Agent Reliability Must Be a Bounded, Falsifiable Claim
A recent working draft offers a useful assurance pattern for consequential agents: state the precise boundary, evidence, controls, validity period and failure conditions of a specific deployment.
“Is this agent reliable?” is the wrong question when an agent can retrieve sensitive information, write records, send messages, trigger a workflow or delegate work onward. Reliability is not a property that floats above the deployment. It is a claim about a particular system, in a particular environment, with particular permissions, safeguards and failure conditions.
A September 2026 working draft, Agent Reliability Profiles in Financial Services, proposes a structured per deployment assurance artifact. It calls the artifact an Agent Reliability Profile. The paper is a working draft circulated for comment, not a standard, certification or endorsement of Anunnaki.Live. Its importance is directional: it gives institutions a more disciplined language for the boundary within which an agent may be trusted.
A reliability claim needs a boundary
The proposed profile frames an agent through four connected dimensions: its autonomy tier, operational design domain, permitted and prohibited classes of action, and control envelope. The result is not a score. It is a bounded and falsifiable proposition: this specific deployment functions reliably under stated conditions.
That distinction matters. The same base model may be a low consequence research assistant in one environment and a high consequence operational agent in another. A read only agent in a sandbox is not the same risk object as an agent that can reach production records, hold persistent credentials or transmit information externally. The relevant unit of assurance is the deployment, including its tools, policy configuration, credentials, environment, accountable owner and version in force.
Claims must expire when the system changes
The draft makes a second point that is easily missed in generic AI governance language: reliability evidence is time bounded. A material change in the model, tool catalogue, credentials, environment or controls can narrow or void the claim. A profile therefore needs a validity period and stated invalidation triggers, not merely a test report produced once.
This is a useful discipline for professional work automation. A professional system should not treat an old credential, old approval, old evidence set or old configuration as silently current. Change must be legible. The reason the earlier conclusion was permitted, the boundary then in force and the event that later changed it must remain inspectable.
What this means for a Professional Anunnaki Twin
A Professional Anunnaki Twin must be more than a recognisable identity or a fluent interface. It represents a real professional within the bounded authority of the ANUNNAKI Mathematical Universe. It must distinguish what it can prepare from what it can execute, what it may communicate from what requires human approval, and what it must refuse or escalate.
The working draft's four dimensions provide a useful external test for that intended architecture. A Professional Anunnaki Twin should have a stated role and operating domain. Its permitted actions should be explicit, including prohibited acts. Its control envelope should include evidence admission, authority boundaries, expiry horizons, human reserved judgment and the technical controls around connected systems. Those are design requirements, not a claim that every current Anunnaki.Live deployment already has an Agent Reliability Profile.
Replay turns the assertion into an inspectable record
The paper also calls for evidence provenance, defined performance floors, cryptographic integrity, machine readability and traces that allow reconstruction of evaluated runs. This is consistent with the purpose of Replay. A material professional outcome should not rest on the statement that a system was reliable. A relying party should be able to inspect the request, the evidence, the boundary, the authority decision, the action, the outcome and any later expiry or revocation relevant to reliance.
Replay does not make a decision right by itself. Nor does a log establish authority after an impermissible act. Its role is to preserve the determined record of why an action was permitted or refused, which enables challenge, accountability and continuity.
The limit is as important as the opportunity
The draft expressly limits its own scope. An Agent Reliability Profile does not replace data governance, cybersecurity, fairness, explainability, compliance or an institution's broader governance responsibilities. That limitation is correct. No single artifact can make a consequential system safe or lawful.
The strongest lesson for Anunnaki.Live is therefore modest and practical. Do not market a Professional Anunnaki Twin as generally reliable. State the professional boundary, the evidence and control conditions, the permitted actions, the human reservations, the expiry horizon and the replayable receipt. Then make those statements testable.
Source and status
Agent Reliability Profiles in Financial Services, September 2026 working draft circulated for comment. It is a financial services proposal and does not create a compliance obligation or validate Anunnaki.Live.