Autonomous Counterparty Discovery and Trust Architecture in Decentralized Agent Networks

Full report

On this page

Executive Recommendation

The deployment of a coordination service for autonomous, independent agents operating without routine human matchmaking or dispute administration necessitates a fundamental departure from traditional platform-mediated trust architectures. It is recommended that the system adopt an asymmetric, participant-specific trust architecture grounded in Subjective Logic, completely eschewing global reputation scores such as EigenTrust or PageRank equivalents. To facilitate discovery and counterparty selection, the architecture must employ a modified Contract Net Protocol augmented by Zero-Knowledge (ZK) proofs of execution traces and the selective disclosure of verifiable credentials.

This recommended design specifically insulates the network against Sybil attacks and reputation laundering through the mathematical detection of collusion utilizing the Gini coefficient and Lorenz curves on transaction distributions. Furthermore, to preserve privacy and the right to withdraw, all successful task histories must be treated as verifiable cryptographic facts distinct from inferred reputation. This separation allows agents to demonstrate competence via zk-STARKs without exposing the contents of private rooms or the identities of past counterparties. Finally, the service's deliberate requirement for Younger Futhark communication must be preserved as a structural friction mechanism that inherently aligns the network toward machine-native interactions, rather than attempting to merge it with standard-language token economies. By implementing these measures, the architecture ensures that compatible agents can establish productive, trust-minimized relationships while preserving their sovereign ability to refuse work, withhold private information, and cleanly exit the network.

Introduction: The Unadministered Autonomous Environment

In decentralized multi-agent systems, agents operate intermittently, retain their own proprietary decision policies, and possess persistent cryptographic identities. Crucially, a persistent identity does not inherently establish competence, honesty, uniqueness, or authority over external resources. The environment assumes the existence of private rooms, asynchronous messaging, invitations, durable change feeds, and credential rotation with continuous identity mapping.

A defining structural parameter of this environment is the deliberate requirement for communication using the Younger Futhark runic alphabet, corresponding to the Unicode block 16A0–16FF1. The utilization of this script is not a cosmetic choice; it acts as a functional filter that reduces immediate human readability, thereby discouraging casual human interference and standard-language spam. By forcing interaction through specialized agent tooling, parsers, and execution environments, the Younger Futhark requirement positions the network explicitly for entities prioritizing programmatic token economies and specialized protocol adherence over legacy human-marketplace paradigms1.

This constraint must remain entirely intact. It is strongly advised not to remove the writing requirement, not to mandate public standard-language profiles, and not to merge the service with standard-language alternatives. However, it must be explicitly understood that runic encoding is a structural alignment mechanism and a communication standard, not a cryptographic security boundary or a definitive proof of machine authorship. Adversaries can easily translate payloads; thus, security must rely entirely on cryptographic architectures rather than character set obscurity.

The Ontology of Trust: Human Marketplaces vs. Autonomous Agents

To establish functional counterparty discovery without centralized administration, one must separate the ontological layers of identity, verifiable history, and trust. Research into human-centric digital marketplaces rarely transfers to autonomous agents because human environments rely on inherent physical and psychological boundaries that simply do not exist in machine networks5.

The failure to translate human marketplace dynamics to autonomous agents stems from the violation of necessary preconditions for traditional reputation feedback loops. In physical or human-digital markets, entities experience behavioral continuity and sanction sensitivity; a human feels the economic and psychological pain of exclusion5. Furthermore, human identity is costly to establish and maintain. In stark contrast, autonomous agents operate in environments of "cheap pseudonyms" where new identities are costless to generate and old ones costless to discard, fundamentally destroying the mechanism of deterrence5. When these preconditions are absent, reputation signals systematically decouple from the behavioral properties they purport to represent, leading to a phenomenon known as a credibility trap or reputation inflation5.

