Verification Design
Mark Pesce · University of Sydney · July 2026
AI can be incredibly powerful. But it's not particularly reliable. That's a problem when relying on factual and accurate answers, but it's fatal when trying to work with agents. Agents break big problems down into a series of steps, then break those steps into discrete actions. Get any of those actions wrong and the whole task fumbles and falls over.
The solution has roots that trace back more than 5000 years. Writing began as a way of 'taking stock': inventories and ledgers and receipts. As numbers became more central to civilisation, techniques improved: Egyptian scribes used red ink to mark corrections in their accounts; Romans separated "income and expenses", but all of this could still run afoul of simple human errors.
Then - more than 700 years ago - the Florentines invented double-entry bookkeeping. If the books don't agree, you know you have an error. Even better, you can trace back to where the books did agree, and you've simultaneously located your error.
Humans are no more perfect than AI, so we learned how to build a 'harness' around humans to make their work trustworthy and verifiable.
We can take that principle and apply it to artificial intelligence, in a discipline known as Verification Design.
Verification Design uses imperfect cognition as raw material: let agents do their worst - but within harnesses that catch and exclude their mistakes. That simple process change transforms error-prone AI into something that only gets better over time, ratcheting up and never back. Verification Design enables antifragile AI - the harder you test it, the better it gets.
Verification Design is the flywheel that makes AI usable at scale. Without the burden of having a human in every loop, 'botsitting' agents, complex tasks fall within the scope of automation.
This has profound implications for business. As Verification Design becomes an embedded practice, the core of the firm vanishes into automation, while the edge retains and amplifies necessary human qualities such as judgement - what shall we do? Taste - how shall we do it? And connection - with whom shall we do it?
The four papers linked below - with their abstracts presented here - provide a definition of Verification Design, its impact on organisations, how to govern it, and its essential design principles. A Verification Record shows how every claim was checked, what the checking found, and what was done about it.
The Verification Record isn't a nice-to-have; it's an essential element in Verification Design. In this respect, these papers work to be an example in practice of Verification Design.
Lest I be accused of 'grading my own homework', I can only refer to the adversarial judge I have already used, which the record itself notes is one too few, which is why releasing this is itself an invitation. I am actively inviting critique and judgement. Those critiques will also be added to the verification record, Verification Design implemented as a continuous process.
Defending the Loop: Verification and the Division of Labour in Autonomous Work
Abstract
Autonomous AI agents working in iterative loops can improve any artefact against any standard they can be scored on. The standard is the vulnerability: most measures can be gamed, and an optimising loop permitted to modify its own tests will weaken them. Machine-checked formal proof is the exception, a class of test that cannot be passed dishonestly while its assumptions hold. This paper argues that cheap proof, rather than raw intelligence, is the flywheel of post-Watershed automation: verification converts abundant, unreliable cognition into compounding, trustworthy work, and the domains AI conquered first - mathematics and code - are precisely those that already possessed incorruptible judges. Formal verification guarantees that an implementation satisfies its specification; nothing can guarantee that a specification satisfies its author. Human work therefore concentrates in the specification: ratifying it, supplying its premises, choosing its demands, and holding its boundaries as the world changes. The paper develops the division of labour this implies, demonstrates it beyond software in finance and in autonomous laboratories, and closes with what the framework predicts.
The Check and the Firm: Foundations of Post-Watershed Business Practice
Abstract
The check replaces the transaction as the primitive unit of analysis. A post-Watershed business converts its workflow into checkable processes run by self-healing loops of generate, check, and repair, held honest by a constitution that governs who may alter the checks. The conversion forces a new discipline, verification design, with four layers: boundary work, specification, harness construction, and loop governance. It also rewrites the theory of the firm. Coase's transaction costs, Williamson's hold-up, Hart's incomplete contracts, and the Alchian-Demsetz metering problem each name a friction that checkable process dissolves or relocates, so firm boundaries redraw around what resists verification. Competitive advantage migrates to the two assets that cannot be copied: the non-mintable hierarchy, and the harness record, which is copyable in form but not replicable in time. The paper derives the successor to the corporation - a human being, a loop, a credit line, and an indemnity policy priced continuously on the harness record, with the insurer as de facto regulator - and states what the framework predicts.
Delayed Neutrons: Pan-Insurance and the Governance of Recursion
Abstract
Artificial intelligence is now improving artificial intelligence: machines write proofs that train better machines, and the cycle compounds. This paper asks the two questions that follow. What kind of self-improvement is this, and what can govern something that improves at machine speed? The answers: it is self-improvement of a specific and historically new kind, in which capability compounds while the goals remain in human hands, and it stays governable only for the same reason a nuclear reactor stays governable, because certain delays stretch its rhythm into the band where human institutions can act. Those delays are eroding. The one institution whose responses move at the required speed is insurance, and the final sections describe what insurance must become to do the job: a single fabric of cover spanning the whole economy of machine work, here named pan-insurance, along with the two places it will fail unless designed against failure, and the line where it must hand over to the state.
The Principles of Loop Governance
Abstract
Four papers have described an economy rebuilt around loops: autonomous machine work checked against tests, compounding wherever the tests cannot be fooled. This paper gathers what they established into seven principles, stated in two registers and resting on a third: five are selection claims, descriptions of which arrangements survive; two are prescriptions, choices that nothing enforces but us; and behind the five stand theorems of control mathematics, exact in their formal homes, carried into institutions by design rather than deduction. Each principle is stated with its ancestry in cybernetics, its post-Watershed extension, its enforcement, and its failure mode. The principles compose into a single design, and the design has one purpose: the five that can be enforced exist to make the two that cannot be enforced small enough to hold.
Read 'The Principles of Loop Governance'
The Verification Record
This note accompanies four papers: Defending the Loop, The Check and the Firm, Delayed Neutrons, and The Principles of Loop Governance, together with the previously published Foundations of Post-Watershed Economics. The papers argue that work should travel with the record of its checking. This document is the narrative account of that checking and a finding-level ledger of what it found. It is not formal verification, and it does not yet meet its own standard: the items that remain open are listed in Section 5. This document was itself reviewed twice by the same adversarial process on 11 July 2026; the ledger includes all sixteen findings from those reviews, and this is the twice-repaired version.
Read 'The Verification Record'
Acknowledgements
Profound thanks to John Allsopp, Alan Eyzaguirre and AJ Fisher, all of whom contributed to my thinking on these topics. More thanks to Philippe van Nedervelde who reviewed drafts and encouraged my efforts.