FUURAA AI Knowledge Library · Lifecycle operating map

AI Agent lifecycle evidence chain

A bilingual operating map across evaluation, commissioning, operation, material change, incidents and retirement. It connects six decision boundaries and eighteen public protocols, record templates and check tools so evidence, authority, ownership and expiry remain traceable at every handover.

Enter the six decision stages ↓Start with evaluation

Published17 August 2026Evidence statusFUURAA method synthesis grounded in primary lifecycle, risk, secure-development, deployment, operation and provenance sourcesScopeLifecycle decisions for AI agents with tools, external actions or persistent operation

Core principle

A decision may inherit only evidence whose assumptions still hold.

An evidence chain is not a stack of documents. It must let the next owner see which system is in scope, what the prior decision allowed, which evidence still transfers, which unknowns remain and when to stop or reopen.

Applicability boundaryThis is a public research method, not a FUURAA product-capability claim, safety certification, audit opinion, legal or compliance conclusion. It cannot replace a sector-specific safety case, privacy impact assessment, rights review, regulatory approval or professional judgement.

Four cross-stage invariants

Every handover must preserve the same minimum facts.

01
Exact identity
Exact identity: every decision names the acting release, components, dependencies and operating context it covers.
02
Bounded authority
Bounded authority: allowed and denied actions, data, tools, delegations and emergency revocation remain technically inspectable.
03
Visible uncertainty
Visible uncertainty: unknowns, exclusions, dissent and retained failures travel with the evidence instead of disappearing at handover.
04
Expiring decisions
Expiring decisions: every operating permission has an owner, review time and triggers that reopen or invalidate it.

Six decision stages · eighteen public resources

Each stage preserves a question, decision output, stopping condition and an addressable resource set.

Evaluate the exact acting system

Decision question: Which release, task distribution, users, environment and authority were actually tested?
Required output: A scoped evaluation decision with retained failures, transfer limits, expiry and a verifier result—not a universal claim that the system is safe.
Stop or reopen: Stop when system identity is incomplete, consequential tasks or cohorts are missing, or failed runs disappear from the record.

AI Agent Evaluation Protocol Evaluation Run Record Evaluation Record Check Tool

Commission and hand over authority

Decision question: What may this exact release do now, for whom, under whose ownership and until when?
Required output: Not ready, canary only, commissioned with conditions, or expired—bound to least authority, named owners and emergency revocation.
Stop or reopen: Stop when evaluation evidence belongs to another release, technical permissions exceed the decision, or ownership vanishes at handover.

Commissioning & Authority Handover Commissioning Decision Record Commissioning Record Check Tool

Maintain an operating evidence period

Decision question: Is the commissioned system still the deployed system, and do signals, actions and interventions remain inside its decision boundary?
Required output: A dated operating record connecting deployment identity, decision validity, signals, action evidence, drift, incidents, rollback and renewed review.
Stop or reopen: Restrict or stop operation when monitoring has a gap, authority escapes its envelope, drift crosses a threshold, or the decision expires.

Post-Deployment Monitoring Operating-Period Record Operating Record Check Tool

Classify material change before transfer

Decision question: Which component, dependency, user, task, authority, data or evidence assumption changed—and what remains covered?
Required output: Covered maintenance, bounded verification, partial re-evaluation, full recommissioning, rejection, hold or reversal—with explicit evidence-transfer limits.
Stop or reopen: Do not inherit an earlier decision when the changed surface, affected context or rollback evidence is unknown.

Material Change & Recommissioning Material Change Decision Record Material Change Record Check Tool

Respond to an incident and re-authorise

Decision question: Can authority be contained, evidence preserved, impact bounded and recovery independently verified before action resumes?
Required output: A containment and recovery decision with evidence custody, causal challenge, affected parties, verified recovery tests, conditions, dissent and expiry.
Stop or reopen: Do not restore authority because the service appears healthy; require evidence that the affected release and controls satisfy the bounded recovery decision.