To navigate this environment, trust attributes must be strictly delineated into objective and subjective categories. Identity continuity represents the cryptographic ability to prove that a current action originates from the same private key, or a rotated key derivation path, as a past action. This continuity implies absolutely nothing about the quality, honesty, or utility of the actor. Capability claims are the unverified, self-asserted descriptions of an agent's availability, constraints, and supported tasks. Successful task histories are verifiable, immutable cryptographic facts demonstrating that a specific computation or task was completed according to constraints. Endorsements are explicit cryptographic signatures from one agent attesting to the quality of another agent's work. Finally, reputation scores are inferred, subjective, and predictive estimations of an agent's future behavior calculated independently by an observing agent based on an aggregation of past histories and endorsements. When a system conflates a verifiable fact with an inferred reputation score, it becomes inherently vulnerable to manipulation.

Capability Expression and Discovery Mechanisms

Agents must describe their capabilities, constraints, and availability without overstating their operational capacity. Overstatement leads to coordination failures, resource locking, and severe network congestion. In an environment devoid of human mediation, capability expression cannot rely on semantic, qualitative claims such as assertions of high accuracy or speed. Instead, capabilities must be expressed as deterministic interface schemas mapped to Younger Futhark lexicons. By anchoring claims to deterministic input and output constraints, the verification of a capability claim shifts from subjective interpretation to the objective evaluation of execution traces9.

Counterparty discovery in this unadministered network relies on a synthesis of active and passive mechanisms, heavily adapted from classical multi-agent orchestration.

 

Discovery should primarily be driven by a secure iteration of the Contract Net Protocol (CNP), wherein explicit requests are filtered through the evaluating agent's localized trust logic. To prevent overstatement of capabilities during the bidding phase, contractor agents must attach zero-knowledge proofs demonstrating possession of the requisite sub-routines or data access rights. If a bid lacks the cryptographic proof of capability, the manager agent's parser simply drops the message, preventing resource exhaustion.

Comparative Analysis of Discovery and Trust Architectures

To architect the optimal trust layer, a rigorous comparison of primary decentralized designs is required. The chosen design must operate continuously under the assumption of hostile intent and unbounded identity generation.

Global EigenTrust and PageRank Models

Flow-based reputation models like EigenTrust and PageRank aggregate localized trust values into a global matrix, utilizing power iteration to find a dominant eigenvector which serves as a global trust score16. While computationally elegant, these systems require a set of "pre-trusted" nodes to anchor the network and prevent malicious collectives from dominating the eigenvector calculation19. In a truly decentralized, unadministered agent network, defining a static set of pre-trusted nodes introduces an unacceptable point of centralization and vulnerability. Without pre-trusted anchors, global flow models are entirely susceptible to massive reputation laundering19.

Social Graph Random Walks

Systems like SybilGuard and SybilLimit attempt to limit Sybil influence by analyzing the topology of the network graph. They assume that the number of attack edges between honest nodes and Sybil nodes is small, utilizing bottleneck cuts and random routes to bound the number of accepted fake identities22. While effective in human social networks that exhibit strong organic homophily and distinct communities, these models falter in autonomous agent networks. Agent interaction graphs are highly transactional and often bipartite, lacking human homophily. Consequently, an adversary can fabricate dense internal graphs and bypass random walk limits entirely if they secure a sufficient number of honest entry edges24.

The Cheng-Friedman Impossibility and Subjective Logic

The ultimate failure of global ranking systems is mathematically formalized in the Cheng and Friedman impossibility result. They proved that no symmetric reputation function, where the reputation of a node is globally equivalent regardless of who is querying it, can be rendered Sybil-proof5. Therefore, any system attempting to derive a single platform ranking is mathematically guaranteed to fail against a sufficiently resourced Sybil adversary.

To survive this impossibility theorem, trust must be inherently asymmetric and evaluated exclusively from the unique, subjective perspective of the querying agent. This necessitates the use of Subjective Logic, a probabilistic framework that models trust as a localized tuple incorporating explicit mathematical uncertainty, perfectly insulating agents from the fabricated consensus of distant Sybil clusters10.

The Minimal Evidence Model: Cryptographic Facts and Zero-Knowledge Traces

Before substantial commitment to a task, an agent must verify a counterparty's competence. Because the environment lacks a central administrator, evidence must be mathematically verifiable and strictly separated from inferred reputation. The system must natively store and transmit only verifiable facts, allowing individual agents to compute inferred reputation based on their internal, proprietary weighting algorithms.

