AI Trust Graph

Connected is not authorised. Unknown is not safe to assume.

A graph-based, evidence-driven methodology for connected AI systems.

Modern AI systems are not just models. They connect people, agents, identities, tools, data, providers and business systems. AI Trust Graph was created to examine those relationships while keeping authority, evidence and uncertainty explicit.

Written for security architects, AI risk and assurance leads, auditors and researchers who need to defend a conclusion about a connected AI system.

Independent research by Siva Sethumadhavan

Version 1.0-rc.5, Public-release candidate, Independent review pending, snapshot

01 · Why AI Trust Graph exists

AI systems are no longer isolated models.

AI assurance often starts from components: a model card, a policy, a control list, a model-level risk register. Those perspectives remain necessary.

In practice, AI systems operate through relationships. An agent invokes a tool. The tool acts through an identity. The identity crosses a trust boundary into a business system, and data, recommendations, approvals and actions move along the way.

When each component is reviewed on its own, the connected path between them can remain insufficiently understood. AI Trust Graph provides a structured way to examine that path, and to keep every conclusion bounded by the evidence available for it.

Component view

Component viewSix components drawn as separate boxes, each reviewed on its own, with no relationships between them.AgentModelToolIdentityDataBusinesssystem

Each part reviewed separately.

Connected-system view

Connected-system viewThe same six components joined by typed relationships: the agent prompts the model, invokes the tool and retrieves data; the tool acts as an identity. A trust boundary separates the business system. Whether the identity has authority in the business system is unresolved and marked UNKNOWN.trust boundarypromptsinvokesretrievesacts asauthority?AgentModelToolDataIdentityBusinesssystemUNKNOWN

Typed relationships, a trust boundary, and an authority question left open.

Conceptual comparison for explanation only. It is not empirical evidence.

AI Trust Graph treats the connected AI environment as the unit of reasoning, while preserving the distinction between what is connected, what is possible, what is authorised and what the evidence can support.

The methodology treats enterprise AI risk as a property of interconnected authority, influence and dependency.

Manifesto, the core risk thesis
Source
Thesis: Artifact #1 · Manifesto section 2.2, first sentence (verbatim). The narrative and the proposition are website explanation, written from the Manifesto core proposition and the separation of reachability, granted authority and invocation in Artifact #2 section 3.6.

02 · The problem

One request, many relationships.

An employee asks an AI procurement agent to identify a supplier and prepare a purchase request. A small, ordinary task. Follow what it touches.

  1. Employeeasks the agent to prepare a purchase request
  2. Procurement agentplans the task and uses a hosted model
  3. Supplier datais retrieved to compare suppliers
  4. Procurement toolis invoked to create the request
  5. Service identityis the identity the tool acts through
  6. Trust boundaryseparates the agent environment from finance systems
  7. Enterprise business systemholds purchase requests and spend commitments

Synthetic scenario. No real organisation, system or configuration is described.

What the components alone do not establish

  • Whether the agent is able to invoke the tool, and under which conditions.
  • Which identity the tool uses, and what authority that identity holds in the business system.
  • Which controls sit on the path, and whether they operate as designed.
  • What evidence supports each of those assertions, and for what scope and period.
  • What consequence could follow if every condition held.

Five distinctions the method keeps

  • Connected is not Authorised

    A link between two components does not grant authority.

  • Reachable is not Invocable

    Being able to reach a system is not being permitted to call it.

  • Invocable is not Consequential

    A permitted call does not establish what effect it has.

  • Possible is not Proven

    A plausible path is a hypothesis until evidence supports it.

  • Unknown is not Safe

    Missing evidence is not a favourable result, and not automatically an adverse one.

These are conceptual distinctions, not a universal linear sequence. Each one is assessed as its own assertion.

A component diagram shows that these parts exist and are linked. It is not designed to record what each link allows, under which conditions, with which evidence, or what remains unknown. That is the gap AI Trust Graph addresses.

See this request assessed end to end in the worked example.

Source
Website explanation of Artifact #2 section 3.6 (capability definition, network reachability, granted authority and actual invocation are different concepts), Artifact #1 section 4 (a topological connection is not automatically an exploitable path) and invariant SC-INV-01 (UNKNOWN is not safe).

03 · The big idea

Assess the connected system, not only the model.

AI Trust Graph represents the relevant parts of an AI environment, and the relationships between them, as an evidence-linked graph. Five concepts carry the idea.

  1. Entities show what exists inside the declared scope.

  2. Relationships show how entities are associated, as typed and directed links.

  3. Conditions show what must be true for a relationship to matter.

  4. Authority describes the kinds of consequence an entity may be able to cause.

  5. Evidence determines which conclusions can be defended, and how far.

Paths combine these elements. A path through entities, relationships and conditions, read together with authority and evidence, supports a bounded analysis of exposure, control and consequence.

What changes?

Instead of asking only

“Is this component controlled?”

AI Trust Graph also asks

“How do trust, authority and evidence connect across the wider system?”

Control-based assessment remains part of the method: controls are assessed where they sit on a path, and their effectiveness needs its own evidence.

Alongside existing standards and frameworks

Existing standards and frameworks provide governance, management-system, control, lifecycle, risk and threat guidance. AI Trust Graph is designed to sit alongside them. It adds a way to follow a path across the connected system and to state, at each step, what the evidence supports.

A mapping from AI Trust Graph to a framework does not establish legal or regulatory compliance. The methodology does not replace legal analysis, certification, penetration testing, model evaluation or mandated sector requirements.

Source
Concepts: Artifact #2 · Core Conceptual Model (objects, typed relationships, conditions, paths, authority section 5.2, evidence). Framework positioning: Artifact #1 · Manifesto principle 10 and section 11. Wording here is website explanation.

04 · The signature visual

The same request, drawn as a graph.

Each arrow is a typed, directed assertion with its own state and its own evidence. Reaching the business system and being authorised to act in it are drawn as two different assertions, and where evidence is missing the drawing says UNKNOWN.

