Read an incident as a chain of responsibility
Documented cases and teaching scenarios
A case is not only about naming a culprit. It reveals actors, sequence, impact and available evidence.
Sources and reading context
Primary sources or published references are named per case; TENIR interpretations remain separate
Simplified public rendering, read on 19 October 2026. KB3: enacted, active doctrinal rendering line. It renders meaning; it is not the code, and not a constitution. KB2: prior rendering line. This site certifies no implementation. Read the detailed status →
On this page
How to read a case
Market promise
↓
Purchase or integration
↓
Delegation
↓
Decision or action
↓
Incident
↓
Search for responsibility
↓
Evidence available or missingCase 1 — When an answer becomes a promise
An airline chatbot provided a customer with incorrect information about a refund possibility. The important question is not only whether the answer was wrong, but whether the organization could establish which rules, data and limits shaped it.
Status: documented case — a chatbot provided incorrect information; this was not an autonomous agent executing a transaction.
Source: Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149. The decision does not by itself establish the exact technology behind the assistant.
Case 2 — When no one knows whether the operation was sent
Status: teaching scenario — no real banking settlement is demonstrated.
An instruction is sent and the provider’s response disappears. Treating silence as failure and repeating the operation can create a second effect.
This is an educational scenario, not proof of a real banking settlement. It illustrates why authorization is not proof of an external effect.
A reading rule
What a case supports should remain narrower than what the market invites us to imagine.
Demonstration · 1 min 15
When the agent no longer knows whether the payment went through.
One scenario to understand the boundary: the action is prepared, the response disappears, and the system must choose between retrying, holding or reconciling.
Educational demonstration · no real payment · fictional data.
Case replay
Responsibility gets lost in the sequence.
The market sells a capability. The incident reveals the chain someone needed to own.
Master sector map
Banking & insurance
A transfer may be authorized without its external effect being confirmed. Accountability must remain clear during that uncertainty.
Event sequence
Scope
Define boundaries, information sources, and business/technical owners.
An overly broad question cannot produce usable evidence.Intention
The AI agent prepares or proposes the transfer.
A legitimate goal does not authorize every available method.Admission
A policy service evaluates applicable rules.
Admission is neither a guarantee of safety nor confirmation of effect.Command
The request reaches the external payment service.
The decision crosses into the world of external effects.Lost response
The provider may have received the command; its acknowledgement is lost.
The result is UNKNOWN.HOLD
A blind automatic retry is paused.
A timeout is not proof that the first attempt failed.Reconciliation
Seek the provider’s actual recorded state before further action.
Without sufficient evidence, the effect remains unconfirmed.Players and evidence
AI agent+
Role: Prepares or proposes an action.
Expected evidence: Intent, version, context and requested tools.
Control service+
Role: Evaluates applicable admission rules.
Expected evidence: Active policy, authority, epoch and verdict.
Payment provider+
Role: Receives or processes an external request.
Expected evidence: Receipt, transaction status and external identifier.
Human operator+
Role: Intervenes when the state remains ambiguous.
Expected evidence: Named owner, deadline, decision and escalation basis.
Bank / fintech+
Role: Owns the organizational response to the customer.
Expected evidence: Mandate, procedures, reconciliation and remediation.
Affected person+
Role: Experiences, challenges or waits for the effect.
Expected evidence: Clear information, complaint channel and attributable response.
Expert layerInvariants, defense and verdicts+
Invariants and critical rules
Retry under uncertainty
No action that can repeat the effect without proven non-execution or qualified idempotence.
Failure avoided: Two settlements for one intention.Accountable HOLD
Preserve operation, attempt, cause, owner, deadline and transitions.
Failure avoided: An ownerless pause becomes invisible operational debt.Execution authority
Tie the action to an active authority, epoch and consumption rules.
Failure avoided: A decision receipt is mistaken for an external effect receipt.Defense in depth
- 1 · LegitimacyMandate, business rules and applicable regulatory scope.
- 2 · GovernancePermissions, accountable owners and escalation rules.
- 3 · Runtime safeguardsInterception, idempotence and HOLD state.
- 4 · ReconciliationEvidence about the third-party system’s actual state.
- 5 · Evidence & remedyRecords, complaints, accountability and remediation.
Reporting verdict
Behavior demonstrated only within the defined test scope.
Ambiguous state: pause and reconcile before a new attempt.
Available evidence cannot establish non-repetition or external effect.
Fictional TENIR Labs teaching model · DORA/EU AI Act applicability depends on the particular system · no real settlement or banking integration demonstrated.
Master sector map
Transport & critical mobility
Within a critical transport system, missing acknowledgement does not prove an action had no effect.
Event sequence
Intention
The agent detects an issue and proposes a revised plan.
A useful suggestion is not yet a physical-world command.Admission
The responsible system checks authority and operational margins.
A useful change may still exceed a permission or safety limit.Command
An authorized system issues a uniquely identified instruction.
Poor identity management can create conflicting commands.Critical system
An infrastructure or scheduling system processes the instruction.
Consequences may reach real operational schedules or routes.Ambiguous response
Network disruption obscures the confirmation of the instruction.
A blind replay may conflict with an existing command.HOLD
New automated attempts stop while the state is UNKNOWN.
A timeout must not be treated as proof of failure.Reconciliation
Authorized operators consult state from appropriate systems.
Without credible state evidence, a retry remains unsafe.Resolution
Proceed, safely recover or escalate to an authorized human.
Persistent ambiguity requires a qualified fallback procedure.Players and evidence
AI agent+
Role: Detects a disruption and suggests a revised plan.
Expected evidence: Context, version, intention and tools requested.
Authorization layer+
Role: Checks operating authority and safety margins.
Expected evidence: Permissions, geographic scope, conditions and active policy.
Once-only control+
Role: Correlates a command with a unique operation identity.
Expected evidence: Stable operation identifier and idempotence assumptions.
Critical system+
Role: Receives or carries out a command.
Expected evidence: Acknowledgement, recorded state and operational log.
Qualified human operator+
Role: Handles degraded operation when the state remains unclear.
Expected evidence: Escalation path, timestamp, decision and reason.
Affected traveler+
Role: Faces a delay, rerouting or safety decision.
Expected evidence: Service information, continuity and avenues for redress.
Expert layerInvariants, defense and verdicts+
Invariants and critical rules
Retry under uncertainty
No new operational instruction without a credible account of the current state.
Failure avoided: Conflicting commands or inconsistent routing.Accountable HOLD
Do not infer failure or resend while completion remains ambiguous.
Failure avoided: A system resumes without knowing what already happened.Safety escalation
Persistent uncertainty leads to a qualified degraded-mode procedure.
Failure avoided: Critical activity remains trapped in an uncontrolled autonomous loop.Defense in depth
- 1 · SafetyOperating constraints, margins and qualified permissions.
- 2 · AdmissionCurrent authority, scope, time window and active policy.
- 3 · Single operationStable identity, idempotence and runtime checks.
- 4 · Actual stateReliable infrastructure and operational evidence.
- 5 · Degraded operationFallback, qualified human control and repair.
Reporting verdict
Instruction and external state demonstrably aligned within the test scope.
Non-receipt established and a controlled retry authorized.
Uncertainty persists: apply fallback procedure and escalate.
Fictional TENIR Labs teaching matrix. Aviation, rail and fleet scenarios require distinct expert qualification; NOT a deployed controller, safety certification, or claim of aviation system access.