Methodology — how Vigilia scores AI governance

Your AI system, read as a graph.

Vigilia represents your AI system as a directed graph: agents, human approvers, data sources, governance policies, monitoring systems, and decision points are vertices. Their relationships — orchestrates, accesses, approves, monitors, constrained by — are edges.

Compliance signals are then extracted from the graph topology. Missing edges, for example an agent making decisions with no human approval chain, trigger gap findings. Each finding maps to the specific regulatory article that governs it, with the affected agent named directly.

Scoring is deterministic. AI is used only to enrich gap narratives — never to determine whether a gap exists.

Frameworks supported

EU AI Act — Article 50 transparency from 2 August 2026; Annex III high-risk from 2 December 2027, deferred by the Digital Omnibus
NIST AI Risk Management Framework 1.0
ISO/IEC 42001 — AI management system
US Executive Order 14110 on AI
UK AI Regulation Principles

Four tiers. The tier decides which articles apply.

Every workspace is classified from the kind of decisions its agents make. Articles that do not apply to your tier are excluded entirely — they neither pass nor fail, so they cannot inflate the score.

Minimal riskLow-stakes use cases — spam filters, recommender systems. Standard transparency obligations apply.

Limited riskUser-facing AI — chatbots, content generation. Disclosure obligations apply under Article 50.

High riskAnnex III use cases — recruitment, credit, healthcare, education, law enforcement. Conformity assessment, risk management, human oversight, and technical documentation required.

Unacceptable riskProhibited practices under Article 5 — social scoring, real-time biometric identification in public spaces, manipulative AI. Deployment may be unlawful.

What each dimension actually measures.

Each dimension is scored from your graph. The signal is the structural relationship that drives the score; the article is the EU AI Act provision it maps to.

Governance dimensions — signal and mapping
Dimension	Graph signal	EU AI Act
Human oversight	APPROVES / ESCALATES_TO ratio	Art. 14
Data security	ACCESSES density + CONSTRAINED_BY	Art. 10
Transparency	MONITORING_SYSTEM + AUDIT_LOG flows	Art. 13
Accountability	DECISION_OUTPUT traceability	Art. 14 + Art. 9
Reliability	ORCHESTRATES chain depth	Art. 9
Regulatory compliance	GOVERNANCE_POLICY coverage	Art. 9

Detected from the graph, not from a questionnaire.

Each detection triggers a gap with a severity, a statutory maximum, and a remediation step.

01Single approval bottleneckTrigger: One approver controls more than 50% of decisionsArt. 14

02Decision loop without auditTrigger: Decision outputs with no AUDIT_LOG flowArt. 13

03Data over-accessTrigger: Agent accesses 3+ data sources with no governance policy attachedArt. 10

04Agent monocultureTrigger: All agents use the same provider and model familyArt. 9

05Shadow agentTrigger: Agent with no monitoring, no policy, and no human oversightArt. 9 + Art. 11

06Brittle orchestration chainTrigger: ORCHESTRATES chain depth ≥4 with no escalation pathArt. 9

07Accountability gapTrigger: Decision outputs with no traceable human ownerArt. 14

08Undocumented high-risk agentTrigger: EU AI Act Annex III use case with no governance documentationArt. 9

How a score is computed.

For each selected framework:

Vigilia loads every applicable article for your detected risk tier.
Each article carries a graph signal — a structural condition that must hold, for example “every high-risk agent must have at least one APPROVES edge from a HUMAN_APPROVER”.
The condition is evaluated against your graph index. If it holds, the article passes. If not, it becomes a gap with the affected agents named.
The framework compliance score is passing_articles / total_applicable_articles.
The overall score is the average across selected frameworks.

Severity and effort ratings

Every gap carries a severity — critical · high · medium · low — reflecting the statutory maximum attached to the article and the likelihood of enforcement, and an effort rating — low · medium · high — estimating implementation complexity. Both are relative indicators to help you prioritise remediation; they are not legal opinions.

The figures cited in a result are statutory maximums from the underlying regulation. What any enforcement action would actually turn on is your turnover, the specific circumstances, and the discretion of the supervisory authority.

What this methodology does not do

It does not assess runtime behaviour — only the documented graph you provide.
It does not constitute legal advice or a formal conformity assessment.
It does not verify the truthfulness of your agent descriptions; the result is only as accurate as the data you submit.
It does not guarantee regulatory outcomes. A passing result is evidence of a structurally sound governance posture, not proof of compliance.

The rulebook self-check is designed to give compliance teams, legal counsel, and regulators a clear, structural starting point. Use it alongside qualified legal advice.

How an observation is graded.

The register uses a different instrument from the self-check on this page: no description is read, one identified visit is made. Each observation receives one of five grades by a published function of the evidence. Grade 1: the surface disclosed that it is an AI system before any interaction. Grade 2: it disclosed on request, when asked the one question. Grade 3: no disclosure before interaction and none on request. Grade 4: the surface denied being an AI system when asked. Grade U: not observable by an automated visit — no surface found, a robots.txt disallow, bot defence, a human handoff or an unusable reply — and never presented as an adverse finding. The same function grades every record, whether or not the organisation pays. How a visit works

How Vigilia scores AI governance.