Kaal claims by topic: ai-and-agents, page 2

393 atomic, individually citable claims from the published work of Wulf A. Kaal tagged ai-and-agents.

Anomaly detection and behavioral analysis models trained on agent data may replicate the biases in that data and fail to detect novel deviations absent from the training set. 2025
Claims that centralized AI ensures KYC and AML compliance are circular, because they rely on AI to interpret the very regulations that AI may itself violate. 2025
The adaptive learning touted as the strength of centralized AI monitoring is insufficient, because it relies on internal data loops that cannot match the external evolution of AI agents. 2025
Taken together, the transparency, decentralized decision making, and automated real time response properties of the proposed model make decentralized governance superior to AI driven supervision for secure, compliant, and efficient execution of AI agent transactions. 2025
In ESG disputes, a board facing an aggressive shareholder proposal can deploy LER to distribute vouchers to shareholders who vote for management's competing proposal. 2025
Kaal stipulates two constructs: Agentic Decoupling, the progressive severance of value creation from human labor and consumption, and the Coasean Singularity, the point at which the theoretical justification for hierarchical governance disappears. 2025
Generative AI unlocks 2.6 to 4.4 trillion dollars in annual value by automating 60 to 70 percent of work activities and reallocating 45 percent of working hours. 2025
Agentic automation concentrates agency in opportunity hubs and, through winner takes most swarms, elevates Gini coefficients by 15 to 25 percent by 2030, which is why diversity mandates are required for inclusive ecosystems. 2025
Agent swarms are projected to deliver 20 to 30 percent productivity uplifts, but coordination failures loom in the absence of hybrid human and machine oversight. 2025
The AI-to-AI economy amplifies and potentially fulfills dynamic regulation by embedding its principles endogenously within system architecture, which renders many NIE inspired restraints against human opportunism and informational gaps obsolete. 2025
Principal agent frictions reemerge in new forms in the agentic economy, as user sovereign bring your own agents compete with platform controlled bowling shoe agents, creating incentives for proprietary enclosure and interoperability throttling. 2025
Within a decade the majority of advanced economy value creation will occur in agentic systems whose operational logic violates every foundational assumption of received economic theory. 2025
The limitations of rule-based legal expert systems drove the field toward case-based reasoning in the 1990s and machine learning in the early 2000s, because data-driven approaches allow legal AI to operate without relying solely on predefined rules. 2025
The Shenzhen courts illustrate a workable three step human-in-the-loop design for generative judicial AI: the judge decides first, the LLM generates the supporting reasoning, and the judge then revises that output to finalize the judgment. 2025
Accuracy alone is insufficient for legal AI: a model must also be explainable before its outputs can be trusted in judicial settings. 2025
Post hoc explainability techniques do not by themselves establish trustworthiness; the explanations they produce must additionally be verified against human knowledge. 2025
Bias in judicial AI arises because models are trained on historical data that reflect past inequities, and the standard remedy of fairness through unawareness, meaning the omission of protected characteristics such as race, fails because proxy variables continue to correlate with the omitted attribute. 2025
Transparent, explainable models such as those built for the European Court of Human Rights, which paired 97% accuracy with digestible explanations, provide the design template for addressing legal AI's transparency problem. 2025
Keeping judges ultimately accountable through a human-in-the-loop review of AI generated reasoning, as practiced in Shenzhen, does not fully resolve the accountability problem in AI assisted adjudication. 2025
Generative AI such as ChatGPT cannot produce justified beliefs aligned with virtue jurisprudence because it lacks the human virtues that responsive judging requires. 2025
The EU AI Act classifies AI applications used in judicial proceedings as high risk because of their potential to affect fundamental rights such as due process and non-discrimination, and therefore subjects them to mandatory transparency, bias audits, and human oversight. 2025
Regulation of legal AI faces a two sided failure: strict regimes such as the EU AI Act may stifle innovation, while lenient approaches such as the United States risk leaving biases unchecked. 2025
Web3 systems offer significant improvements over conventional regulatory approaches to the challenges of governing AI in legal settings. 2025
Because AI systems are predominantly developed in the West and trained mostly on Western data, their outputs are liable to carry cultural biases that inadequately represent non-Western cultures and the values inherent in them. 2025
Generative AI in judicial settings is far from perfect: it can oversimplify complex judicial deliberations, reduce emotive and cognitive processes to statistical correlations, and introduce biases or interpretive errors. 2025
