claimant. Even then, enforcement must remain purpose-protective and relief should run to the charitable object, not Musk personally. The article introduces ‘epistemic infrastructure’ as a functional analytic concept — describing AI systems that mediate knowledge, safety, access, release, and institutional reliance — and explains why control over such systems may be legally material where a charitable purpose concerning their governance is proved. It does not create a new cause of action. Musk v OpenAI may become an early test of whether charitable law can enforce restricted public-benefit commitments attached to AI infrastructure. A comparative annex examines how English charity law would approach the same facts.

This article examines multi-agent architectures in institutional asset allocation—exemplified by ‘self-driving portfolio’ systems—through the lens of evaluative structure. While these systems distribute forecasting and allocation across specialised agents, selection remains governed by a fixed Investment Policy Statement (IPS). The central claim is that such architectures expand search over the portfolio space without addressing the specification or revision of the objective function. The analysis formalises a distinction between optimisation and evaluation. In current systems, the evaluative structure remains external, so errors and disagreement are processed relative to a fixed objective rather than as evidence of its mis-specification. Recasting multi-agent systems as ensemble optimisers, the article shows that model diversity does not imply outcome independence under shared evaluative criteria. A stylised simulation isolates the consequences of shared evaluation. Increasing similarity produces greater portfolio overlap, higher concentration, and stronger drawdown correlation under stress, with clustered adaptation dynamics. The article defines the objective problem, distinguishes parametric adaptation from evaluative revision, introduces evaluative crowding, and proposes Objective-Layer AI as a minimal extension in which objectives become operational variables within the system. The frontier lies in designing and governing the objective layer itself.

Financial markets exhibit a recurring pattern in which strategies built on heterogeneous data, models, and signals converge in positioning and unwind in a highly synchronized manner under stress. Existing accounts attribute such comovement to liquidity constraints, leverage, and overlapping holdings, but these mechanisms explain amplification rather than the emergence of alignment ex ante. This article advances a complementary explanation, evaluative crowding, in which convergence arises from shared structures of evaluation rather than shared information. Agents optimize under evaluative structures l in L, comprising risk operators, admissibility constraints, and objective functions. When these are partially aligned across institutions, heterogeneous signals are mapped through a common transformation Phi_l, inducing clustering in admissible strategies without coordination or common beliefs. A formal model shows that when the evaluative operator has low effective dimensionality, optimal portfolios concentrate in a restricted subspace. Alignment is therefore a property of shared mappings rather than shared inputs. The analysis further distinguishes between first-order failure (performance deterioration within a fixed evaluative structure) and second-order failure (misalignment of the evaluative structure itself). Because evaluative structures are typically fixed over the timescales relevant to within-episode adjustment, perturbations generate synchronized responses across agents. The framework yields testable implications on convergence conditional on evaluative similarity, break-type adjustment dynamics, and regime-dependent persistence. A stylized simulation, matched to an analytic benchmark, shows in the maintained setting that compression and alignment are determined by the spectral properties of the evaluative operator under isotropic inputs. The argument situates market comovement within a broader epistemic framework: systemic fragility depends not only on liquidity and balance sheet constraints, but also on the structure of evaluation governing decision-making.

Machine learning has substantially expanded optimisation capacity in financial systems. It has not resolved alpha compression, strategy crowding, or regime-driven instability. This article argues that these phenomena share a common structural source: optimisation under fixed evaluative criteria. Financial systems optimise over strategies while holding constant the criteria that define what improvement means — Sharpe ratio, volatility, drawdown, liquidity — inducing evaluative compression, a progressive contraction in the set of strategies admissibly superior under a shared evaluative ordering. As optimisation becomes more efficient, this contraction accelerates. Machine learning amplifies the problem it is presumed to solve. The article introduces a distinction between first-order failure — performance degradation under a given objective — and second-order failure — inadequacy of the objective itself. Contemporary systems are architecturally capable of detecting and responding to the former. They have no category for the latter. Analysis of the 2007 quant equity crisis demonstrates that shared evaluative criteria generate structural coupling across independently developed strategies, producing correlated losses that recalibration alone cannot resolve. Similar dynamics are identified in the March 2020 liquidity dislocations and the 2022 UK LDI crisis. To address this structural limitation, the article proposes evaluative operation: a minimal, constrained architectural extension in which performance criteria become admissible objects of system-level transformation, triggered by persistent signals of evaluative misalignment. A six-stage procedural pipeline specifies how such transformation operates in practice. The article develops evaluative accountability as the corresponding governance requirement, introduces evaluative arbitrage as a primary governance risk, and proposes three empirically testable predictions. The central implication is that, as machine learning scales in finance, the binding constraint is no longer optimisation efficiency but control over the criteria under which optimisation proceeds.

Contemporary AI systems are typically described in terms of behaviour, such as generation or autonomy. This paper argues that such classifications obscure a deeper structural distinction. Across machine learning, finance, and governance, systems optimise under fixed objective functions that remain external to their domain of operation. The paper introduces Objective-Layer AI as a category in which the objective function itself becomes an admissible object of system operation. This condition is not satisfied by existing approaches, including meta-learning and preference-based optimisation, where objectives remain externally governed. Using finance as a test case, the paper shows that optimisation under fixed objectives is structurally bounded. It further argues that a subset of out-of-distribution failures reflects mis-specified objectives rather than generalisation error. The inclusion of evaluation within system operation shifts governance from output supervision to the specification of constraints on admissible objective transformation. The paper advances a limited claim: systems operating under fixed objectives remain dependent on externally specified evaluation, regardless of behavioural sophistication. Objective-Layer AI is proposed as a structural criterion distinguishing optimisation-bound systems from those in which evaluation becomes operational.

Out-of-distribution (OOD) failure is usually framed as a problem of generalization: models degrade when deployment inputs differ from those represented in training. This framing is useful but structurally incomplete. It captures failure relative to a fixed evaluative standard, typically formalized as a loss function, reward signal, or learned preference model, while leaving unaddressed a prior question: whether failure can function, for the system itself, as evidence against that evaluative standard. This paper argues that contemporary machine learning systems cannot internalize that second-order possibility. Where the evaluative function L lies outside the system’s domain of operation, failure can only be processed as deviation under L, not as evidence that L is itself inadequate for the encountered domain. On this basis, the paper distinguishes between first-order failure, understood as degraded performance relative to L, and second-order failure, understood as inadequacy of L as an evaluative ordering over candidate outputs. It then introduces a minimal coherence-monitoring function C to distinguish performative failure from normative absence, and argues that current architectures reliably instantiate the former while lacking any system-internal mechanism for the latter. OOD failure is therefore reinterpreted not merely as a performance defect, but as a structural signal of epistemic enclosure: a condition in which evaluative standards govern behavior without themselves being subject to system-level revision. The contribution is classificatory rather than architectural. It clarifies a structural limit of contemporary AI systems, identifies the threshold at which epistemic enclosure would be breached, and situates OOD behavior within broader questions of evaluative control, epistemic agency, and responsibility.