Published October 2, 2026 | Version 1.0

AI Trust Graph: A Graph-Driven, Evidence-Based Methodology for AI Assurance

Description

Whitepaper version 1.0 · methodology bundle 1.0-rc.4 (public-release candidate).

AI Trust Graph (ATG) is an open, vendor-neutral and product-independent methodology for AI assurance across connected AI systems, developed as independent research by Siva Sethumadhavan.

ATG represents an AI-enabled environment as a directed, labelled multigraph and reasons along one chain: Objects → Relationships → Conditions → Paths → Authority and Influence → Consequence → Controls → Evidence → Decision. It is designed to help assess how trust, delegated authority, access, data flows, tools, agents, providers and control breakpoints combine across connected system architectures.

The methodology is organised across six assurance domains and 72 canonical controls. These are supported by a cumulative, evidence-gated and non-compensating M1–M5 maturity model, an E0–E5 evidence model, explicit uncertainty handling through the UNKNOWN state, a 13-phase assessment lifecycle, path-oriented analysis, reporting requirements and governance.

A central principle is that evidence strength and control effectiveness are separate. Strong evidence may confirm an adverse state, while weak evidence cannot justify a strong assurance conclusion. ATG therefore preserves uncertainty instead of silently converting absent, insufficient or materially conflicting evidence into zero, safe, passed, failed, low-risk, Not Applicable or Not Tested states.

The whitepaper also explains the Path Exposure Index (PEI), which may be calculated for triage only, on determinate eligible active paths. PEI is not a probability, an expected-loss model, a certification score, an overall trust score or proof of exploitability.

This publication is a non-normative whitepaper that introduces and explains the methodology. The governed methodology artifacts in the public GitHub repository remain the canonical source for definitions, controls, evidence grades, maturity rules, scoring, assessment procedures, ontology, reporting and governance.

The methodology remains subject to independent methodology/architecture review, independent AI-security review, an inter-assessor reproducibility study, field calibration and legal approval of the licence/trademark position. An author-performed PEI sensitivity analysis has been completed; it is not an independent review.

Supplemented by : Canonical methodology repository (bundle 1.0-rc.4): https://github.com/Sivas1187/Ai-trust-graph

Files

AI-Trust-Graph-Whitepaper-v1.0.pdf

Files (678.8 kB)

Name Size
md5:6e8c075c3489cbf50169f8fe8fcce66c
678.8 kB Preview Download

Additional details

Additional titles

Subtitle
Reasoning about trust, authority, paths, controls, evidence, and accountable decisions across connected AI systems

Related works

Is documented by
Other: https://aitrustgraph.org (URL)
Is supplemented by
Other: https://github.com/Sivas1187/Ai-trust-graph (URL)