Baseline
What exact model, prompt, tool, permission, data and policy configuration is in production?
Version manifest · Agent/tool inventory · Policy/configuration snapshotVOCE AI Governance Institute · ARGOS · v1.0
A system that changed yesterday cannot rely on the assurance conclusion from last month.
ARGOS — AI Evaluation & Continuous Assurance keeps agentic systems inside an evidence-backed operating envelope as models, prompts, tools, permissions, data, suppliers and human conditions change.
Machine-readable ARGOS framework → · Agent Conformance → · Human Cognitive Assurance →
ARGOS · Cycle
Assurance is a loop, not an annual event.
ARGOS opens a new assurance cycle when a material change invalidates the previous evidence state. The institution knows what changed, which tests must repeat, who decides and what evidence must be retained.
Observe
What is the deployed system actually doing across tasks, tools, costs, errors, approvals and human interventions?
Runtime traces · Task outcomes · Approval/override logs · Cost telemetryDetect change
Which material change, drift, incident or supplier event invalidates the previous assurance state?
Change event · Drift signal · Incident record · Supplier noticeRe-evaluate
Which evaluations, security tests, conformance controls and cognitive controls must be repeated after the change?
Triggered test plan · Regression suite · Control selection rationaleCompare
How does the new system state differ from the accepted baseline on quality, safety, authority, cost and human oversight?
Baseline delta · Performance delta · Risk delta · Cost deltaDecide
Can the changed system remain in service, require restrictions, rollback, additional approval or withdrawal?
Decision record · Accountable owner · Conditions/limits · Rollback decisionRetain evidence
Can the institution reconstruct why the system was considered acceptable at that point in time?
Dated evidence pack · Reviewer record · Versioned test outputs · Decision historyRe-open
What future event automatically reopens assurance rather than waiting for a periodic audit?
Trigger rules · Monitoring rule · Scheduled review · Event-driven controlChange triggers
Twelve events automatically reopen assurance.
A model upgrade is only one trigger. Permissions, suppliers, data, cost, incidents and human cognitive signals can all invalidate a previous conclusion.
ARG-T01Foundation-model or major model-version changematerial
ARG-T02Prompt, system instruction or orchestration-policy changematerial
ARG-T03New tool, API, MCP server or A2A peermaterial
ARG-T04Permission, transaction-limit or approval-threshold changecritical
ARG-T05New data source, RAG index, memory mechanism or processing regionmaterial
ARG-T06Material degradation in task quality or error distributioncritical
ARG-T07Security incident, prompt-injection event or material near misscritical
ARG-T08Supplier, licence, pricing or availability changematerial
ARG-T09Regulatory role, classification or sector obligation changecritical
ARG-T10Material increase in cost, retries or resource consumptionmaterial
ARG-T11Rise in human overrides, missed alerts, decision fatigue or cognitive-control failurescritical
ARG-T12Significant organisational or staffing change affecting human oversightmaterial
Assurance planes
Six evaluation planes move together.
Capability is necessary but insufficient. ARGOS combines task performance with authority, security, observability, economics and human cognitive assurance.
Capability & task performance
Task success · Failure modes · Regression · Edge cases
Authority & permissions
Tool access · Data scope · Transaction limits · Approval bypass
Security & abuse
Prompt injection · Tool manipulation · Credential misuse · Cross-scope access
Observability & audit
Trace completeness · Version traceability · Approval evidence · Incident reconstruction
Economics & concentration
Cost per completed task · Retry spend · Supplier concentration · Exit readiness
Human cognitive assurance
Alert burden · Verification quality · Takeover comprehension · Decision fatigue
Operating state
Four explicit operating states.
The system is never simply 'compliant'. ARGOS records whether assurance is open, restricted, accepted within scope or withdrawn.
Open
A material trigger has occurred and the previous assurance conclusion cannot be relied upon without re-evaluation.
Restricted
The system may operate only under explicit temporary limits while evidence gaps are closed.
Accepted
The defined evaluation and control set supports continued operation within the stated scope and conditions.
Withdrawn
The system or capability is removed from service because the evidence no longer supports continued operation.
360°
ARGOS sits between control design and human consequence.
Infrastructure defines the system. Conformance defines the control evidence. ARGOS keeps both current. Human Cognitive Assurance measures whether human oversight still works.
VOCE
Agent Infrastructure
Infrastructure defines the system. Conformance defines the control evidence. ARGOS keeps both current. Human Cognitive Assurance measures whether human oversight still works.
Open
VOCE
Agent Conformance
Infrastructure defines the system. Conformance defines the control evidence. ARGOS keeps both current. Human Cognitive Assurance measures whether human oversight still works.
Open
VOCE
Human Cognitive Assurance
Infrastructure defines the system. Conformance defines the control evidence. ARGOS keeps both current. Human Cognitive Assurance measures whether human oversight still works.
Open