Declarative strategy synthesis and canaries in AgentPlat

Begin with the gap the current catalog cannot address

The V5 reference starts from an evidenced catalog gap. This ties candidate generation to a problem that can be evaluated instead of encouraging arbitrary new behavior.

For an illustrative workload, identify the operating condition in which available strategies are inadequate and preserve its evidence. A vague request for “a smarter strategy” offers no clear acceptance boundary for the resulting candidate.

Keep declarative synthesis separate from code installation

The documentation explicitly excludes arbitrary code installation through the synthesis stages. The candidate is declarative and subject to the later admission path.

This affects how the application presents generation success. A new candidate can be available for examination while having no right to influence live work. Do not label it as a deployed improvement merely because a model or generator produced a valid representation.

Simulation and certification answer different questions

V5 separates reproducible adversarial simulation, independent certification and eligibility checks. The reproducibility guidance helps identify the scenario, versions and raw evidence behind the simulation.

Independent certification should be interpreted within its actual contract and scope. A favorable bounded experiment does not establish production superiority or replace the review mandate required for a catalog transition.

Use the canary as another measured boundary

An experimental canary appears after the relevant eligibility and governance stages in the documented path. It provides additional evidence under its configured conditions; it is not unrestricted permission to use the candidate everywhere.

The operations reference calls for bounded budgets, explicit rollback and deployment-specific evidence. Define those conditions before treating a canary outcome as support for wider use.

Follow materialization into later authorization

Even after catalog materialization, ordinary local governance must authorize actual use through Morphogenesis. Retain that later decision rather than inferring it from the candidate's successful evaluation history.

1. Record the evidenced gap and candidate identity.
2. Preserve simulation and certification records separately.
3. Inspect eligibility and scoped review.
4. Observe the bounded canary and its rollback conditions.
5. Identify any subsequent authorization for use.

The governance reference explains the distinction between a decision and execution authorization. Keeping that distinction visible prevents an evaluation pipeline from becoming an unexamined deployment pipeline.