[Paper Review] Local and Global Trust Based on the Concept of Promises

J.A. Bergstra, Mark Burgess|ArXiv.org|Dec 23, 2009

Logic, Reasoning, and Knowledge21 references13 citations

TL;DR

This paper proposes a trust model based on promises—autonomous commitments between agents—where local trust is the expectation that a promise will be kept. It introduces a global trust metric via eigenvector centrality in a promise graph, showing trust is inherently heuristic and decouples reliability from behavior types, offering a formal foundation for models like the Web of Trust and trusted third parties.

ABSTRACT

We use the notion of a promise to define local trust between agents possessing autonomous decision-making. An agent is trustworthy if it is expected that it will keep a promise. This definition satisfies most commonplace meanings of trust. Reputation is then an estimation of this expectation value that is passed on from agent to agent. Our definition distinguishes types of trust, for different behaviours, and decouples the concept of agent reliability from the behaviour on which the judgement is based. We show, however, that trust is fundamentally heuristic, as it provides insufficient information for agents to make a rational judgement. A global trustworthiness, or community trust can be defined by a proportional, self-consistent voting process, as a weighted eigenvector-centrality function of the promise theoretical graph.

Motivation & Objective

• To resolve ambiguity in trust by grounding it in the concept of promises, which are autonomous, behavior-specific commitments.
• To decouple agent reliability from the specific behaviors being judged, enabling fine-grained trust assessment.
• To formalize global or community trust as a self-consistent, proportional voting process over a promise graph.
• To show that trust is fundamentally heuristic and insufficient for rational decision-making due to incomplete information.
• To provide a unifying framework that explains existing models like the Web of Trust and trusted third parties as special cases.

Proposed method

• Defining trust as the expectation value that a given promise will be kept, with promises being autonomous declarations of intended behavior.
• Modeling trust relationships as a directed graph where nodes are agents and edges represent promises, forming a promise-theoretical graph.
• Deriving global trust as a weighted eigenvector-centrality function of the promise graph, ensuring self-consistency and proportionality.
• Using probabilistic rules for combining multiple promises, analogous to fault tree analysis, to model bundles of commitments.
• Distinguishing reputation as a path-dependent, epidemic-like propagation of trust estimates across the network.
• Applying the model to analyze and formalize existing architectures such as the PGP Web of Trust and trusted third-party systems.

Experimental results

Research questions

• RQ1 How can trust be formally defined in a way that respects agent autonomy and avoids ambiguity in common usage?
• RQ2 What is the relationship between local trust (individual expectation) and global trust (community-wide reputation) in a network of autonomous agents?
• RQ3 Can a self-consistent, scalable measure of global trust be derived from promise relationships without relying on centralized authorities?
• RQ4 Why is trust inherently heuristic, and what limitations does this impose on rational decision-making in trust-based systems?
• RQ5 How do existing trust models like the Web of Trust and trusted third parties emerge as special cases of the proposed promise-based framework?

Key findings

• Trust is best defined as the expectation that a promise—an autonomous commitment—will be kept, providing a clear, behavior-specific, and policy-sensitive foundation.
• Global trust can be computed as a weighted eigenvector-centrality of the promise graph, ensuring proportionality and self-consistency across the network.
• Reputation is not inherently tied to trust but emerges as a path-dependent, epidemic-like propagation of trust estimates, making it sensitive to network structure.
• The model shows that trust is fundamentally heuristic because insufficient information exists to make fully rational judgments, making all trust valuations policy-dependent and ad hoc.
• The PGP Web of Trust and trusted third-party models are shown to be special cases of the proposed framework, though they lack full visibility and consistent trust measurement.
• The framework enables direct measurement of host trustworthiness in systems like cfengine by observing whether agents keep security-related promises, validating the model in practice.