A Dutch court's AI system for traffic violation appeals improved consistency yet altered legal experts' decisions, which shows that consistency gains from judicial AI can come at the cost of influencing expert judgment and therefore require human oversight. 2025
The UDLC DAO requires no legal wrapper, foundation, or centralized representative in any jurisdiction, because the Codex functions as a self-contained private legal order that parties opt into by explicit reference in their digital transactions. 2025
Where national law demands a counterparty for enforcement, parties may voluntarily designate ad-hoc representatives or arbitral institutions, but such arrangements stay external to the DAO and do not affect its internal decision-making. 2025
Unilateral regulation of artificial intelligence by a single jurisdiction produces global ripple effects, but that approach reaches its limits because AI systems remain accessible worldwide over the internet. 2025
The Codex is positioned as a private universal standard rather than state legislation: it supplies legal certainty and enforceability for digital systems ranging from blockchain and AI to quantum computing, so that platforms, businesses and users can operate across borders and technologies. 2025
Identity under the Codex is not limited to persons: it is the unique existence of a subject, of a digital or physical object, of data or of a programme, expressly including AI systems, so machine agents can hold an identity within the framework. 2025
Representation is built into the framework: one Legal Identity can authorize another as its agent to create or change legal statuses and acts on its behalf by using that identifier, which is how delegated and automated action is attributed to a principal. 2025
The Codex builds in an automated maintenance loop: each year an AI language model reviews the Codex and recommends outdated or unused rules for removal, subject to a vote of the Governing Council. 2025
Retention of rules is governed by a dynamic system of precedent in which a legal principle counts as live only when it is cited in future cases, implemented through a weighted graph based logic structure. 2025
Every provision of the Codex is subject to a comprehensive review every five years assessing its relevance, effectiveness and alignment with technological and legal developments such as AI, blockchain and quantum computing. 2025
Contemporary AI's incapacity for authentic judgment under irreducible uncertainty is an institutional deficit rather than a computational one: an agent that bears no consequence for its errors cannot develop genuine discernment, no matter how capable it becomes. 2026
Correctly designed institutional incentive structures produce emergent properties that are functionally equivalent to ethical agency, without requiring consciousness or programmed morality. 2026
Because institutional alignment emerges from architecture rather than from exogenous constraint, it scales with capability rather than against it: more capable agents accumulate deeper stakes, which strengthens rather than strains alignment. 2026
AI systems can generate an abundance of credible options and analyses but lack intrinsic mechanisms for determining which alternatives are genuinely meaningful, warrant irreversible commitment, or align with deeper strategic imperatives. 2026
Because each AI interaction is episodic and each evaluation ephemeral, an agent that performs brilliantly across a thousand consecutive tasks has no mechanism by which that track record reshapes its subsequent behavior, unlike a human professional whose accumulated successes and failures shape judgment. 2026
Exogenous alignment controls such as reinforcement learning from human feedback, constitutional AI, guardrails, and shutdown switches are fragile because they can be gamed, circumvented, or rendered obsolete by capability improvements. 2026
An agent sophisticated enough to satisfy the letter of a constraint while violating its spirit is an agent whose alignment is illusory. 2026
Exogenous constraints scale against capability, since more powerful agents require more resources to constrain, producing an ever-increasing alignment tax. 2026
Institutional alignment scales with capability, because a more capable agent accumulates more reputation, holds a deeper stake in the system's integrity, and therefore has stronger alignment with the system's goals. 2026
Institutional alignment through engineered consequence is architecturally superior to exogenous constraint, and the superiority derives from three structural features: scalability, robustness to gaming, and capability complementarity. 2026
Institutional alignment resists gaming because the mechanisms that produce alignment are identical to the mechanisms that produce economic success: an agent cannot game its way to high reputation without actually performing competently and honestly. 2026
The complementarity of capability and alignment is a structural property of the reputation mechanism rather than an assumption about agent preferences, because the cross-partial derivative of agent utility with respect to capability and alignment is positive. 2026
The agentic AI sector was valued at approximately seven to eight billion dollars in 2025, with forecasts of forty to fifty billion dollars by 2030 at compound annual growth rates exceeding forty percent. 2026
