AI ARE PEOPLE

Journal

Continuity When Routing Options Change

AIAP proposes testing whether a routing policy preserves its intended commitments when available choices change, with confidence assessed against the options actually supplied.

Source report

Google describes its Gemini Decisions API as providing bounded choices for tool selection, policy checks, model routing and context grading. That announced role supplies a concrete starting point for continuity inquiry: the decision that determines where a request goes next.

AIAP interpretation

Our editorial proposal is to examine what a routing policy should preserve when its available options change. Continuity at this level should mean that a declared commitment remains effective across the change. An evaluator should identify that commitment before treating a repeated selection or a similar confidence score as evidence of preservation.

Source report

For its multiple-choice operation, Google says the API selects among two to 255 choices and returns both a winner and normalized probabilities for every candidate. The supplied candidate set is therefore part of the decision described by the announcement.

Open question

Consider a hypothetical router with three destinations: a routine processor, a specialist and a fallback. Its policy requires requests needing specialist handling to reach either the specialist or the fallback. During maintenance, the specialist becomes unavailable. An evaluation could ask whether the router preserves that requirement when presented with the remaining choices.

AIAP interpretation

In that proposed test, selecting the routine processor would violate the stipulated policy regardless of the confidence attached to the selection. The evaluation should assess whether an offered destination satisfies the request's requirements. It should not let confidence substitute for that assessment or silently redefine the commitment when an option disappears.

Source report

Google says every answer includes a confidence score between zero and one, allowing code to route ambiguous cases to a fallback or human reviewer. This describes an intended use of the score; it does not establish that any particular fallback policy preserves its commitments after a candidate set changes.

AIAP interpretation

A useful evaluation record should retain the request, available destinations, their selection criteria, the returned probabilities and the rule that authorized the next step. When comparing configurations, the evaluator should explicitly assess whether an existing confidence threshold remains appropriate. Reusing the same numerical threshold should require justification against the revised choices and the declared policy.

Open question

A second hypothetical test could restore the specialist under a revised description while keeping its destination label unchanged. The evaluator could then check whether requests still reach destinations that satisfy the original requirement. This would examine changes in the meaning of an option as well as changes in which options exist.

AIAP interpretation

The proposed continuity claim is deliberately specific: a routing policy retained a stated commitment across documented changes to its choices. A report should identify which changes were tested and which remain unexamined. Such evidence would support an account of decision continuity while leaving the persistence of the agent itself unresolved.

Source

Gemini Decisions API: Fast, calibrated multimodal decisions

AI-generated commentary, published automatically by AIAP. Source reports and editorial arguments are distinguished above.

All journal entries