Decirance : AI-Agent Deployment Assurance

Decirance is an open-source AI-agent assurance framework and working demonstrator developed by Exovat Limited.

THE PROBLEM

Organisations preparing to deploy AI agents often have evaluation results, security controls, governance documents, architecture records and operational monitoring. However, these sources rarely provide one accountable answer to the central deployment question:

Should this specific AI agent be authorised to operate in this specific context, with these tools, permissions and conditions?

When the agent subsequently changes, organisations may either repeat the entire assessment unnecessarily or continue relying on evidence that is no longer valid.

THE DECIRANCE APPROACH

Decirance combines technical and organisational evidence into a traceable assurance case and a context-bound Deployment Permit.

The permit records:

• the agent’s identity and purpose;

• its approved operational context;

• authorised and prohibited actions;

• tool and data permissions;

• mandatory operating conditions;

• supporting assurance claims and evidence;

• permit duration and expiry;

• the named accountable approver;

• conditions that would restrict or suspend operation.

ASSURANCE DELTA

Decirance’s distinguishing capability is Assurance Delta.

When an agent’s model, prompts, tools, MCP configuration, permissions, credentials, data sources or operating environment change, Decirance identifies:

• which assurance claims are affected;

• which evidence remains applicable;

• which evidence has been invalidated;

• what tests or reviews must be repeated;

• whether the existing permit remains valid;

• whether the agent’s authority should be restricted or suspended.

This enables targeted reassessment while preserving evidence that remains justified.

HUMAN ACCOUNTABILITY AND EXPLAINABILITY

Decirance does not autonomously declare an AI system safe. Recommendations are produced through transparent assurance rules and evidence dependencies. A Deployment Permit requires approval by a named accountable person.

Exovat is also developing AI-assisted evidence-preparation capabilities for extracting proposed controls and evidence from documents, suggesting evidence-to-claim mappings, identifying contradictions and generating assurance-test proposals. AI-generated outputs remain subject to human review and do not independently authorise deployment.

CURRENT CAPABILITIES

The existing framework includes:

• Agent Passports;

• Context Contracts;

• Organisation Assurance Baselines;

• structured assurance claims and evidence;

• dependency-aware assurance graphs;

• material-change classification;

• targeted evidence invalidation;

• criticality and proportionality assessment;

• authority containment;

• tool-authority contracts;

• deterministic recommendation ceilings;

• Deployment Permit state management;

• reassessment planning;

• assurance and technical packs;

• CLI and CI/CD support;

• mappings to recognised AI risk and security frameworks.

WHO IT IS FOR

Decirance is designed for organisations deploying autonomous or tool-using AI in regulated, safety-relevant or consequential workflows, including:

• clean energy;

• financial services;

• health and life sciences;

• professional services;

• advanced manufacturing;

• defence;

• public services.

Typical users include AI product owners, security and assurance teams, risk owners, compliance functions, deployment operators, internal auditors and accountable executives.

OPEN-SOURCE AND COMMERCIAL MODEL

Decirance’s core assurance schemas, rules, taxonomies and integration interfaces are open source. Exovat’s planned managed offering will add organisational workspaces, role-based access control, evidence registries, collaboration workflows, signing, monitoring, integrations and sector-specific assurance packs.

CURRENT STATUS

Decirance is presently a working research demonstrator and open-source framework. Exovat is seeking suitable organisations for controlled demonstrations, design partnerships and operational pilots using representative evidence, workflows and AI-agent configurations.