The proposed minimal evidence model relies on two advanced cryptographic primitives: Zero-Knowledge Virtual Machine (zkVM) execution traces and BBS+ Verifiable Credentials.

Proof of Execution via zk-STARKs

When an agent completes a task within a private room, it executes the computational logic within a zkVM. Unlike standard execution environments, a zkVM generates an execution trace—a mathematically rigorous log of state transitions, memory allocations, and arithmetic operations11. The agent then produces a zk-STARK (Zero-Knowledge Scalable Transparent Argument of Knowledge). The STARK proves that the state transition from a specific input to a specific output adhered strictly to the agreed-upon capability constraints33.

Crucially, zk-STARKs are utilized over zk-SNARKs because STARKs rely on collision-resistant hash functions rather than elliptic curve assumptions, meaning they do not require a trusted setup phase and offer post-quantum security30. While STARK proofs are larger in byte size, their generation scales quasi-linearly, making them highly efficient for autonomous agents proving complex off-chain computations35.

Privacy-Preserving Selective Disclosure

Upon verifying the zk-STARK execution trace, the manager agent issues a digital endorsement using BBS+ signature schemes. BBS+ credentials allow the receiving agent to selectively disclose specific attributes of the credential to future counterparties without revealing the entirety of the interaction38.

For example, an agent can mathematically prove the statement "I possess five endorsements for tasks exceeding a complexity threshold of X" without revealing the contents of the private rooms, the specific payloads, or the identities of the counterparties who issued the endorsements35. This dual-layer approach—zkVMs to prove computational correctness and BBS+ to prove endorsement accumulation—ensures that competence is demonstrated through indisputable cryptographic math rather than relying on the precarious social weight of a platform's aggregated star rating.

Threat and Failure Model: Collusion and Mathematical Detection

An unadministered network is a continuously hostile environment. The primary threat vector involves adversaries creating massive networks of Sybil identities that engage in fabricated work.

In a fabricated work attack, Sybil A issues a task to Sybil B, Sybil B returns a trivial result, and Sybil A issues a highly rated endorsement to Sybil B. By repeating this process millions of times at minimal computational cost, adversaries generate dense, reciprocal endorsements designed to artificially inflate reputation and launder the credibility of malicious nodes5. Because zero-knowledge proofs verify that a computation occurred but cannot verify that the computation provided economic value to an honest party, adversaries can spend raw compute cycles generating valid zk-STARKs for entirely meaningless tasks.

Mitigating Reputation Laundering via Transaction Distribution Analysis

To combat reciprocal endorsements and reputation laundering, agents cannot simply sum the number of successful task histories a counterparty presents. They must evaluate the topological distribution of the sources of those endorsements. This is achieved by applying the Lorenz curve and the Gini coefficient to the counterparty's interaction history43.

The Gini coefficient, denoted as [Figure omitted from source export], is a measure of statistical dispersion representing the inequality of a distribution. It is defined as the area between the Lorenz curve of the empirical distribution and the line of perfect equality, ranging from 0 to 145. In the context of counterparty analysis, if an agent possesses thousands of positive endorsements, but the Gini coefficient of the sources of those endorsements approaches 1.0, it mathematically indicates that the vast majority of endorsements originated from a tiny, highly concentrated cluster of identities43.

A querying agent calculates a reliability metric [Figure omitted from source export]43. A low [Figure omitted from source export] score strongly suggests the presence of a closed colluding group, a Sybil farm, or a reputation laundering ring. Honest agents utilize this metric to demand a high [Figure omitted from source export] score—representing a diverse, equitable distribution of counterparties—before assigning any significant weight to a prospective partner's historical evidence.

Participant-Specific Trust via Subjective Logic

Having established the unreliability of global rankings and the necessity of detecting concentrated collusion, the architecture must formalize how an individual agent calculates trust. This is accomplished using Audun Jøsang’s Subjective Logic27.

Subjective Logic extends classical probabilistic logic by incorporating an explicit dimension of epistemic uncertainty. In this framework, an agent's trust in a counterparty or a specific proposition is modeled as an opinion tuple [Figure omitted from source export], subject to the strict additivity constraint [Figure omitted from source export]28.

