Kognita 0.6 'Claims': a statement has to earn its way to a client

Context

Up to 0.5, Kognita governed what an agent was allowed to do. 0.6 is about what an agent is allowed to say. The goal in the roadmap is plain: every material statement should be traceable, typed, current and checked before a human relies on it. Agents reason over trusted data. They do not decide what is true.

0.6.0 is released on PyPI. Kognita is still alpha, and this release is about putting that idea into code, not claiming it is finished.

Decision: every material claim carries its own record

A claim is stored with its type, its source, the calculation behind any figure, its verification status and the business definition it uses. Its content sits in the retention store and its hash sits in the evidence chain, so a reconstruction shows exactly what a relationship manager saw.

Release checks run in a fixed order: stale status, audience, label, approved source, calculation, verification, then definition. The first failure is the reason given. A fixed order means two people looking at the same refusal get the same answer.

Decision: the gateway refuses prose that was never released

For a high-materiality use case, the AI gateway does not return model text unless that exact text was recorded and released for the audience. Anything else is refused, and so is anything the gateway cannot check because its store is unavailable. That is the line that turns the claim record from good housekeeping into a control.

Decision: stale data fails closed

Facts arrive with a contract: value, source, owner, as-of date and a freshness threshold set by policy. When a fact is stale, the use case decides whether to refresh it, flag it or remove it, and the default is to suppress it. 'Only use current data' is enforced in code rather than asked of the model.

Decision: sources, figures and definitions are registered, not trusted

• Each use case has a register of approved sources. The model's own knowledge is never a source for a regulated claim.
• When a house view is superseded, dependent open recommendations and claims are marked stale, and the kognita evidence affected command lists the clients, past recommendations, past communications and open opportunities touched by it.
• Kognita does no financial maths. An authoritative figure must cite a registered calculation service, or it is blocked.
• Governed business terms, such as assets under management, are registered with an owner and version. A claim using one must cite the version in force, and the affected-records query works for definition changes too.

Decision: the checker cannot be the author

From tier 2 upwards, a registered verifier has to try to prove a claim wrong before it reaches a relationship manager or client. The verifier may not share the producer's model or provider, and both models are pinned in evidence, so independence is checked at the decision rather than asserted. No model marks its own homework.

Decision: learning is proposed, never applied

• An AI inference about a client becomes a relationship record only after a relationship manager verifies it.
• A learned change to a standard needs Knowledge Lead approval and produces a new, effective-dated standard version. Outputs record which version produced them.
• Exporting interactions for training or evaluation is a governed request with its own approval, and client exclusions hold for every future export.
• Sensitive-category inferences are blocked unless the use case permits them, with a pluggable detector. A vulnerability flag goes to a relationship manager and cannot change treatment without a recorded human decision.

Decision: evaluation and reasoning are evidence, not permission

A tier 2 or higher agent cannot be activated until recorded evaluation results meet the use case's thresholds. A fairness report compares recommendation, contact and prioritisation rates across segments the bank defines, including proxy attributes, and flags disparities to the risk owner.

Where a provider returns a reasoning trace, it is retained and shown as the model's own account. No policy may allow an action because the model's reasoning says it is permitted.

Known limits

• Fairness disparities are flagged to the risk owner. They are not registered as monitored metrics.
• The gateway records the use-case standard version in force. A version supplied by the caller is ignored.

Take this with you

If an agent's words can reach a client, treat each statement like a record: where did it come from, is it current, and who other than its author checked it? If you cannot answer, it should not be sent.

Sources

Kognita is an open-source alpha. This note comes from kognita 0.6.0 on PyPI, Kognita v0.6.0 release notes, Kognita CHANGELOG: 0.6.0, Kognita README: Claims, Kognita ROADMAP: 0.6 'Claims' in the Kognita repository.