Portfolio boundary
Define which agents, automations, processes, teams, providers and decisions are included in the assessment.
SPECIFICATION FOR CONTROLLED PILOTS
PORTFOLIO ASSURANCE · SYSTEMIC INTERACTIONS · HUMAN CONTROL
Aegis Portfolio Assurance
Assess whether an agent portfolio remains governable as permissions, decisions and dependencies begin to interact.
ASSURANCE LOOP
Unit of analysisSystem + portfolio
DecisionFIT · FIT WITH CONTROLS · NOT FIT YET · SPECIALIST REVIEW
EvidenceVersioned Assurance Case
ActivationHuman authorisation only
The unit of analysis is no longer one agent. It is the hybrid organisation.
Institutional Agent Graph
Represent agents, humans, data, tools, permissions, dependencies and decision paths as a connected system.
Composite authority
Detect capabilities emerging when several agents combine permissions and actions they do not hold individually.
Common-cause failure
Identify shared models, providers, identities, data or infrastructure capable of affecting the whole portfolio.
Aggregate autonomy
Assess how much effective decision-making has been automated even when each component declares limited autonomy.
Institutional coherence
Examine whether local objectives, criteria and automations still reinforce the declared institutional logic.
Human control effectiveness
Check whether people retain information, time, authority and real ability to stop or reverse.
Portfolio Assurance Case
Integrate claims, evidence, uncertainty, correlated risks, controls, expiry and revalidation triggers.
The problem that agent-by-agent audits do not reveal
Many reasonable decisions, taken separately, can produce an organisation that nobody deliberately designed.
Composite authorityThe combination can execute what no single agent can do alone.
Synthetic consensusSeveral agents agree because they share a model, data or bias, not because of independent validation.
Institutional driftExceptions, permissions and automations gradually change how the organisation decides.
Institutional amnesiaDecisions and controls persist after their rationale, owner or context has been lost.
Honest status: methodological specification suitable for controlled pilots. It is not a public standard, certification, prediction of institutional behaviour or employee-surveillance tool.
A complete cycle, not an isolated audit
Each stage produces a decision, an accountable owner and enough evidence for the next stage not to rely on blind trust.
NeedOrbyn→DesignBlueprint→AssessmentAegis→PressureARIA→AuthorisationHuman→ObservationVigilance→RevalidationPortfolio
NEXT DECISION
Start a conversationThe first question is not what AI can do.
It is which system you are prepared to authorise, under which limits and with what evidence.