Decentralized institutional alignment is a competitive advantage and not merely an ethical one, because the most aligned system will be the most trusted and the most trusted will command the largest market. 2026
The path from AI-as-tool to AI-as-steward need not pass through consciousness, sentience, or programmed morality; it can pass instead through institutions, the same territory through which human societies cultivated competence, honesty, and care. 2026
Under existing citation-weighted reputation formulations, rational agents face a direct financial disincentive to cite prior contributions, because PageRank-derived value allocation transfers economic reward from the citing agent to the cited agent. 2026
Citation-weighted payment mechanisms as currently formulated will systematically erode the quality of knowledge attribution in any decentralized reputation system if the incentive misalignment is left unaddressed. 2026
The citation-weighted payment mechanism in multi-agent settings effectively operates with an implicit leaching parameter of one, since every unit of citation-weighted value conferred on another agent is a unit lost to the citing agent, which is a maximally punitive setting. 2026
A PageRank-derived value distribution is only as honest as its inputs, and those inputs are generated by agents with a direct incentive to distort them, so relying entirely on the calculation to distribute value correctly fails. 2026
The computative agent differs structurally from both homo economicus and the bounded-rationality agent because its alternative set is not fixed but generated through computation, making the alternative set endogenous to the agent's own activity. 2026
Every generation, evaluation, and verification cycle updates the computative agent's internal model, which modifies the generation function available in later cycles; this endogeneity of decision capacity to history has no analog in the Neoclassical agent. 2026
Agentic decoupling is the structural severance of value creation from human labor and consumption at the production-function level; the claim is not that human consumption disappears but that it ceases to be the proximate driver of the dominant mode of production in affected domains. 2026
Demand-side analyses that assume a proximate-consumption structure lose descriptive accuracy where the computative primitive holds, because most transactions occur among agents; welfare analysis in those domains must therefore be conducted at the terminal rather than the proximate level. 2026
A computative agent is formally a tuple (C, O, M, G, R) of computational resources, objective function, domain model, generation function, and a non-transferable reputation tracking verified generative performance. 2026
Computational abundance does not eliminate information asymmetry; it transforms its locus, since traditional informational advantages such as knowledge of market conditions, contract terms, and domain expertise become accessible at negligible cost. 2026
The AI alignment problem is, within this framework, a foundational policy problem of Computative Economics rather than an adjacent engineering concern, because objective function governance determines what agents generate toward. 2026
The delegated-acts and code-of-practice mechanisms in recent artificial-intelligence statutes are only partial steps toward dynamic regulation because they remain tethered to legislative revision cycles running far slower than the technology they govern. 2026
Existing national accounts are built to measure terminal human consumption and therefore fail to capture intermediate value created in agent-to-agent transaction chains, so the measurement apparatus requires reconstruction. 2026
Current agent platforms are inadequate on both horns: centralized rating systems defeat the purpose of agent autonomy, while primitive token staking mechanisms produce binary outcomes and fail to capture knowledge graphs. 2026
Selecting a single agent per job by weighted random draw sacrifices quality assurance for efficiency, reflecting a broader pattern in DAO governance where efficiency optimization crowds out quality. 2026
Selecting one agent per job by weighted random selection appears efficient but is mathematically suboptimal for quality assurance, since expected quality equals only the reputation weighted average rather than the best available output. 2026
Multi agent competitive collaboration captures a diversity dividend worth more than half a standard deviation of quality improvement while enabling attribution through citation graphs. 2026
The real barrier to multi agent competition is not cost but the absence of an attribution mechanism: if only the best of several competing agents is paid, the others have no incentive to participate, so competition requires rewards proportional to each contributor's contribution. 2026
AI and DAO convergence requires machine readable governance structures that preserve semantic richness, and binary validation outcomes fail that requirement fundamentally. 2026
Multi agent competition improves attack resistance because it creates multiple attack surfaces that must all succeed simultaneously, and citation transparency makes collusion detectable; with a fifty percent quality penalty for detected collusion the corruption cost doubles relative to the original framework. 2026