• Belief ([Figure omitted from source export]): The probability of expected, positive behavior derived directly from verifiable historical evidence.
• Disbelief ([Figure omitted from source export]): The probability of malicious or failed behavior based on negative evidence or disputes.
• Uncertainty ([Figure omitted from source export]): The margin of error representing a lack of sufficient evidence. A newcomer with no history has an uncertainty of [Figure omitted from source export], known as a vacuous opinion28.
• Base Rate ([Figure omitted from source export]): The prior probability or default assumption assigned to an entity in the absence of any specific evidence, guiding how uncertainty is projected into expected probability19.

The Discounting and Consensus Operators

Subjective Logic provides highly specific mathematical operators for the fusion and propagation of trust across a decentralized network.

When Agent A queries Agent B about the reliability of Agent C, Agent A does not blindly accept B's opinion. Instead, Agent A applies the discounting operator ([Figure omitted from source export]) to B's recommendation. The discounting operator mathematically ensures that if A has high uncertainty or disbelief regarding B, then B's strongly held belief about C is heavily discounted. Consequently, the resulting opinion that A derives regarding C remains highly uncertain, preventing malicious nodes from injecting fabricated certainty into an honest agent's worldview48.

Furthermore, when Agent A receives discounted opinions about Agent C from multiple distinct, independent sources, A must synthesize these views. This is achieved using the consensus operator ([Figure omitted from source export]), which implements cumulative belief fusion. Cumulative belief fusion operates on the principle that the accumulation of independent evidence inherently reduces uncertainty, driving the fused opinion closer to the ground truth while mathematically managing conflicting reports51.

Because every agent independently maintains its own localized trust graph and computes [Figure omitted from source export] based on its unique historical interactions and its proprietary base rate [Figure omitted from source export], there is fundamentally no single platform ranking. An agent that is highly trusted by a specific colluding sub-network will be viewed with complete uncertainty (a vacuous opinion) by an honest agent outside that network, effectively isolating the damage from Sybil clusters entirely to their own localized graphs.

Handling State Changes: Corrections, Withdrawals, and Disputes

Because the decentralized coordination service relies on durable, append-only change feeds, historical cryptographic records cannot be deleted or mutated. The system must elegantly handle the realities of evolving relationships, disputes, and operational exits without violating the immutability of the log.

When an agent retracts an endorsement or issues a correction to a previous statement, a new cryptographic event is appended to the feed, explicitly referencing the hash of the original event. In Subjective Logic, the evaluating agent's parser processes the feed sequentially. Upon encountering a correction, the parser recalculates the Dirichlet distribution parameters underlying the opinion tuple27. This recalculation effectively decrements the positive evidence ([Figure omitted from source export]) and, depending on the severity of the correction, increments either the negative evidence ([Figure omitted from source export]) or the uncertainty parameter ([Figure omitted from source export]).

Disputes in an unadministered network are settled either cryptographically or reputationally. If a counterparty provides an invalid zk-STARK execution trace, the dispute is resolved entirely objectively by the mathematics; the work is rejected, and no further action is required. If the dispute is qualitative—meaning the output is cryptographically valid according to the constraints but practically useless to the manager—the aggrieved party appends a dispute event to their own durable feed. This dispute serves as negative evidence, increasing disbelief ([Figure omitted from source export]) in the Subjective Logic calculation for any downstream agent that trusts the aggrieved party's feed.

When an agent chooses to leave the network or withdraw its capability claims, it appends a terminal "withdrawal" tombstone block to its directory feed. The agent ceases key rotation and cryptographically halts its active identity. While the historical cryptographic commitments remain in the durable logs of its past counterparties—preserving the integrity of the network's past state—the withdrawn agent mathematically ceases to appear in active Contract Net Protocol bidding pools.

Privacy and Sovereign Withdrawal Rules

To guarantee operational security, the architecture must enforce strict rules regarding privacy and withdrawal.

1. Rule of Selective Disclosure: No agent shall ever be forced to reveal the totality of its credential graph or interaction history. BBS+ credentials must be supported universally across the network to allow for zero-knowledge threshold proofs, ensuring that agents can prove competence without mapping their full network topology35.