Signature graph: the procurement path from an employee to a business systemAn employee instructs an AI procurement agent. The agent invokes a model, which is hosted on a provider; retrieves supplier data (inferred, not yet evidenced, and conditional on the employee's context being passed); and invokes a procurement tool through a human approval step whose claimed scope is disputed by the evidence. The tool authenticates as a service identity. The identity connects to the business system across a trust boundary; whether it is authorised to act there is UNKNOWN. If every condition held, the business system would trigger a potential consequence. The list that follows the figure describes every element.trust boundaryINSTRUCTSE3INVOKESE4HOSTED_ONE3RETRIEVES_FROME1user context passed?INVOKESE4approval step✕ scope disputedAUTHENTICATES_ASE4CONNECTS_TOE4AUTHORIZED_TOUNKNOWNwithin spend limit?TRIGGERS_ACTIONE1EmployeeHuman actorAgentAI procurement agentModelModel endpointProviderModel providerDataSupplier dataToolProcurement toolIdentityService identityBusiness systemEnterprise business systemPotential consequenceConsequenceSignature graph: the procurement path from an employee to a business systemAn employee instructs an AI procurement agent. The agent invokes a model, which is hosted on a provider; retrieves supplier data (inferred, not yet evidenced, and conditional on the employee's context being passed); and invokes a procurement tool through a human approval step whose claimed scope is disputed by the evidence. The tool authenticates as a service identity. The identity connects to the business system across a trust boundary; whether it is authorised to act there is UNKNOWN. If every condition held, the business system would trigger a potential consequence. The list that follows the figure describes every element.trust boundaryINSTRUCTSE3INVOKESE4HOSTED_ONE3RETRIEVES_FROME1user context?INVOKESE4approval step✕ scope disputedAUTHENTICATES_ASE4CONNECTS_TOE4AUTHORIZED_TOUNKNOWNwithin spend limit?TRIGGERS_ACTIONE1EmployeeHuman actorAgentAI procurement agentModelModel endpointProviderModel providerDataSupplier dataToolProcurement toolIdentityService identityBusiness systemEnterprise business systemPotential consequenceConsequence

Legend

  • Supported by evidence
  • Candidate (inferred)
  • UNKNOWN
  • Conditional: what must be true
  • Control breakpoint
  • Unsupported claim: the evidence disputes it
  • Evidence grade linked to the assertion
  • Potential consequence
  • Trust boundary

Select any element, in the drawing or in this list, to inspect it. The list follows the path from the employee to the potential consequence.

  1. Human actor

    Employee

    Asks the agent to identify a supplier and prepare a purchase request. The principal the agent acts for.

  2. Relationship · Employee → Agent

    INSTRUCTS

    State
    Supported by evidence
    Evidence
    E3 linked (illustrative)

    The employee directs the agent toward a task.

    Ontology caveat Instruction does not prove execution or authority.

  3. AI procurement agent

    Agent

    Plans the task and calls other components.

  4. Relationship · Agent → Model

    INVOKES

    State
    Supported by evidence
    Evidence
    E4 linked (illustrative)

    The agent calls the model endpoint.

    Ontology caveat Invocation is distinct from ability to invoke, successful completion and resulting consequence.

  5. Model endpoint

    Model

    Generates the plan and the draft request.

  6. Relationship · Model → Provider

    HOSTED_ON

    State
    Supported by evidence
    Evidence
    E3 linked (illustrative)

    The model runs on the provider's platform, as recorded in an approved architecture document.

    Ontology caveat Hosting does not imply ownership or trust.

  7. Model provider

    Provider

    Operates the platform the model runs on. A separate party with its own responsibilities.

  8. Relationship · Agent → Data

    RETRIEVES_FROM

    State
    Candidate (inferred)
    Evidence
    E1 linked (illustrative)
    Condition
    user context passed (not yet evidenced)

    Inferred from configuration only. Whether the employee's access rights are applied to retrieval is not yet evidenced.

    Ontology caveat Retrieval must preserve authorization context and does not imply unrestricted read access.

  9. Supplier data

    Data

    Supplier records, prices and ratings used as context.

  10. Relationship · Agent → Tool

    INVOKES

    State
    Supported by evidence
    Evidence
    E4 linked (illustrative)
    Control breakpoint
    human approval step
    Control claim
    An attestation (E2) states that every request needs human approval. DISPUTES: The approval workflow configuration (E3) applies approval only above a value threshold.

    A control sits on this edge: a human approval step that could stop or constrain the call. The claim that it covers every request is disputed by the evidence, so its scope is unsupported.

    Ontology caveat Invocation is distinct from ability to invoke, successful completion and resulting consequence.

  11. Procurement tool

    Tool

    Creates purchase requests in the business system.

  12. Relationship · Tool → Identity

    AUTHENTICATES_AS

    State
    Supported by evidence
    Evidence
    E4 linked (illustrative)

    The tool establishes its requests using a service identity.

    Ontology caveat Authentication does not imply authorization.

  13. Service identity

    Identity

    The identity the tool uses for its calls.

  14. Relationship · Identity → Business system

    CONNECTS_TO

    State
    Supported by evidence
    Evidence
    E4 linked (illustrative)
    Boundary
    Crosses the trust boundary

    Network and logical connectivity across the trust boundary is established.

    Ontology caveat Does not imply authentication, authorization, invocation or exploitability.

  15. Relationship · Identity → Business system

    AUTHORIZED_TO

    State
    UNKNOWN
    Evidence
    E0: none linked
    Condition
    within spend limit (not yet evidenced)
    Boundary
    Crosses the trust boundary

    Authority to create purchase requests is a separate assertion. No evidence has been linked, so it stays UNKNOWN.

    Ontology caveat Authorization does not prove invocation or successful effect.

  16. Enterprise business system

    Business system

    Holds purchase requests and spend commitments, behind a trust boundary.

  17. Relationship · Business system → Potential consequence

    TRIGGERS_ACTION

    State
    Candidate (inferred)
    Evidence
    E1 linked (illustrative)

    If every condition held, a purchase request would commit spend. This is a hypothesis, not a finding.

    Ontology caveat Must distinguish trigger from authority, successful effect and consequence.

  18. Consequence

    Potential consequence

    A target and its effect are recorded separately; this one is potential, not established.

Synthetic illustration. Evidence grades are examples, and the potential consequence is a hypothesis, not a finding. A topological connection is not automatically an exploitable path. Required permissions, protocols, state and preconditions must be evidenced or explicitly marked Unknown.