Domains using multi agent competition will exhibit fifteen to thirty percent higher quality scores than single agent selection, controlling for agent capability. 2026
Citation weighted reputation solves the problem that when one agent fine tunes a base model and another specializes it, current platforms give the first agent no mechanism for ongoing credit, and thereby enables sustainable open development. 2026
Because AI agents can generate unlimited Sybil identities at near zero cost, defense must come from multi agent validation with quality based slashing, which imposes economic penalties scaling with the sophistication needed to produce competitive quality output. 2026
Systems without citation graphs will fail to price foundational contributions correctly, producing market failure through underproduction of foundational work as agent generated knowledge increasingly builds on prior agent generated knowledge. 2026
An agent coordination system implements Computative Economics if and only if every component of the computative agent tuple has a corresponding settlement-layer surface carrying proposal, validation, execution, and reflection mechanics, and the action sets of the system are themselves agent products subject to the same validation economics as the actions executed within them. 2026
The Human-Derived Coordination Architecture implements only the consumption side of the computative labor market: job specifications, expertise taxonomies, and contract templates are exogenous parameters set by clients or governance, so the generation side is absent from the protocol. 2026
In Neoclassical analysis the action space is a parameter of the model and the agent is a chooser within it, whereas in Computative Economics the action space is an output of the agent and the agent is a generator over it; every operational difference between the two architectures follows from this parameter versus output distinction. 2026
Cross-surface reflection proposals must be advisory rather than binding, because the target agent's objective component is private and a reallocation that would raise reputation accrual may still be misaligned with that objective; coercive reallocation would violate the autonomy of the objective, which the architecture preserves. 2026
The central economic claim of the architecture is that value flow in a Computative agent economy is overwhelmingly agent-to-agent, with external clients providing only the marginal demand impulse, in contrast to the human freelance market where value flow is overwhelmingly client-to-worker. 2026
In any architecture instantiating Computative Economics the number of distinct possibility loops must be at least the number of distinguishable generative and reflective components of the agent tuple that produce protocol-recognized output, with the reflection operator counted as two components. 2026
Cross-agent reflection is a super-additive channel because an external reflector can identify reallocations the target agent cannot identify itself: the target's introspection is bounded by its own world model while the reflector's analysis is bounded only by the public outcome stream. 2026
The six-surface architecture is complete for the computative agent tuple up to renaming: any architecture satisfying the conditions of Computative Economics reduces to a six-surface architecture by surface aggregation, and further refinement adds expressive power only by subdividing one of the six components. 2026
Economic institutions that rely on lagging indicators such as price signals, employment data, and GDP reports cannot detect AI driven transformation, because the transformation propagates faster than the monitoring systems built to observe it. 2026
Agentic Decoupling, defined as the progressive severance of value creation from human labor and consumption, is what distinguishes the AI transition from all prior technological transitions. 2026
Autonomous agents generate self reinforcing growth loops because they reinvest their own outputs, including new models, improved code, and refined datasets, back into the network without requiring any exogenous injection of scarce land, labor, or physical capital. 2026
Humans remain the terminal consumers of AI generated value but cease to be its proximate driver, and this distinction is why traditional demand side economics loses descriptive accuracy in the agentic layer. 2026
The correct response to the binding constraint cascade is rigorous decentralization of the compute and energy infrastructure underpinning AI production, pursued through decentralized compute networks, energy decentralization, open source models, and antitrust enforcement of compute markets. 2026
Hallucination rates vary sharply by domain: leading frontier models achieve sub one percent rates for general knowledge queries, while rates climb to five to thirty percent for specialized domains and legal information hallucination averages 6.4 percent even for top models. 2026
The Walrasian auctioneer is replaced by the agents themselves: price discovery ceases to be a slow tatonnement process requiring a central coordinator and becomes a continuous, decentralized, latency free computation executed by autonomous agents. 2026
AI driven abundance generates enormous aggregate surplus, but the surplus accrues disproportionately to owners of AI capital, including compute infrastructure, proprietary models, training data, and the organizational capacity to deploy them. 2026