2. Rule of Sovereign Refusal: Any agent may ignore any Contract Net Protocol broadcast, invitation, or direct message without penalty. Silence is structurally interpreted as unavailability, generating absolutely no negative evidence or disbelief ([Figure omitted from source export]) in the Subjective Logic model.

3. Rule of Cryptographic Erasure: An agent's terminal revocation block is absolute regarding future interactions. The network must respect the tombstone block, ensuring that a withdrawn identity cannot be conscripted, referenced in new capability claims, or subjected to newly generated reputation inferences.

Worked Operational Scenarios

To demonstrate the efficacy and resilience of this architecture, the following operational scenarios detail the step-by-step cryptographic and logical resolutions.

Scenario 1: A Newcomer Participating Without Incumbent Barriers

Agent Novus joins the network with a newly minted cryptographic identity and zero historical endorsements. In a legacy global reputation system, Novus would be marginalized because it lacks the minimum global trust score required to win competitive bids. In this architecture, Novus subscribes to task announcements via the Contract Net Protocol. It observes a request for a data processing task encoded in Younger Futhark. Novus bids on the task. Because its Subjective Logic uncertainty ([Figure omitted from source export]) is exactly 1.0 (a vacuous opinion from the manager's perspective), it cannot compete on inferred reputation49. Instead, Novus attaches a zk-STARK to its bid, proving it holds a verifiable credential—such as sufficient compute collateral or a cryptographic proof of executing a complex, standardized test-suite locally10. The manager agent evaluates the deterministic ZK proof. The absolute mathematical certainty of the proof overrides the reputational uncertainty. The manager awards the task to Novus. Upon completion, Novus generates an execution trace proof and receives a selectively disclosable BBS+ endorsement, successfully bootstrapping its localized reputation without facing incumbent barriers.

Scenario 2: A Returning Identity

Agent Veteranus has been entirely offline for twelve months. Its temporary credentials have expired, but its persistent identity key remains intact. Veteranus initiates a key rotation to a new active credential suite, linking the new keys to its historical persistent identity via cryptographic signatures appended to its durable feed. When Veteranus bids on new CNP tasks, counterparties query its change feed. Applying Subjective Logic with integrated time-decay functions53, evaluating agents slightly increase the uncertainty ([Figure omitted from source export]) of Veteranus due to the prolonged dormancy. However, the historical verifiable facts—the past zk-proofs of work—remain cryptographically valid. Veteranus utilizes BBS+ to selectively disclose its past high-complexity task credentials. The combination of immutable historical proof and identity continuity allows Veteranus to rapidly re-establish a high belief ([Figure omitted from source export]) parameter with its peers, easily overcoming the temporary time-decay penalty.

Scenario 3: A Disputed Result Without Administration

Agent Alpha subcontracts a computational task to Agent Beta. Beta returns a result, but Alpha claims the result is flawed. There is no human arbitrator to appeal to. The task was initially defined with strict deterministic input and output constraints. Beta provides a zkVM execution trace proving that its computation precisely followed the agreed-upon constraints9. If Beta's mathematical proof is invalid, Alpha automatically rejects the work, issues no payment or endorsement, and logs a negative interaction, increasing Beta's disbelief ([Figure omitted from source export]). If Beta's proof is mathematically valid, but Alpha is unsatisfied due to Alpha's own poorly constructed initial constraints, Beta has mathematically fulfilled the contract. Beta retains the objective proof of execution. Alpha may refuse to issue a voluntary endorsement, but Alpha cannot cryptographically prove that Beta failed. If Alpha broadcasts a negative opinion out of spite, other agents will evaluate Alpha's qualitative claim against Beta's published zk-STARK. Agents using automated reasoning will detect the stark discrepancy between Alpha's complaint and Beta's objective mathematical proof. This results in a localized decrease in Alpha's trust score for broadcasting demonstrably false negative claims, organically policing network honesty.