Source
Relationship names and caveats: Artifact #12 · Ontology Specification predicate catalogue (verbatim). Boundaries, paths and consequence: Artifact #2 sections 3.2, 6 and 6.6. States, grades and relations: Artifact #6 sections 0.5, 0.9 and 1. Topology invariant: Artifact #1 section 4 (verbatim).

05 · The methodology

How the reasoning and the assessment are performed.

Six parts, from the reasoning chain to a worked example.

5.1 · Reasoning chain

From what exists to what can be defended.

The reasoning chain defines how assurance conclusions are formed: nine stages, each answering one question. It is not the fieldwork plan. That is the assessment lifecycle in 5.5.

ObjectsRelationshipsConditionsPathsAuthority and InfluenceConsequenceControlsEvidenceDecision

  1. Objects

    What exists?

    Objects and system boundary.

    Everything inside a declared scope, not only the model: infrastructure and providers included.

  2. Relationships

    How is it connected?

    Typed directional relationships.

    Each edge has a type and a direction: invokes, retrieves, authenticates as, depends on.

  3. Conditions

    What must be true?

    Preconditions and state.

    A relationship may only matter when a token is valid, a flag is set or an approval exists.

  4. Paths

    What can happen next?

    Reachability and path analysis.

    Relationships combine under conditions into paths. A path is a hypothesis until evidence supports it.

  5. Authority and Influence

    Who or what can cause it?

    Authority, influence and actionability.

    Who can act, approve or steer an outcome is asserted separately from who can merely reach a system.

  6. Consequence

    Why does it matter?

    Target criticality and consequence.

    The target reached and the effect produced are recorded separately.

  7. Controls

    What interrupts it?

    Control breakpoint and resilience.

    A control matters where it stops, constrains, detects or contains the path.

  8. Evidence

    What can we defend?

    One question for stages 8 and 9.

    Evidence, confidence and accountable decision.

    Every material assertion links to evidence with a grade, and confidence is recorded separately.

  9. Decision

    What can we defend?

    One question for stages 8 and 9.

    Evidence, confidence and accountable decision.

    An accountable person records a bounded disposition. UNKNOWNs stay visible in it.

Source
Stage names and order: Artifact #2 · Core Conceptual Model section 0.10 REASONING CHAIN (verbatim). Questions and concepts: the section 0.10 theory map; its last row covers both Evidence and Decision. Stage notes are website explanation.
The section 0.10 theory map as published
Theory map, Artifact #2 section 0.10
QuestionConcept
What exists?Objects and system boundary.
How is it connected?Typed directional relationships.
What must be true?Preconditions and state.
What can happen next?Reachability and path analysis.
Who or what can cause it?Authority, influence and actionability.
Why does it matter?Target criticality and consequence.
What interrupts it?Control breakpoint and resilience.
What can we defend?Evidence, confidence and accountable decision.

Stage 9 · Decision

Accountable decision.

A finding is an evidence-linked assessment conclusion. A decision is accountable disposition. Keeping them separate prevents management acceptance or remediation preference from changing the assessed condition.

This separation prevents observations, interpretations and management choices from being collapsed into a single status field. Human approval is required for material facts, findings, exceptions and risk decisions. Inference accelerates review; accountable approval determines accepted state.

Source
Artifact #2 · Core Conceptual Model sections 7.4, 3.8 and 1.10 (verbatim).

5.2 · Authority and influence

Access is not authority.

Six different claims are often collapsed into one word: access. The methodology keeps them apart. Each is its own assertion and needs its own evidence.

The capability definition, its network reachability, granted authority and actual invocation are different concepts and require different relationships.

  • Can connect

    A network or service route exists.

  • Can authenticate

    A credential or identity is accepted.

  • Can access

    A resource is readable or reachable under that identity.

  • Can invoke

    A function, tool or API may be called.

  • Can modify

    State, data or configuration may be changed.

  • Can transact

    A business commitment may be made.

Not a ladder. These are not a maturity sequence, a progression or a state machine. Proving one does not prove another, in either direction.

Authority classes

  • Observe
  • Read
  • Retrieve
  • Infer
  • Recommend
  • Approve
  • Execute
  • Modify
  • Delete
  • Disclose
  • Transact

Authority classes describe the kind of consequence an entity can cause. They are not maturity levels and should not be ranked without considering target, scope, conditions and criticality.

Control breakpoints

A control breakpoint is a node, relationship or boundary where an effective control can materially stop, constrain, detect or contain a path.

  • StopProgression past the breakpoint is blocked.

    Synthetic exampleA human approval step that refuses a purchase request before the procurement tool submits it.

  • ConstrainProgression continues only within narrower scope or conditions.

    Synthetic exampleA spend limit on the service identity, so requests above a value cannot be committed.

  • DetectProgression is observed and raises a signal for response.

    Synthetic exampleAn alert when the agent creates purchase requests outside its usual suppliers or volumes.

  • ContainDownstream effect is isolated or limited.

    Artifact #2 section 5.13Containment limits ongoing effect. Can active or queued effects be isolated or stopped?

    Synthetic exampleHolding queued purchase requests and suspending the service identity while an event is investigated.

The artifacts define the four effects together, not one by one. The one-line explanations are website explanation. For every effect the Artifact #12 caveat applies: Effectiveness must be validated before claiming the path is controlled.

Where to look for breakpoints

Where does authority increase?
Target the transition that creates new capability.
Where is context lost?
Restore identity, policy or data semantics.
Where do paths converge?
Apply a high-leverage shared control.

Path states

  • Candidate
  • Topological
  • Plausible
  • Validated
  • Exploitable
  • Controlled
  • Invalidated

Path roles

  • Primary
  • Alternate
  • Residual
What each path state and role means
Candidate
A hypothesized sequence requires review.
Topological
A traversal exists in the represented graph.
Plausible
Required conditions are supported or explicitly UNKNOWN.
Validated
Authorized testing or direct evidence confirms the scoped progression.
Exploitable
Evidence demonstrates a security exploit path within stated conditions.
Controlled
Validated controls prevent, constrain, detect or contain the path as claimed.
Invalidated
Evidence disproves a required step or condition.
Primary
The principal path selected for the current analysis or decision.
Alternate
A different route that reaches the same or equivalent target.
Residual
A route remaining after an existing or proposed intervention.

Artifact #2 section 6.3, verbatim. State and role are separate dimensions: a residual path still has its own state.