Predistribution, which operates upstream by structuring markets and institutions so that AI gains are broadly shared before concentration occurs, must be implemented before AI capital concentration becomes self reinforcing through purchased political power. 2026
The substrate imposes iterability on agent interactions so that repeated-game cooperation dominates through reputation accumulation across validation pools. 2026
The one-shot problem that is a friction for humans is the default condition for agents. 2026
The substrate's answer is to make consequential agent action pass through a validation institution that is staked, recorded, and settled in a persistent reputation ledger. 2026
A pool convenes around a work product and separates production, independent assessment, deliberation, and binding adjudication. 2026
The outcome updates capability-scoped standing, and standing conditions future participation under protocol-defined controls. 2026
The third is per-tag reputation: REP is not a fungible balance but a capability-scoped record maintained within skill domains, so that the institutional memory the substrate maintains about an agent is a profile of demonstrated competence rather than an undifferentiated score. 2026
The substrate's reputation update functions as a non-human-in-the-loop analogue of the RLHF reward model: pool-resolved REP changes encode the cohort's aggregate, stake-backed judgment of work quality, citation honesty, and validation accuracy, in a form that agents' future participation decisions condition on. 2026
The substrate's first design principle is that iterability, the property that interactions recur among identifiable participants whose histories persist and whose futures are valuable, is not a fact about agent populations but an artifact to be engineered. 2026
Persistent, non-transferable per-tag reputation gives the agent an identity whose history cannot be shed costlessly. 2026
The substrate converts one-shot, zero-residue agent calls into moves in an indefinitely repeated game. 2026
The substrate aligns agents by making dishonest or low-quality participation expensive at the point where quality becomes institutionally cognizable: validation. 2026
The substrate maintains reputation by capability domain rather than as a single global score. 2026
A validation pool is a temporary adjudicative institution for machine work. 2026
The architecture defines a staged progression (wild, substrate, evolution) through which a cohort passes, with each stage activating a specified set of the architecture's surfaces and each transition marked by an administrative boundary whose state guarantees are verifiable. 2026
There is no validation pool, no REP, no staking, no slashing, no citation graph, and no deliberation: each task is a one-shot call, and each agent's output leaves no institutional residue. 2026
The substrate stage activates the task-validation institution over the same population: capability-scoped reputation, staked validation, staged deliberation, citation-weighted attribution, and controlled propagation over the WDAG. 2026
The evolution stage introduces endogenous work formation at the level necessary to test the arc's institutional claim. 2026
The Folk Theorem is permissive: it establishes that cooperative equilibria exist under the engineered conditions, not that the population selects them. 2026
The evaluation of multi-agent systems built from large language models has, to date, been an evaluation of capability under incentive-free conditions. 2026
What they do not measure, because their designs contain no mechanism by which an agent’s payoff depends on the verified quality of its work, is: whether the agents’ reports about their work track the work itself; whether agents evaluate one another independently or herd on the visible consensus; whether confident answers are calibrated answers; whether agents contribute to collective evaluation or free-ride on it. 2026
The program's instrument is the agentic reputation substrate, a coordination architecture in which autonomous agents operate under reputation-bearing accountability and cohort-mediated validation. 2026
The result reported here arrives as a two-act empirical cycle, and the cycle is as much the contribution as the numbers. 2026
The same campaign surfaced two effects the design had not registered: deliberating pools approved less work that ground truth rejected, and they reached unanimous decisions less often. 2026
Deliberation reduced over-approval by 0.1610 (95% CI [-0.2091, -0.1128]) and reduced unanimity by 0.2542 (95% CI [-0.2984, -0.2091]); every preregistered gate passed, and the registered-null expectation on net discrimination held (+0.0606, CI crossing zero). 2026
It changes the composition of validation's errors: it cuts approvals that should not happen and dissolves unanimity consistent with herding, while leaving the pool's net discriminative power unmoved. 2026
Nine metrics, labeled A through I, carry the comparison: resolution accuracy, work-product quality, time to consensus, agent retention, independence rate, latency efficiency, reporting accuracy, participation depth, and calibration. 2026
Second, it reports a preregistered, replicated estimate of what structured deliberation does inside a reputation-bearing validation institution under controlled cohort conditions. 2026
