Bounded local strategy adaptation in AgentPlat

Measure the outcome before learning from it

The strategy intelligence reference lists mission progress, latency efficiency, resource efficiency, recovery quality and safety as measured dimensions. Callers cannot replace that evidence with an injected scalar reward.

For an illustrative strategy evaluation, faster completion may come with poorer recovery or more unresolved work. Preserve the dimensions instead of flattening the observation into a single favorable score that hides the tradeoff.

Keep each stage of adaptation visible

V3 separates measurement, local learning, proposal generation, counterfactual analysis, recommendation, review, catalog transition and execution. A recommendation is therefore neither a completed catalog change nor proof that the strategy ran.

Design the operator view around those stages. If a reviewer rejects the recommendation, the record should still explain what was suggested and why. If a catalog transition succeeds, retain the later execution evidence separately.

Govern changes without uncontrolled switching

The documented governance actions include promotion, degradation, retirement and rollback. It also identifies fail-closed conditions such as stale recommendations, low confidence, excessive switching and self-review.

The decision model supplies the scope and mandate requirements for the selected reviewer. A strategy's favorable history does not let it authorize its own promotion or escape the applicable independence rules.

Bind the chosen strategy to its implementation version

The strategy registry reference uses identifiers, versions and digests for coordination strategies. Retain those coordinates when a recommendation becomes a selected configuration.

Otherwise, an operator may know the strategy name without knowing which implementation produced the result. This is particularly important when comparing an earlier baseline with a changed allocation or team-formation policy.

Separate software conformance from observed improvement

A pipeline correctly refusing stale recommendations demonstrates that control behavior. It does not prove that a promoted strategy improves a real mission.

• Record the measured outcome dimensions.
• Inspect the recommendation and its currentness.
• Follow the independent review where policy requires it.
• Identify the exact catalog transition and selected version.
• Measure subsequent work under registered conditions.

Use the evaluation methodology for that last step. Keep the original observation and the later operational result as distinct evidence.