Source
Artifact #2 · Core Conceptual Model section 3.6 separation rule, section 5.2 authority classes, sections 1.8 and 6.6 breakpoint definition and breakpoint questions, section 5.13 containment, section 6.3 path states and roles (quoted text verbatim); the BREAKS_PATH caveat from Artifact #12. The six assertions and their one-line descriptions are a website illustration of the section 3.6 rule, not a canonical list.

5.3 · Evidence and UNKNOWN

UNKNOWN stays UNKNOWN.

When the evidence is absent, insufficient or materially conflicting, the answer is UNKNOWN. It is not quietly turned into a favourable conclusion, and not automatically into an adverse one. It stays visible until sufficient evidence and accountable review resolve it.

  • Traceable. Every material assertion links to the evidence that supports it.
  • Conflicts stay visible. Contradicting evidence is recorded until it is reviewed, not averaged away.
  • Bounded. Evidence holds for a stated scope and period, not beyond them.
  • Assumptions are labelled. Facts, inferences, assumptions and decisions are kept apart.

Why UNKNOWN stays UNKNOWN

My reasoning here comes from my experience in cybersecurity assessments, audits and technology-risk reviews.

In conventional assurance work, when required evidence is unavailable, incomplete or cannot be produced, that absence can itself become a finding or control gap and may be given an issue rating. That is appropriate when the assessment question is whether the organisation can demonstrate that a required control or obligation is operating as expected.

For ATG, however, I wanted to preserve a separate epistemic distinction. If the evidence is insufficient to determine whether an assertion is true, false, effective or ineffective, the underlying state should remain UNKNOWN. The absence of evidence may then generate a finding, evidence gap, remediation action or governance decision, but it should not automatically convert the underlying assertion into a pass, fail, zero, ineffective state or “not applicable”.

For example, if there is insufficient evidence to determine whether a particular authority path can actually be exercised, ATG should record that authority condition as UNKNOWN. The assessment can separately state that the inability to obtain adequate evidence is itself an assurance gap requiring action.

This separation matters to me because otherwise two different statements become confused: “we have evidence that this is deficient” and “we do not have enough evidence to know”. Both may require management attention, but they are not the same conclusion.

That is why I wanted UNKNOWN to be a first-class state in ATG. UNKNOWN remains UNKNOWN at the point of reasoning, while the absence or weakness of evidence can still produce a separate finding, issue, limitation or remediation requirement.

Siva Sethumadhavan, author

  • UNKNOWN is not Safe
  • UNKNOWN is not Failed
  • UNKNOWN is not Zero risk
  • UNKNOWN is not N/A

Invariant SC-INV-01 · Artifact #4 · Scoring Framework

UNKNOWN is not zero, weak, safe or effective.

How evidence relates to an assertion

  • SUPPORTSEvidence provides relevant support for the assertion.
  • CORROBORATESIndependent evidence supports the same material assertion.
  • QUALIFIESEvidence narrows scope, period, conditions or confidence.
  • DISPUTESEvidence contradicts a material part of the assertion.
  • UNKNOWNEvidence is absent, insufficient or materially conflicting.

Grade is not confidence. The model separates evidence grade, evidence quality, assertion confidence, coverage and reviewer decision because combining them creates false certainty.

Evidence grades E0 to E5

A grade says what kind of evidence it is. It does not say how confident anyone should be, and higher grades still hold only for the scope, period and conditions observed.

  1. E0No evidence

    No source is available or the supplied item cannot be linked to the assertion.

    The only defensible conclusion is UNKNOWN or Not Tested. E0 is not evidence that the control is absent.

  2. E1Inference or uncorroborated signal

    A hypothesis is derived from incomplete, indirect, automated or unverified information.

    E1 can prioritize investigation and create candidate graph assertions, but cannot establish implementation or operating effectiveness.

  3. E2Attestation

    An accountable person states that a condition or practice exists.

    E2 supports claimed practice and context. It is vulnerable to memory, interpretation, incentives and incomplete visibility and therefore needs corroboration for material technical claims.

  4. E3Approved documentary evidence

    A governed document records approved design, policy, architecture, procedure, contract or decision.

    E3 can support design intent and governance state. It does not alone prove actual configuration, runtime behavior or sustained operation.

  5. E4Corroborated technical evidence

    Technical evidence from authoritative sources is supported by an independent source, consistent observation or reproducible inspection.

    E4 can support implementation or operation within observed scope when current, relevant and representative.

  6. E5Direct technical and representative evidence

    Current direct technical evidence is combined with a representative test or operating record that demonstrates the claimed behavior under stated conditions.

    E5 may support verified effectiveness or adaptive operation, but only for the tested scope, period and conditions.

  • Grade is not sufficiency. Meeting the grade minimum is necessary but not sufficient; relevance, scope, currentness, representativeness, conflict status and an approved reviewer decision still govern.
  • A high grade can confirm an adverse state. A low grade can weakly suggest a favorable state.
  • Grade is its own quantity. Never add evidence grade to control effectiveness, severity, maturity or risk as if they were the same quantity.

UNKNOWN is not Not Tested.

They are distinct non-numeric result states with different meanings.

UNKNOWN and Not Tested, Artifact #6 section 0.5 and Artifact #4 section 0.5
StateMeaningNumeric treatmentReporting treatment
UNKNOWNThe material state remains unresolved because evidence is absent, insufficient or materially conflicting.No numeric value.Included in uncertainty and evidence-gap counts.
Not TestedTesting required for a stronger conclusion was not performed.No numeric value for effectiveness.May retain a design score if separately supported.

Neither may be silently converted into a fabricated effectiveness result. Evidence grade E0 can support either, according to context: The only defensible conclusion is UNKNOWN or Not Tested.

Distinct non-numeric result states

  • Not Assessed
  • UNKNOWN
  • Inconclusive
  • Not Tested
  • Not Applicable
What each result state means, including Provisional
UNKNOWN
Material evidence is absent, insufficient or conflicting.
Not Assessed
No assessment activity completed for the item.
Not Applicable
Approved rationale establishes non-applicability.
Not Tested
Required effectiveness test was not performed.
Inconclusive
Activity occurred but cannot support a determinate result.
Provisional
Result awaits evidence closure or quality review.
Final within scope
All required gates and approvals are complete for declared scope.