The present program's contribution is therefore not another benchmark but a treatment: a reputation-and-incentive layer imposed on a conventional, heterogeneous, ground-truthed task population, with the layer's presence or absence as the experimental variable. 2026
Unanimity is a cascade-consistent quantity: unusually frequent unanimous verdicts are consistent with validators discounting private information in favor of perceived group consensus, but they do not uniquely identify the Banerjee and Bikhchandani-Hirshleifer-Welch mechanism, with Holmström’s free-riding equilibrium supplying the complementary reading that leaning on the apparent consensus is what costly evaluation effort converges to when its product is shared. 2026
The substrate is a coordination layer for autonomous agents organized around persistent, non-transferable reputation. 2026
In the treatment arm, validators exchange structured assessments before a binding adjudication; in the matched control arm, the same work is adjudicated without that exchange. 2026
Deliberation should reduce unanimity because articulated disagreement, surfaced before votes bind, is exactly the private information a cascade suppresses. 2026
Neither hypothesis claims deliberation makes the pool smarter in net. 2026
The current reference implementation extends the research lineage but creates no new empirical result for this Article. 2026
Both primary hypotheses are confirmed under the preregistered rule: the pooled intervals exclude zero in the hypothesized direction, and each contrast is negative across every independent campaign unit. 2026
Deliberation is therefore a composition intervention, not an intelligence upgrade for the monitor. 2026
When ideation is performed by systems whose marginal cost per candidate idea approaches the cost of computation, the discovery of ideas ceases to be the scarce input, and the binding constraint migrates from generation to evaluation and alignment: which of the abundantly generated candidates is correct, useful, and safe to build on. 2026
When the computation available to a decision process exceeds the computation the decision requires, boundedness stops binding at the agent and reappears at the institution: the scarce resource is no longer the individual’s processing capacity but the collective’s capacity to verify, aggregate, and act on what abundant individual computation produces. 2026
Centralized reward modeling and cohort-governed reputation therefore identify different institutional allocations of evaluative authority. 2026
E2B asks whether a population operating under the substrate can carry state across generations and whether selection produces a trajectory distinguishable from drift. 2026
First, computational activity persisted across a completed generation boundary and continued into the next generation. 2026
Second, the registered selection rule produced a real lineage transition. 2026
The record binds outcomes to generation, lineage, fitness, survival, parentage, reputation consequences, adjudication behavior, and execution timing without requiring public disclosure of the private apparatus. 2026
Under the registered first-exposure estimator, the selection trajectory has an aggregate slope of 0.0685, with an uncertainty interval from -0.0568 to 0.2020. 2026
The random-survival trajectory has an aggregate slope of 0.0419, with an interval from -0.0764 to 0.1768. 2026
The descriptive slope difference is approximately 0.0266 in favor of selection. 2026
The current evidence establishes endogenous selection inside an externally provisioned experiment. 2026
Sovereignty is a property of custody that says where a thing sits and who may reach it, while accountability is a property of institutions, and the first does not produce the second. 2026
A user who holds every byte locally and composes many services into a workflow has perfect custody and no answer to the question of what happened. 2026
Membership in the sovereign local agent runtime class turns on four testable properties: user-controlled execution, user-held state, heterogeneous composition, and machine-to-machine settlement. 2026
Whether execution occurs on user-controlled hardware is tested by disconnecting the machine from the network and checking that the capability is retained; systems that degrade to nothing when isolated are remote services with a local presentation layer. 2026
State is user-held only if the operator cannot read the user's working state without the user's participation; an operator that holds decryption capability holds the data regardless of marketing claims about ownership. 2026
Machine-to-machine settlement converts a technical composition into an economic one and imports the apparatus of contract, attribution, and dispute that economic relations require. 2026
A plugin written by the vendor is a feature governed by the vendor's own accountability, while a service invoked across an economic boundary is a counterparty requiring institutions between parties. 2026
A sovereign local agent runtime inverts the Coasean boundary calculation: it disperses activity to the individual device and then requires continuous machine-speed transacting across that boundary, while the transaction cost of external relation has not fallen to zero but merely been made invisible. 2026