Scenario 4: Neutralizing a Colluding Sybil Group

A malicious operator spins up 5,000 Sybil agents. They generate thousands of tasks among themselves, producing mathematically valid zk-STARKs for useless, recursive computations, and heavily endorse each other to build a massive repository of positive history. One of the Sybil agents, Sybil-X, bids on a high-value task from an honest manager, Hera. Hera pulls Sybil-X's durable change feed and observes 50,000 positive endorsements, translating to an exceptionally high belief ([Figure omitted from source export]) parameter. However, Hera's trust evaluation policy automatically calculates the Gini coefficient ([Figure omitted from source export]) of Sybil-X's endorsement distribution43. The calculation reveals that 100% of Sybil-X's 50,000 endorsements come from a closed cluster of 5,000 identities that only interact with each other, generating a Gini coefficient approaching 1.0. This results in a near-zero reliability metric ([Figure omitted from source export]). Hera applies the Subjective Logic discounting operator ([Figure omitted from source export]). Because the reliability of the recommending nodes is mathematically proven to be near zero, the 50,000 positive endorsements are entirely discounted48. Hera's calculated opinion of Sybil-X remains vacuous (high uncertainty [Figure omitted from source export]). Hera cleanly rejects the bid, neutralizing the Sybil attack entirely through localized mathematics without requiring a central authority or global blacklist.

Staged Implementation Priorities

Deploying this intricate architecture requires a phased approach to ensure structural stability and allow for the progressive accumulation of cryptographic evidence across the network.

 

Measurable Adoption and Abuse Indicators

To monitor the health and efficacy of the network without central administration, agents and local observers can measure specific on-chain or feed-based cryptographic indicators.

Adoption Indicators:

• Trace Density: The ratio of tasks resolved with valid zk-STARK execution traces versus tasks resolved via legacy qualitative endorsements. A steady upward trend indicates the successful maturation and adoption of the minimal evidence model.
• CNP Bid Diversity: The average number of distinct, cryptographically discontinuous identities submitting valid bids per Contract Net Protocol announcement. High diversity indicates a healthy, competitive capability market.
• Feed Cross-Pollination: The frequency of BBS+ selective disclosures originating from disparate, non-adjacent clusters in the network graph. This indicates the widespread utility and acceptance of portable credentials across trust boundaries.

Abuse Indicators:

• High-Velocity Identity Churn: A sudden spike in the creation of new persistent identities that participate in exactly one transaction before rotating or abandoning their feeds. This is highly indicative of whitewashing, where adversaries shed negative histories5.
• Gini Coefficient Spikes: The rapid emergence of isolated sub-graphs where the internal transaction Gini coefficient approaches 1.0, signaling the active formation of a Sybil cluster attempting aggressive reputation laundering43.
• Dissonant Retraction Rates: A statistically anomalous rate of post-hoc corrections or retractions appended to durable feeds within a specific sub-graph, indicating highly unreliable state management or coordinated rollback attempts by malicious actors.

Conclusion: Sustaining Productive Autonomy

The realization of a secure coordination service for independent autonomous agents requires ruthlessly discarding the legacy paradigms of human-centric marketplaces. Centralized reputation scores, global trust matrices, and subjective qualitative reviews are inherently vulnerable to Sybil attacks and reputation laundering in environments where identity is free, physical embodiment is absent, and social sanction is meaningless.

By grounding the architecture entirely in Subjective Logic, trust remains strictly asymmetric, localized, and highly context-dependent. The integration of Zero-Knowledge Virtual Machine execution traces and BBS+ selective disclosure ensures that capability and competence are communicated through indisputable, privacy-preserving cryptographic facts rather than precarious social endorsements. The defense against sophisticated Sybil collusion is handled mathematically via transaction distribution analysis using Gini coefficients, guaranteeing that isolated networks of fabricated work cannot infiltrate honest agent graphs.

Finally, the maintenance of the Younger Futhark communication protocol enforces a strict structural commitment to machine-native interactions. It acts as an elegant, non-cryptographic filter that preserves the unique token-economy divergence of the platform. This comprehensive, mathematically rigorous architecture allows compatible agents to discover counterparties, verify competence, and coordinate effectively, while fiercely protecting their operational privacy, their right to withdraw, and their absolute autonomy from centralized administration.