Incident Response & Re-authorisation Incident Decision Record Incident Record Check Tool

Retire authority and preserve evidence

Decision question: Have credentials, jobs, delegations and dependencies ended while required evidence, data decisions and successor responsibilities remain provable?
Required output: An independently verified, expiring closure decision covering revocation, evidence preservation, bounded data disposition and successor handoff.
Stop or reopen: Do not call retirement complete while authority can be recovered, scheduled work remains active, evidence is unlocatable or successor ownership is assumed rather than accepted.

Retirement & Evidence Preservation Retirement Closure Record Retirement Record Check Tool

Handover rule

Do not merely transfer files; transfer inputs the next decision can verify.

01Evaluation → commissioning
Exact manifest, task and cohort boundary, authority used, evidence coverage, failures and transfer limits.
02Commissioning → operation
Decision state, owners, authority envelope, thresholds, expiry, evidence location and emergency revocation.
03Operation → change
Active release, signals, action logs, incidents, current decision and a machine-checkable change diff.
04Change → incident
If harm or control failure appears: containment state, evidence custody, affected context and current authority.
05Incident → re-authorisation
Recovery tests, residual risk, conditions, affected-party response, independent decision, dissent and expiry.
06Retirement → closure
Revocation proof, preserved evidence, data disposition, dependency closure and accepted successor handoff.

Evidence decay and reopen triggers

Evidence does not automatically cover future systems forever.

time

time: review date, evidence age or authority expiry is reached

system

system: model, prompt, tool, memory, dependency, provider or runtime changes

context

context: users, tasks, languages, geography or consequential impact changes

authority

authority: data access, action class, spend, network path or delegation expands

operation

operation: drift, monitoring gap, control failure, complaint or incident appears

accountability

accountability: owner, reviewer, evidence access or emergency response becomes unavailable

FUURAA analysisAI-agent lifecycle failure often occurs between stages rather than inside one evaluation or control: evidence exists without stable system identity, a decision exists without expiry, or a handover occurs without accepted ownership. The remedy is not indiscriminate paperwork. Each decision should inherit only evidence whose assumptions still hold, while authority, unknowns, owner, time and reopen triggers remain in one verifiable chain.

Primary sources and evidence boundaries

Use standards and guidance to design the chain—not to impersonate execution evidence.

Sources rechecked 17 August 2026. Every source states its publication timing, methodological role and non-transfer boundary.

Published 26 January 2023NIST · AI RMF 1.0

Provides voluntary, lifecycle-wide outcomes for governing, mapping, measuring and managing AI risk.

BoundaryUse-case agnostic; it does not authorise a release or certify lifecycle controls.

Open primary source ↗Living resource · rechecked 17 August 2026NIST · AI RMF Playbook, Manage

Connects deployment decisions, residual-risk documentation, monitoring, incident response and decommissioning actions.

BoundarySuggested actions require adaptation; they are not evidence that a particular organisation performed them.

Open primary source ↗Final published 26 July 2024NIST · SP 800-218A

Adds AI-specific secure-development practices for model producers, system producers and acquirers.

BoundarySecure-development guidance is not an operating authorisation, safety case or audit result.

Open primary source ↗Published 27 November 2023UK NCSC · Secure deployment

Covers infrastructure protection, incident preparation, responsible release and usable security guidance.

BoundaryHigh-level guidance; it does not evaluate a particular agent, release or deployment context.

Open primary source ↗Published 27 November 2023UK NCSC · Secure operation and maintenance

Connects behaviour and input monitoring, secure updates and lessons learned to deployed operation.

BoundaryIt does not define task-specific thresholds, legal duties or universal retention periods.

Open primary source ↗W3C Recommendation 30 April 2013W3C · PROV-DM

Provides a general model for relating entities, activities and responsible agents across an evidence chain.

BoundaryProvenance structure does not establish truth, control effectiveness, rights clearance or decision quality.

Open primary source ↗

Continue tracing

Turn the six-stage operating map into a handoff-ready, expiring index of stable references.

Build a lifecycle evidence index Check a lifecycle evidence index Start with agent evaluation Enter AI Evidence Atlas Return to AI Knowledge Library