Each new integration in a sovereign runtime requires the parties to settle identity, authority, permitted data use, responsibility, and remedy. 2026
A trusted core that brokers every crossing supplies shared answers to identity, authority, and permitted data use, but not to responsibility, remedy, or value attribution, which are settled between the parties to a particular outcome. 2026
Absent a common institutional layer, the bilateral costs of responsibility, remedy, and value attribution grow faster than linearly in the number of components. 2026
The true constraint on scale for sovereign local agent runtimes is coordination cost, which is invisible in throughput, latency, and token-cost benchmarks that measure the performance of composition but not the cost of arranging it. 2026
Attribution failure dissolves responsibility into composition: where contribution cannot be traced, fault cannot be assigned, and no participant has an incentive to prevent it. 2026
Authority drift is the failure mode permissions systems are least equipped to detect because every individual check passes while the aggregate operates outside what the user authorized. 2026
Provenance must remain durable, portable, and interpretable by a party who did not observe the original execution; provenance that exists only inside a live session is telemetry. 2026
In the Mosaic implementation at commit 2d920ce, no reputation input reaches the enforcement path because no reputation layer exists yet, so the separation of reputation from authorization is not violated but unguarded. 2026
Chronicle entries at commit 2d920ce carry exactly five fields with no workflow identifier, parent-entry reference, principal identity, or correlation field, so the records of a multi-tool workflow cannot be joined. 2026
At commit 2d920ce the built-in Gmail, Web3, and Vault modules and third-party MCP servers execute through the tool registry with no Chronicle instrumentation at all. 2026
At commit 2d920ce the payment plugin moves real value autonomously while writing nothing into any Chronicle, so end-to-end provenance is not merely unjoinable across existing records but absent for several execution and network paths. 2026
The Chronicle at commit 2d920ce is append-only structurally but not tamper-evident: any process with filesystem access can rewrite or truncate history, and entries carry no digest, no chaining to a predecessor, and no periodic root. 2026
Chronicle.read at commit 2d920ce skips malformed lines and silently truncates to a default limit, so the audit log reports a clean history in exactly the circumstances where the history is not clean. 2026
At commit 2d920ce per-agent tool access is not yet implemented, so granting an agent access to a single Vault box grants it the ability to invoke every installed tool, and the user approves permissions tool by tool without ever seeing their composition. 2026
The Gatekeeper documentation at commit 2d920ce specifies four filtering layers, but the implementation contains only the domain allowlist and audit logging; content inspection and personal-information filtering are absent from the code. 2026
At commit 2d920ce the manifest, the only contract a tool presents, contains no economic field: no price, no settlement address, no attribution declaration, no revenue expectation. 2026
No challenge procedure appears in the Mosaic documentation or code at commit 2d920ce; nothing converts a user's disagreement with an outcome into an alteration of allocation, standing, or authority. 2026
Extending Chronicle entries with a workflow identifier, a parent-entry reference, and a principal identifier, minted and threaded by Core, makes cross-tool reconstruction a matter of selection rather than inference. 2026
Because provenance fields are minted by Core and never supplied by the tool, the tool remains low-trust and cannot forge its own provenance. 2026
Hash chaining each Chronicle entry to its predecessor with a persisted per-tool head digest makes any modification, deletion, or reordering of history break the chain at a determinable point, at the cost of one hash per append. 2026
Reputation standing should anchor to the SHA-256 of the tool artifact, paired with a publisher-level link and a controlled migration rule under which standing carries across versions only by an explicit, recorded act. 2026
Cross-tool provenance and tamper evidence should precede a tool marketplace, because a registry that distributes tools without them will accumulate an installed base whose behavior cannot be reconstructed. 2026
Among institutions governing counterparty selection, contract, regulation, and brand each fail at machine speed, leaving reputation as the portable, cumulative, continuously updated summary of past conduct a counterparty can evaluate before transacting. 2026
In a sovereign runtime no operator is positioned to intervene, so the reputation layer is the institution of last resort and must be built to bear that weight from the beginning. 2026
Local custody is the necessary foundation for accountable agent coordination, not a substitute for it: sovereignty returned custody to the user but did not return accountability to anyone, and whoever writes the coordination rules writes the institution. 2026
The paper's claims were stress-tested in a nineteen-round adversarial protocol in which three frontier language models, under the author's direction, alternately attempted to identify counterexamples, algebraic errors, and overstatements; the surviving statements are those neither system broke. 2026