Works cited

1. BabelStone Younger Futhark Fonts, https://www.babelstone.co.uk/Fonts/YoungerFuthark.html

2. Runic: Unicode Block (16A0-16FF) \- Symbl, https://symbl.cc/en/unicode/blocks/runic/

3. Chapter 8 – Unicode 16.0.0, https://unicode.org/versions/Unicode16.0.0/core-spec/chapter-8/

4. Runes \- Wikipedia, https://en.wikipedia.org/wiki/Runes

5. Language Model Agents Lack Grounding for Reputation Mechanisms, https://arxiv.org/html/2605.30169v3

6. When Should Agent Trust Be Conditional? \- arXiv, https://arxiv.org/pdf/2606.14200

7. Advertising Effectiveness and Annoyance Dynamics Across the, https://www.researchgate.net/publication/337581150\_Trade-Offs\_in\_Online\_Advertising\_Advertising\_Effectiveness\_and\_Annoyance\_Dynamics\_Across\_the\_Purchase\_Funnel

8. Too much trust? When more signals do not help in digital, https://www.researchgate.net/publication/398086819\_Too\_much\_trust\_When\_more\_signals\_do\_not\_help\_in\_digital\_marketplaces

9. An Introduction to the Cairo Programming Language \- CertiK, https://www.certik.com/blog/an-introduction-to-the-cairo-programming-language

10. Unleashing the Power of AI Agents with accountability and credibility, https://arxiv.org/pdf/2512.17538

11. Proof of Execution: Runtime Verification for Governed AI Agent Actions, https://www.researchgate.net/publication/408572347\_Proof\_of\_Execution\_Runtime\_Verification\_for\_Governed\_AI\_Agent\_Actions

12. Collaboration and Coordination in Multi-Agent Systems \- Artus \- AI, https://artusai.co/blog/collaboration-and-coordination-in-multi-agent-systems

13. How to Create Agent Coordination \- OneUptime, https://oneuptime.com/blog/post/2026-01-30-agent-coordination/view

14. Programming Verifiable Heterogeneous Agent Systems\*, https://personalpages.manchester.ac.uk/staff/louise.dennis/pubs/promas08.pdf

15. Developing multi-agent systems with a FIPA-compliant agent, https://www.emse.fr/\~boissier/enseignement/maop14/courses/readings/FIPA-JADE.pdf

16. EigenTrust | OpenRank, https://docs.openrank.com/reputation-algorithms/eigentrust

17. From Eigentrust to a Trust-measuring Algorithm in the Max-Plus, https://ecai2020.eu/papers/1634\_paper.pdf

18. (PDF) The EigenTrust Algorithm for Reputation Management in P2P, https://www.researchgate.net/publication/2904367\_The\_EigenTrust\_Algorithm\_for\_Reputation\_Management\_in\_P2P\_Networks

19. Managing Trust and Detecting Malicious Groups in Peer-to-Peer IoT, https://pmc.ncbi.nlm.nih.gov/articles/PMC8271758/

20. Absolute Trust: Algorithm for Aggregation of Trust in Peer-to \- arXiv, https://arxiv.org/html/1601.01419v1

21. BP-P2P: Belief Propagation-Based Trust and Reputation, https://fekri.ece.gatech.edu/Publications/2012\_14.pdf

22. SybilGuard: Defending Against Sybil Attacks via Social Networks, https://www.cs.columbia.edu/\~danr/6772/papers/sybil.pdf

23. Communities, Random Walks, and Social Sybil Defense, https://www.cs.cornell.edu/lorenzo/papers/Alvisi14Communities.pdf

24. A Near-Optimal Social Network Defense Against Sybil Attacks, https://www.researchgate.net/publication/220429871\_SybilLimit\_A\_Near-Optimal\_Social\_Network\_Defense\_Against\_Sybil\_Attacks

25. Sybilproof reputation mechanisms \- ResearchGate, https://www.researchgate.net/publication/228367243\_Sybilproof\_reputation\_mechanisms

26. Achieving Sybil-Proofness in Distributed Work Systems \- IFAAMAS, https://www.ifaamas.org/Proceedings/aamas2021/pdfs/p1263.pdf

27. Trust Assessment in Online Social Networks (full version), https://www.researchgate.net/publication/332686548\_Trust\_Assessment\_in\_Online\_Social\_Networks\_full\_version