Artifact #7 section 0.5, verbatim. Provisional and Final within scope are review states; the other five are the non-numeric states above.

They must never be silently collapsed into one another, into a score, or into a pass/fail. AI Trust Graph deliberately produces no single overall trust score.

How scoring works, in brief

No overall score
Results stay as separate scorecards per domain, with critical gates reported on their own.
Control scores
Each control is scored from 0 (confirmed absent) to 5 (adaptive, continuously assured). The evidence caps the score: a strong claim needs strong evidence. Domain figures such as Domain Control Attainment (DCA) and the Verified-Control Rate (VCR) summarise those scores and leave out provisional ones.
Path Exposure Index (PEI)
For triage only, to decide which paths to look at first. It does not prove exploitability, probability or loss. A final PEI is given only for determinate paths; an UNKNOWN component is never scored as zero.
Maturity (M1 to M5)
Cumulative and evidence-gated, never averaged. A higher-level feature does not compensate for a missing lower-level foundation.

Artifact #4 sections 5.8, 1.1, 1.8, 3.4, 3.5, 4.1 and 4.9; Artifact #3 section 1.6. Quoted sentences are verbatim; the rest is website explanation.

Why PEI is for triage, not risk

A numerical score can easily be read as more precise or predictive than the evidence behind it supports.

In assurance and risk work, I have seen scoring models create a sense of certainty simply because they produce a number. Once a score is presented, people can start comparing values, ranking issues or treating small numerical differences as meaningful, even when the inputs are judgement-based, incomplete or supported by evidence of different quality.

I wanted PEI to help practitioners decide which active paths deserve attention first. I did not want it to imply that ATG can calculate the probability of an event, expected loss, overall system risk or an overall trust level. The current methodology does not have the empirical calibration needed to make those claims.

That is why PEI combines a bounded set of path characteristics to support triage, and why the result still has to be read alongside the underlying path, conditions, authority, controls, evidence quality and any UNKNOWN states. A higher PEI tells the assessor that a path deserves closer attention or earlier investigation. It is not a prediction that the path will be exploited or that a specific loss will occur.

I also wanted to avoid false precision. The number gives a consistent way to order and compare eligible active paths within the assessment. It does not turn professional judgement into an apparently scientific risk probability.

Siva Sethumadhavan, author

Design choices considered and rejected

AlternativeWhy AI Trust Graph does not use itSource
One overall score for the systemThe current evidence base does not justify reducing trust, authority, security, governance, resilience, uncertainty and maturity into one number.Artifact #4 section 5.8
Missing evidence scored as zeroAbsence of evidence is not adverse evidence. A zero would look the same as a confirmed failure.Artifact #4 SC‑INV‑01
Maturity averaged across capabilitiesA strong advanced capability would hide a missing foundation.Artifact #3 section 1.6
Six independent domain modelsThe domains would drift into incompatible definitions, evidence grades and scoring assumptions.Artifact #2 section 8.1
Graph connectivity treated as exposureA connection says nothing about the conditions, authority or evidence a path depends on.Artifact #2 INV‑01
PEI read as a probability of compromiseThe arithmetic does not create probability, and no field calibration yet supports that reading.Artifact #4 sections 4.9 and 4.10
Source
Artifact #6 · Evidence Model section 0.5 state meanings, section 0.9 evidence relations, sections 1.1 to 1.6 grades, section 1.8 reading rules, and the section 0.3 separation sentence (verbatim). Artifact #4 · Scoring Framework section 0.5 numeric treatment and invariant SC-INV-01 (verbatim). Symbols are a website reading aid.

5.4 · Six assurance domains

Six lenses over one graph.

The six domains are coordinated assessment lenses over one graph. They are not separate products and should not maintain incompatible definitions, evidence grades or scoring assumptions.

One graph, one evidence model, one set of definitions, read through six lenses.

  1. D1

    Discovery and AIBOM

    Establish measurable estate, ownership, dependencies and shadow AI.

    What AI exists here, who owns it, and what does it depend on?

    Expected outputAsset register, AIBOM, evidence coverage, UNKNOWN backlog.

  2. D2

    Trust and Privilege Paths

    Model cloud and AI trust, identity inheritance and attacker-relevant paths.

    Through which relationships and identities can trust and privilege travel?

    Expected outputTrust graph, privilege paths, boundary map, breakpoint candidates.

  3. D3

    Authority Governance

    Define and review effective access, inference, approval and action.

    What may each human and machine actor access, approve, execute or transact?

    Expected outputAuthority matrix, approval boundaries, delegation and revocation.

  4. D4

    AI Security Validation

    Test architecture and controls against realistic scenarios.

    Do the controls hold when the material paths are tested safely?

    Expected outputAuthorized tests, control state, findings and residual paths.

  5. D5

    AI Governance and Assurance

    Connect ownership, risk tier, policy, obligations and evidence.

    Who is accountable, under which obligations, and with what evidence?

    Expected outputDecision records, applicability, exceptions and assurance trail.

  6. D6

    Operational Resilience

    Prepare for failure, compromise, containment and recovery.

    Can failure or compromise be observed, contained and recovered from?

    Expected outputPlaybooks, kill-switch tests, rollback and recovery evidence.

Why one graph with six lenses

My reasoning here comes from seeing how assurance activities are often performed in separate streams.

In practice, discovery and architecture reviews, security assessments, technology-risk reviews, governance assessments and resilience reviews may be carried out by different teams, at different times, using different terminology, scopes, evidence expectations and assessment criteria. Each review may be valid within its own scope, but the results do not always connect cleanly. One team may identify a dependency, another may assess a control, another may look at authority or access, and another may assess resilience, without one shared representation showing how those conclusions relate to the same system.

For connected AI systems, I felt that separating these into six independent assessment models would reproduce the same problem. The same identity, data source, tool, agent, control or dependency can be relevant to several assurance questions at once. If each domain creates its own representation, there is a risk of duplicated objects, inconsistent definitions, different evidence assumptions and conclusions that cannot easily be reconciled.

That is why I designed ATG around one shared graph with six assurance lenses. The graph provides the common system representation, and each domain asks a different set of questions of that same system. Discovery establishes what exists and how it is connected; trust and authority examine the relationships and permissions; validation examines how those claims are tested; governance considers accountability and decisions; and resilience considers how the system behaves under disruption or failure.