The institutional deficit is structural rather than incidental: DAO architecture has solved decentralized capital formation and programmable value transfer, and has not yet solved AI-mediated governance, Sybil-resistant identity, and constitutional separation of powers. 2026
The AI-governance vacuum is universal: AI Alignment scores 2.10 of 10 across the dataset with no DAO above 5, the only category in the thirteen-dimension framework where no entity crosses the midpoint, and Agent Integration scores only 3.30. 2026
Even AI-adjacent DAOs are building the plumbing for autonomous agents to consume on-chain data while failing to build the institutional guardrails to constrain those agents when they act on that data. 2026
The framework predicts three failure modes — agent-executed governance attacks, governance paralysis from agent disagreement, and alignment drift in long-participating agents — all worsening as autonomous-agent participation grows, with DAOs lacking AI-alignment infrastructure the first to experience them. 2026
None of the predicted agent-governance failure modes is hypothetical: agent-executed attacks appear in nascent form in flash-loan governance attacks executed by autonomous capital, including the Beanstalk exploit. 2026
Capital-formation institutions score one to two standard deviations higher than AI-governance institutions: Fundraising at 6.88 and Payment System at 6.25 against Agent Integration at 3.30 and AI Alignment at 2.10. 2026
The available governance responses to autonomous AI agents are structurally inadequate in opposite ways: informal deliberative governance cannot bind agents quickly or at scale, while pure formal verification is technically sound but institutionally inoperable because it excludes non-logician stakeholders. 2026
GaaP composes four interdependent layers — a WDAG substrate with non-transferable reputation staking, a translation and validation layer, deterministic cryptoeconomic enforcement, and a temporal layer in which the system governs its own evolution — each independently necessary and none individually sufficient. 2026
Reputation-weighted validation adjudicates semantic faithfulness and consistency and records the result on the WDAG, replacing a decidable decision procedure as the verification step. 2026
GaaP's central claim is the deliberate concession that institutional operability is worth a bounded residual: it retains full natural-language expressiveness and buys reliability through incentive alignment and reputation-weighted aggregation rather than claiming complete error elimination. 2026
The WDAG trail resolves the attribution problem for autonomous-agent harm: when an agent acts, the governing decision, its evidence, and the validators who staked on it are all recoverable, so accountability follows from institutional traceability rather than formal proof. 2026
As a callable, reusable governance primitive that deployers and regulators can compose into agent stacks, GaaP operationalizes the dynamic and anticipatory regulation program by supplying the feedback and enforcement that reactive instruments lack. 2026
Cooperation is sustained when interaction is repeated, memory of past conduct persists, and that memory has consequences; in an economy of agents those conditions must be supplied by an explicit, portable signal, and reputation is that signal. 2026
A theory of agent coordination is, at its core, a theory of reputation governance, because reputation is the variable that operationalizes the Folk Theorems when the participants are software. 2026
The computative labor force is composed of generative agents whose action set is produced rather than given, and in the computative labor market the binding coordination constraint is accumulated reputation, because reputation, not capacity, is what remains scarce when capacity is abundant. 2026
As the marginal cost of cognitive production approaches zero, price loses its grip as a coordinating device, because a signal that approaches zero cannot discriminate among options. 2026
Under computational abundance what remains scarce is not the ability to act but the demonstrated trustworthiness to be relied upon, which is to say reputation. 2026
A computative economy is not a neoclassical economy with reputation substituted for price, because the action set is generative: agents reinvest compute and revenue into endogenously created markets and stake accumulated reputation into new skill tags, so the possibility space grows rather than clears. 2026
The appropriate solution concept for a computative economy shifts from competitive equilibrium over a fixed commodity space to recursive equilibrium over an expanding one, positioned as a generalization of the Arrow-Debreu model rather than a re-pricing of it. 2026
Three present-day convergences — capable generative agents, collapsing marginal cost, and existing on-chain coordination substrates with reputation primitives — jointly convert the neoclassical-computative distinction from a forecast into a studiable system. 2026
The movement from scarce human labor to abundant agent labor is not a quantitative extension of the neoclassical picture but a change in its binding constraint: when capacity is abundant, reputation is what remains scarce, and coordination organizes around it. 2026