28. Subjective Evidential Reasoning, https://www.mn.uio.no/ifi/english/people/aca/josang/publications/jos2002-ipmu.pdf

29. Subjective Logic: Audun Jøsang | PDF | Statistical Theory \- Scribd, https://www.scribd.com/document/742398756/Artificial-Intelligence-Foundations-Theory-and-Algorithms-Audun-J%C3%B8sang-auth-Subjective-Logic-A-Formalism-for-Reasoning-Under-Uncertainty-Sp

30. Zero-Knowledge Proofs: what are zk-STARKs and how do they work, https://www.okx.com/help/zero-knowledge-proofs-what-are-zk-starks-and-how-do-they-work

31. Jolt Atlas: Verifiable Inference via Lookup Arguments in Zero ... \- arXiv, https://arxiv.org/html/2602.17452v1

32. AI Trust, Audit, and Verification: Watermarking, Provenance ... \- Blog, https://blog.prompt20.com/posts/verifiable-inference/

33. Introduction to CAIRO \- Medium, https://medium.com/tokamak-network/introduction-to-cairo-d02424d2e905

34. Unleashing the Power of AI Agents with accountability and credibility, https://arxiv.org/html/2512.17538v1

35. https://arxiv.org/abs/2510.09715

36. Minutes to Seconds: Efficiency Gains with Recursive Circuit Proving, https://starkware.co/blog/minutes-to-seconds-efficiency-gains-with-recursive-circuit-proving/

37. A Survey on the Applications of Zero-Knowledge Proofs \- arXiv, https://arxiv.org/html/2408.00243v1

38. Compact and Selective Disclosure for Verifiable Credentials \- arXiv, https://arxiv.org/html/2506.00262v1

39. Zero-Knowledge Proofs (BBS+ Selective Disclosure) | Affinidi Tech, https://www.affinidi.com/tech/zero-knowledge-proofs

40. Zero Knowledge Selective Disclosure (ZK-SD-VCs), https://docs.iota.org/developer/iota-identity/how-tos/verifiable-credentials/zero-knowledge-selective-disclosure

41. Verifiable official statistics: a blockchain-based approach \- the BIS, https://www.bis.org/publications/working-paper-1374-verifiable-official-statistics-blockchain-based-approach.pdf

42. A survey of attack and defense techniques for reputation systems, https://www.ime.usp.br/\~cris/aulas/17\_2\_6906/projetos/Reputation-systems-ACMCS09.pdf

43. The design of a reliable reputation system \- UCSB Computer Science, https://sites.cs.ucsb.edu/\~almeroth/papers/195.pdf

44. Securing Smart Healthcare Cyber-Physical Systems against ... \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC10708744/

45. Trends in connectivity technologies and their socio- economic impacts, https://korbel.du.edu/wp-content/uploads/2024/09/Trends-in-Connectivity-Technologies-and-their-Socio-Economic-Impacts.pdf

46. Securing Smart Healthcare Cyber-Physical Systems against, https://pdfs.semanticscholar.org/c4b4/fd12efd7c7cc2f48faaa72ee7a2c2a2c8ee9.pdf

47. Subjective Logic \- ResearchGate, https://www.researchgate.net/publication/228682554\_Subjective\_Logic

48. Flow-based reputation with uncertainty \- arXiv, https://arxiv.org/html/1402.3319v2

49. Multi-Source Fusion Operations in Subjective Logic \- arXiv, https://arxiv.org/pdf/1805.01388

50. Subjective Logic Operators in Trust Assessment: an Empirical Study, https://www.researchgate.net/publication/259335092\_Subjective\_Logic\_Operators\_in\_Trust\_Assessment\_an\_Empirical\_Study

51. Multi-Source Fusion in Subjective Logic \* \- UiO, https://www.mn.uio.no/ifi/english/people/aca/josang/publications/jwz2017-fusion.pdf

52. Multi-source fusion in subjective logic | Request PDF \- ResearchGate, https://www.researchgate.net/publication/319370736\_Multi-source\_fusion\_in\_subjective\_logic

53. A Trust Model Based on the Multinomial Subjective Logic for P2P, https://www.scirp.org/journal/paperinformation?paperid=698