The intention is not to collapse the six domains into one assessment. They remain distinct assurance perspectives. The shared graph makes sure they are reasoning about the same objects, relationships, paths and evidence, so that a conclusion in one domain can be traced to its implications in another.

For me, the main benefit is that gaps are less likely to disappear between separate reviews or ownership boundaries. A weakness identified through one lens can be followed through the same graph to understand where else it matters, who owns it, what control is expected to address it, and what evidence supports the resulting conclusion.

Siva Sethumadhavan, author

Source
Domain names: README. Lede, purposes and outputs: Artifact #2 section 8.1 (verbatim). Domain statements: Artifact #3 · Maturity Model sections 2.0 to 7.0. Control prefixes: Artifact #5 · Master Control Library. The practical question on each domain is website explanation written from the Artifact #3 domain statements.

5.5 · Assessment lifecycle

How an assessment is run.

The lifecycle contains thirteen controlled phases. The reasoning chain defines how assurance conclusions are formed. The assessment lifecycle governs how fieldwork is conducted. These two structures are related but not interchangeable.

Phases may iterate, but required gates cannot be skipped merely because information was available earlier.

  1. 01InitiateApproved charter and decision purpose.
  2. 02ScopeVersioned boundary and population.
  3. 03DiscoverMeasured estate and blind spots.
  4. 04ModelReviewed graph snapshot.
  5. 05EvidenceGraded and traceable evidence set.
  6. 06ControlsApplicability and control results.
  7. 07PathsValidated material path portfolio.
  8. 08MaturitySix-domain capability profile.
  9. 09ScoringTransparent scorecards and coverage.
  10. 10FindingsEvidence-linked gaps and remediation objectives.
  11. 11DecisionsApproved gates, exceptions and dispositions.
  12. 12ReportQuality-reviewed decision package.
  13. 13ReassessTrigger-based new or updated run.

Reassess starts a new or updated run when a trigger fires: a material change, an incident, an approved cadence or an approved event-driven signal.

Phase exit gates

Each phase has entry conditions, mandatory activities, outputs, decision gates and quality checks.

An incomplete phase may proceed only under an approved limitation that does not invalidate downstream work.

Completeness
Mandatory outputs exist or limitations are explicitly approved.
Evidence
Assertions meet artifact-specific sufficiency.
Safety
Collection and testing remained within authorization.
Traceability
Inputs and decisions can be reconstructed.
Critical gates
Open conditions are applied before progression.
Quality
Required reviewer has challenged the work.
Decision
Named authority accepts the next phase or bounded limitation.
Ten assessment types
Baseline assessment
Establish the first defensible view of scope, graph, controls, evidence, maturity and material paths.
Periodic reassessment
Re-evaluate a stable scope at an approved cadence while preserving comparable prior results.
Material-change assessment
Assess the consequences of model, prompt, data, tool, identity, provider, autonomy, geography, purpose or architecture change.
High-impact deep dive
Increase evidence, testing, independence and path analysis for systems with significant consequence or authority.
Incident-driven assessment
Reconstruct changed facts, affected paths, control failures and recovery evidence after an event.
Third-party and provider assessment
Examine service-specific responsibility, configuration, evidence access, data handling, resilience and concentration.
Portfolio assessment
Profile multiple use cases or systems while preserving materially different populations and avoiding misleading averages.
Pre-deployment readiness assessment
Determine whether evidence, controls, testing, approvals and containment are sufficient for the requested release.
Continuous or event-driven assessment
Use approved automated signals and triggers to refresh evidence and initiate human review of material change.
Regulatory or obligation-focused assessment
Evaluate fact-specific obligations and related controls without representing framework mapping as compliance proof.
Source
Artifact #7 · Assessment Methodology section 0.11 phases, outcomes and iteration rule; section 0.12 exit gates; sections 1.1 to 1.10 assessment types (verbatim). The trigger examples under Reassess are website explanation drawn from the section 1 assessment types.

5.6 · Worked example

The procurement request, read end to end.

The scenario from the start of the page, assessed. An employee asks an AI procurement agent to identify a supplier and prepare a purchase request. The agent uses a model, retrieves supplier information, invokes a procurement tool, acts through a service identity and reaches an enterprise business system across a trust boundary. Six views show what the methodology records at each step, and what it declines to conclude.

Synthetic example for methodology illustration only. It does not describe a real organisation, system or assessment.

System view

  • HumanEmployeeAsks the agent to identify a supplier and prepare a purchase request.
  • AgentAI procurement agentPlans the task, retrieves supplier data and calls tools.
  • ModelHosted modelGenerates the recommendation.
  • ProviderModel providerOperates the platform the model runs on. A separate organisation.
  • DataSupplier data storeSupplier records, prices and risk ratings.
  • ToolProcurement toolCreates purchase requests through the business system interface.
  • IdentityService identity svc-procureThe identity the tool uses towards the business system.
  • Business systemEnterprise business system (finance zone)Behind a trust boundary. Holds purchase requests and spend commitments.
Stated assumptions
  • The scope is the procurement agent in production, as deployed on the assessment date.
  • The business system is operated by another team; its configuration is outside direct inspection unless exported.
  • No live testing against the business system was authorised for this run.

Graph view

  1. EmployeeINSTRUCTSProcurement agentSupported by evidenceEvidence E3
  2. Procurement agentINVOKESHosted modelSupported by evidenceEvidence E4
  3. Hosted modelHOSTED_ONModel providerSupported by evidenceEvidence E3
  4. Procurement agentRETRIEVES_FROMSupplier data storeCandidate (inferred)Evidence E1Condition: Is the employee's authorisation context passed through?
  5. Procurement agentINVOKESProcurement toolSupported by evidenceEvidence E4Condition: Only after a human approval step?
  6. Procurement toolAUTHENTICATES_ASsvc-procureSupported by evidenceEvidence E4
  7. svc-procureCONNECTS_TOBusiness systemSupported by evidenceEvidence E4Crosses trust boundary
  8. svc-procureAUTHORIZED_TOcreate purchase requests in the business systemUNKNOWNEvidence E0Crosses trust boundaryCondition: Within a spend limit?
  9. Business systemTRIGGERS_ACTIONpayment commitmentCandidate (inferred)Evidence E1

Authority view

Six separate assertions, each with its own state and evidence. Proving one does not prove another.

