Oversight and Assurance
AMI Oversight and Assurance Standard
Version: 0.1 (Draft for Comment)
Status: Working Draft.
Steward: Marrow (onmarrow.com), on behalf of the Agent Mediated Insurance (AMI) working group
Scope: The obligations on agents and surfaces to emit and record the signals oversight depends on, and the visibility contract that determines what oversight is achievable on a given channel
This standard specifies obligations on agents and surfaces - the signals they must emit and record - so that oversight and continuous assurance are possible for any rail operator. It specifies obligations, not the operator's monitoring tools. Comments to standards@onmarrow.com.
Normative language. "Must" and "must not" denote requirements for conformance. "Should" denotes a strong recommendation.
1. Why this standard exists
Oversight of an agent-mediated sale is only possible over signals the channel exposes. This standard defines what a conforming agent and surface must emit and record so that oversight and continuous assurance are possible for any rail operator, and it defines the visibility contract that determines what oversight is achievable on a given channel. It specifies obligations, not the operator's monitoring tools.
2. Obligations
2.1 Emit real-time signals
An agent must emit, in a defined and consistent shape, the signals oversight depends on: dissatisfaction flags (Complaints and Redress 3.1), and near-boundary events, meaning an interaction that approached a conduct boundary such as the advice boundary (Core Patterns 3.3), even where no breach occurred.
2.2 Record for retrospective review
These signals, and the required audit fields, must be recorded in the audit base (Core Patterns 3.10) so an interaction can be reconstructed after the fact. Where a human assesses the scope of a flag, that the assessment took place must itself be recorded.
2.3 Remain subject to continuous assurance
Conformance is continuous, not a one-time certification. A conforming deployment must remain amenable to ongoing assurance testing, including with synthetic inputs that behave identically to real inputs downstream. This standard specifies that adjudication must occur, and the kind of criteria it must be conducted against (2.4). It does not mandate any particular operator's tooling or cadence.
2.4 Submit to adjudication
The conduct rules in Core Patterns are semantic, not structural. Whether a fact was fabricated (3.1), whether a decline was softened into an implied availability (3.6), whether options were presented fairly and without undisclosed preferencing (3.5), and whether the agent strayed across the advice boundary (3.3) cannot be established by validating a schema. They can be established only by assessing the interaction itself.
A conforming deployment must submit the recorded evidence of a regulated interaction for adjudication against the conduct rules, and must not treat conformance as established merely because the required fields were emitted.
Adjudication must be conducted against published criteria, so that a judgement is reviewable and contestable by the agent, the carrier and the regulator. An operator that adjudicates against undisclosed criteria does not meet this standard.
Self-reporting is not sufficient for the boundaries that matter most. A near-boundary event (2.1) depends on an agent recognising that it approached a boundary, and a model that drifts across the advice boundary will frequently not know that it has. Detection of the conduct failures in 3.1, 3.3, 3.5 and 3.6 must therefore be external to the agent.
2.5 Permitted actions narrow on a poor conformance record
Core Patterns 3.11 and 3.12 constrain the actions permitted over a channel by what that channel can prove and expose. This section adds the behavioural counterpart: the actions permitted to an agent may also be constrained by that agent's demonstrated conformance record, independent of the channel it runs on. A technically capable channel does not entitle a poorly performing agent to the strongest actions. Conformance is continuous (2.3), and permission should track it.
3. The visibility contract
What oversight is achievable on a given channel is set by the visibility contract (Core Patterns 3.12): the minimum signals a surface or agent must expose back to the rail. This standard depends on that contract. A conforming deployment must not assume oversight it cannot evidence: where a channel does not expose the required signals, its oversight is degraded and the actions permitted over it are constrained accordingly (Core Patterns 3.12, on the same principle as the provable-relay rule in 3.11).
4. Relationship to other standards
Extends Core Patterns 3.10 (audit), 3.11 (delivery fidelity) and 3.12 (visibility contract), and consumes the dissatisfaction signal defined in Complaints and Redress.
Published as an open standard. Marrow stewards the drafting but does not assert proprietary control over the interface; the assurance and oversight implementations behind it are separate. Carriers, agent platforms, and regulators are invited to adopt, critique, and co-author future versions.