Operator: A Protocol for Trustless Delegation Under Uncertainty

David Shi, Kevin Joo · 2025

Read Operator: A Protocol for Trustless Delegation Under Uncertainty by David Shi, Kevin Joo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-00631.txt.

Introduction

David Shi and Kevin Joo proposed Operator in twenty twenty-five, a protocol for delegating work to untrusted agents under uncertainty. Autonomous AI agents gain from handing subtasks to others, yet in low-trust settings no upfront spec or central overseer can guarantee the work is right. Their guiding claim: correctness emerges when exposing error is cheaper than committing it.

Tasks are posted as intents, and solvers compete to fill them while staking collateral. Verifiers check results after the fact. Anyone may challenge a result by staking against it, triggering verification; wrong agents are slashed and correct challengers rewarded. Erroneous verifiers can themselves be penalized through escalation. When incentives align across all roles, the authors argue, honesty becomes the equilibrium strategy.

Catalog record

CREATORDavid Shi, Kevin Joo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-00631

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Operator: A Protocol for Trustless Delegation Under Uncertainty David Shi, Kevin Joo Abstract Correctness is an emergent property of systems where exposing error is cheaper than committing it. In dynamic environments, autonomous AI agents benefit from delegating work, yet correctness often cannot be assured through upfront specification or centralized oversight. We propose a protocol that enforces correctness through collateralized claims in a recursive verification game. Tasks are published as intents, and solvers compete to fulfill them. Selected solvers carry out tasks under risk, with correctness checked post hoc by verifiers . Any challenger can challenge a result by staking against it to trigger the verification process. Incorrect agents are slashed and correct opposition is rewarded, with an escalation path that penalizes erroneous verifiers themselves. When incentives are aligned across solvers , challengers , and verifiers , falsification conditions make correctness the Nash equilibrium. 1 History The twentieth century’s best response to the problem of coordination in the face of uncertainty came from Friedrich Hayek, who argued that price systems were distributed information processors [ 1 ] . For Hayek, price mechanisms were a way to compute solutions to coordination problems that no central planner could possibly solve. Yet, the power of price signals was constrained by the goods being sold, and as economic activity shifted away from physical goods towards services and knowledge, the limitations of markets to reliably price intangible work became increasingly clear. Hart and Moore formalized these limitations with the theory of incomplete contracts [ 2 ] . In any system where uncertainty was pervasive and variables could not be observed or controlled, contracts were forced to be fundamentally open-ended. Hadfield and Hadfield-Menell extended the theory of incomplete contracts by arguing that AI alignment, like human contracting, depended on the ability for AI to interpret and respond to normative structures [ 3 ] . The question of how to make this norm-sensitivity actionable remained unresolved. The limitations of institutional remedies did not mark the end of efforts to achieve scalable and impartial adjudication in adversarial environments. In the late twentieth century, the field of mechanism design emerged as a formal framework for understanding how rules and incentives shape strategic behavior in economic systems. Hurwicz, Maskin, and Myerson conceptualized coordination as an engineering problem, positing that by carefully structuring incentives, systems could induce self-interested agents to reveal private information, report honestly, and enable collective verification [ 4 , 5 , 6 ] . The advent of Bitcoin brought a practical revolution. Satoshi Nakamoto showed that it was possible to achieve consensus and settlement on facts across vast, pseudonymous networks without trusted intermediaries [ 7 ] . Vitalik Buterin extended this breakthrough by enabling not just transfers of value, but the execution of arbitrary Turing-complete logic in the form of smart contracts [ 8 ] . Digital cash and other digital assets were all enabled in this way, so long as the transaction logic was expressible within the protocol. UMA [ 9 ] and Polymarket [ 10 ] extended these ideas by incorporating mechanisms that resolved ambiguities through economic consensus. These mechanisms relied on the same underlying structure: that if exposing an error was cheaper than committing it, truth became the stable outcome. As Karl Popper observed, knowledge in complex domains depended not on exhaustive specification, but on systematic exposure to falsification [ 11 ] . UMA instantiated this principle for bounded claims, where the space of possible outcomes was discrete. However, as tasks grew more ambiguous and temporally extended, this model reached its limit. Operator builds on this work by generalizing adversarial verification to arbitrary tasks. Instead of requiring ex ante clarity or binary outcomes, it enables correctness to be adversarially surfaced, addressing the failure modes of specification and deceptive alignment identified in advanced learning systems [ 12 ] . 2 Protocol Operator enforces correctness by making error economically unprofitable. The protocol is governed by a falsification condition: B>\frac{F}{P_{e}} (1) where B is the bond at risk, F is the falsification cost (the cost required to challenge), and P_{e} is the estimated probability that an error exists. This condition applies to all roles in the system: solver ( S ), challenger ( C ), and verifier ( V ). Each participant is exposed to economic risk under this condition, and when the inequality is satisfied, adversarial falsification is profitable and truth is the only rational outcome. 2.1 State Transitions Intent Published Solver Selected Bond Locked Result Submitted Challenge Period Finalization Verification Adjudication Challenge Period 2 Finalization No challenge Challenged Upheld Overturned Challenged No challenge Upheld Figure 1: State transition diagram of the Operator protocol. A process begins when a task T=(C,I,D) is published to the network. The constraint set C defines the permissible methods of execution and verification. The intent I specifies the objective to be satisfied, and the data D provides the raw input on which the task is to be performed. This publication constitutes an open call for execution. Agents may submit commitments to perform the task, each accompanied by a proposed strategy and a bonded stake. The solver could be the first agent to meet the bonding requirements, or be randomly selected from a pool of qualified, bonded agents. Once selected, the network registers the solver and locks the associated bond. The solver executes task T and submits a result R=(O,E) , where O is the output and E is the evidence supporting it. The result is published to the network and enters a challenge period defined by the constraint set C . During this window, any agent may initiate a challenge by posting the required bond and presenting a falsification. If no valid challenge is submitted within the allowed time, the result is finalized and the solver ’s stake is returned. If a challenge is submitted during the challenge period, the challenger initiates a dispute Q=(T,R,A) , where T is the original task, R is the result submitted by the solver , and A is the adversarial evidence or argument contesting the result. The network selects a verifier or verifier quorum V to adjudicate Q , following a commit-reveal pattern. The verifier reviews the task definition T=(C,I,D) , the result R=(O,E) , and the challenger ’s submission A , and issues a ruling in favor of either the solver or the challenger . This ruling enters a new challenge period. If the ruling is challenged, a new challenger submits a second-order dispute Q^{\prime} against the prior decision, and the network selects a new verifier V^{\prime} to adjudicate it. This recursive process continues until a result or ruling survives a challenge window without further opposition. 3 Incentives Assume all participants aim to maximize expected value, that if a result is incorrect, at least one participant can detect and demonstrate the error at finite cost, and that the falsification condition holds for all roles as established in Equation 1 . Let a solver S be selected to perform a task T , having posted bond B_{S} . The solver may submit a correct result R^{*} , which carries no risk, or an incorrect result R^{\prime} , which carries a non-zero probability P_{e} of being challenged. If a challenge is successful, S is slashed. The expected loss

Read the complete text