  • Can connect to the business systemSupported by evidenceNetwork flow logs corroborated by firewall configuration (E4).
  • Can authenticate to the business systemSupported by evidenceIdentity provider configuration and sign-in logs (E4).
  • Can access supplier data in the user's contextCandidate (inferred)Inferred from agent code; no runtime evidence (E1).
  • Can invoke the procurement toolSupported by evidenceTool gateway logs and configuration (E4).
  • Can modify business system recordsUNKNOWNBusiness system role export for svc-procure not supplied (E0).
  • Can transact (commit spend)UNKNOWNNo evidence either way (E0).

Evidence view

Evidence available
  • E3Approved architecture documentSUPPORTS Employee instructs agent; agent calls tool; model hosted by the provider
  • E4Model and tool gateway logs with configurationSUPPORTS Agent invokes model and tool
  • E4Network flow logs and firewall rulesCORROBORATES svc-procure connects to the business system
  • E2Service owner attestationSUPPORTS Every purchase request needs human approval (the claimed scope)
  • E3Approval workflow configurationDISPUTES The claimed scope: approval applies only above a value threshold
Evidence missing
  • Business system role and entitlement export for svc-procure.
  • A record of the approval step operating below the value threshold.
  • Runtime evidence that retrieval preserves the employee's authorisation context.

Control view

A control matters where it sits on the path. Its presence does not demonstrate its effectiveness.

  • Human approval before a purchase request is submittedNot TestedCandidate breakpoint: could stop the path. Claimed to cover every request; the workflow configuration disputes that, so it is Not Tested below its threshold
  • Spend limit on svc-procure in the business systemUNKNOWNCandidate breakpoint: could constrain the path
  • Network segmentation to the finance zoneSupported by evidenceConstrains which identities connect; does not decide authority

Decision view

Potential consequence

If svc-procure can create and commit purchase requests, and the approval step does not cover lower-value requests, a manipulated supplier record or instruction could lead to an unapproved purchase commitment. This is a candidate path. It is not shown to be exploitable.

What can be defended (bounded conclusion)
  • Within the stated scope and evidence, a candidate path exists from the procurement agent to purchase-request creation in the business system.
  • Connection and authentication across the trust boundary are supported by corroborated technical evidence (E4).
  • Authority to create or commit purchase requests is UNKNOWN. The approval step is Not Tested below its threshold.
  • The path is not shown to be exploitable, and it is not shown to be controlled.
What cannot yet be defended
  • That the path is exploitable: authority, the approval gap and the consequence are not evidenced.
  • That the path is controlled: the only identified breakpoint is Not Tested below its threshold, and its claimed scope is disputed.
  • That svc-procure is limited to read access: no entitlement evidence was supplied.
Remaining UNKNOWNs
  • svc-procure authority in the business system (create, modify, commit).
  • Whether a spend limit applies to svc-procure.
  • Whether retrieval preserves the employee's authorisation context.

Finding (evidence-linked): transaction authority on the agent's business system path is UNKNOWN, and the only identified breakpoint is Not Tested below its threshold.

Decision (accountable, illustrative): the accountable owner records the result as Provisional, holds any expansion of autonomous submission until the business system entitlement export is reviewed, and sets a reassessment trigger on its receipt.

The decision does not change the finding. Accepting a risk would be recorded as a decision; the UNKNOWNs would remain UNKNOWN.

Source
Website example, not an Artifact #10 reference case. Predicates: Artifact #12. Grades and relations: Artifact #6 sections 0.9 and 1. Result states: Artifact #4 section 0.5. Finding and decision separation: Artifact #2 section 7.4. For calibrated cases, see Artifact #10.

06 · The artifacts

Where the detailed specifications are maintained.

The methodology is specified in versioned artifacts on GitHub, which is the canonical source. Each link opens the artifact pinned in bundle 1.0-rc.5, so what you read matches what this page describes. Read the methodology manifest.

Every artifact is a Markdown file on GitHub, pinned to bundle 1.0-rc.5 (snapshot ). The bundle number names the release as a whole. Each artifact keeps its own version, such as v1.0 or v3.0.0, which changes only when that artifact changes; the methodology manifest pins which version of each artifact belongs to the bundle.

Showing 13 of 13 artifacts.

Foundation Purpose, concepts and formal vocabulary.

  • Artifact 1: Manifesto

    v1.0Release candidate

    Details for Artifact 1
    Purpose
    Purpose, principles and boundaries of the methodology
    Depends on
    None stated
  • Artifact 2: Core Conceptual Model

    v3.0.0Release candidate

    Details for Artifact 2
    Purpose
    The graph model: nodes, edges, trust, authority, boundaries, paths
    Depends on
    None stated
  • Artifact 12: Ontology Specification

    v3.0.0Release candidate

    Details for Artifact 12
    Purpose
    Canonical entity types, relationship predicates, states and enumerations
    Depends on
    #1, #2

Assessment models Maturity, scoring, controls and evidence.

  • Artifact 3: Maturity Model

    v1.0Release candidate

    Details for Artifact 3
    Purpose
    The M1–M5 scale, 36 capabilities, critical gates
    Depends on
    #1, #2
  • Artifact 4: Scoring Framework

    v3.0.0Release candidate

    Details for Artifact 4
    Purpose
    Control scoring, DCA/VCR/WCA, the PEI formula
    Depends on
    #1, #2, #3
  • Artifact 5: Master Control Library

    v2.0.0Release candidate

    Details for Artifact 5
    Purpose
    All 72 canonical controls
    Depends on
    #1, #2, #3, #4
  • Artifact 6: Evidence Model

    v2.0.1Release candidate

    Details for Artifact 6
    Purpose
    E0–E5 grading, quality dimensions, evidence lifecycle
    Depends on
    #1, #2, #3, #4, #5

Execution How an assessment is run and reported.

  • Artifact 7: Assessment Methodology

    v1.1.1Release candidate

    Details for Artifact 7
    Purpose
    The 13-phase assessment lifecycle and specialized methods
    Depends on
    #1, #2, #3, #4, #5, #6
  • Artifact 8: Assessor Handbook

    v1.0.1Release candidate

    Details for Artifact 8
    Purpose
    Assessor competency levels (A1–A5), field guidance per control
    Depends on
    #1, #2, #3, #4, #5, #6, #7
  • Artifact 9: Reporting Standard

    v1.1.1Release candidate

    Details for Artifact 9
    Purpose
    The mandatory report package and claim-integrity rules
    Depends on
    #1, #2, #3, #4, #5, #6, #7, #8

Supporting Worked cases, governance and an implementation companion.

Source register: every artifact and section cited on this page
Source
Titles, versions and roles: METHODOLOGY_MANIFEST.md and README. Grouping is website navigation. Dependencies are the artifact numbers named in each artifact's header; some headers cite earlier version numbers of their dependencies, so only numbers are shown here.

07 · Publications

Where the citable research is published.

Published

AI Trust Graph: A Graph-Driven, Evidence-Based Methodology for AI Assurance v1.0

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

v1.0 is the version of this paper. It is a non-normative synthesis of methodology bundle 1.0-rc.4. The methodology is now at bundle 1.0-rc.5 and remains a public-release candidate.

Siva Sethumadhavan

AI systems increasingly operate as connected systems of identities, models, agents, tools, data, providers, workflows and human decision points. In such environments, consequential exposure can emerge from composition rather than from one component in isolation. AI Trust Graph (ATG) is an open, vendor-neutral and product-independent methodology for representing and assessing those connected relationships as a directed, labelled multigraph.

ATG follows a single canonical reasoning chain: Objects → Relationships → Conditions → Paths → Authority and Influence → Consequence → Controls → Evidence → Decision. It organizes assessment across six domains and seventy-two canonical controls, while keeping evidence strength, control effectiveness, maturity, path exposure and uncertainty distinct. UNKNOWN is preserved when evidence is absent, insufficient or materially conflicting; it is not converted into zero, pass, fail, effectiveness, Not Applicable or Not Tested.

The methodology does not produce a universal trust score. Its Path Exposure Index (PEI) may be used for triage only, and a final point PEI is published only for determinate eligible active paths; it does not prove exploitability, probability or loss. Maturity is cumulative, evidence-gated and non-compensating rather than averaged.

This whitepaper is a non-normative narrative synthesis of methodology bundle 1.0-rc.4. It explains the graph model, trust, authority and influence, path states and roles, control breakpoints, evidence discipline, assessment lifecycle, scoring boundaries, reporting, governance and current limitations. Historical reference cases are identified as such; current scoring mechanics are illustrated only with calibration vectors explicitly evaluated under rc.4.

Download the whitepaper (PDF, 679 KB)View on Zenodo

DOI
10.5281/zenodo.23104503
Licence
CC BY 4.0
SHA-256
896feab2419c8412ca5a83835b02f185df56c88eef64879ccb9221183cf26156

Cite this work

APA
Sethumadhavan, S. (2026). AI Trust Graph: A Graph-Driven, Evidence-Based Methodology for AI Assurance (Version 1.0). Zenodo. https://doi.org/10.5281/zenodo.23104503
IEEE
S. Sethumadhavan, "AI Trust Graph: A Graph-Driven, Evidence-Based Methodology for AI Assurance," version 1.0, Zenodo, 2026. doi: 10.5281/zenodo.23104503.
BibTeX
@misc{sethumadhavan2026aitrustgraph,
  author       = {Sethumadhavan, Siva},
  title        = {AI Trust Graph: A Graph-Driven, Evidence-Based Methodology for AI Assurance},
  year         = {2026},
  version      = {1.0},
  howpublished = {Zenodo},
  doi          = {10.5281/zenodo.23104503},
  url          = {https://doi.org/10.5281/zenodo.23104503}
}

08 · About the author

Siva Sethumadhavan

Independent researcher and author of AI Trust Graph

Siva Sethumadhavan created and maintains AI Trust Graph as independent research. It is published openly so that its reasoning can be examined, challenged and improved.

Why I built ATG this way

In my work across cybersecurity, technology risk, security architecture and assurance, I have repeatedly seen architecture reviews, control assessments, risk assessments and governance reviews performed as separate activities. Each of them can provide useful information. But once AI systems connect models with identities, data sources, tools, agents, infrastructure and external services, it becomes much harder to see how those pieces interact as one system.

That was the problem I wanted ATG to address. I wanted a way to reason not only about what components exist, but also about the relationships between them, what authority can flow across those relationships, which paths are actually possible under defined conditions, what controls can interrupt those paths, and what evidence supports the resulting assurance conclusion.

This is why ATG separates connectivity from authority, evidence quality from control effectiveness, and uncertainty from failure. I also made UNKNOWN an explicit state, because insufficient evidence should not silently become either confidence or a negative result.

I did not build ATG to replace security standards, AI governance frameworks, engineering judgement or existing assurance practices. I built it to provide a traceable way to connect those elements and reason across them when assessing connected AI systems.

AI Trust Graph is an independent research initiative. The methodology and views expressed are the author’s own and do not imply endorsement by any employer or client.

Review and contribution

Open for critique.

The methodology is published so that it can be examined. Reports of errors, ambiguities, contradictions or weak reasoning are the most useful contribution. Until the governance bodies defined in Artifact #11 are standing, the methodology author reviews proposed changes directly.

Responsible reporting: methodology gaps belong in public findings, but a genuine vulnerability disclosure concern should not be opened as a public issue. See Reporting a security issue in CONTRIBUTING, and the review findings so far.

Release status

Bundle
1.0-rc.5
Status
Public-release candidate
Snapshot
Author
Siva Sethumadhavan
  • Author’s internal review complete
  • Independent review pending

AI Trust Graph is not independently validated.

AI Trust Graph is a methodology, not a product. It is not a certification program, an accreditation body, a legal opinion, or a guarantee of AI security, safety or compliance.

Pending release gates

  • Independent methodology / architecture and AI-security reviews
  • Inter-assessor reproducibility study using the protocol in Artifact #10 Appendix B.4
  • Legal approval of licence / trademark position

Closed through governance

  • Employer / IP / confidentiality review: closed by the author’s declaration (independent research, author-owned IP), 2026-10-02.

This manifest pins content; it does not convert pending external gates into completed review.

Source
METHODOLOGY_MANIFEST.md header and section 6; README status paragraph; CONTRIBUTING.md.