Policy for impossible per-application allocation

Determine what value and persistence behavior the Productivity Apportionment skill should use when enterprise data do not support a defensible per-application split of EC2 memory or cost.

Background

The evaluator expects zero for certain EC2 memory and cost allocations when the underlying tools provide host-level or otherwise insufficient data for a defensible application-level split. Observed executions sometimes instead allocate nonzero values, exposing an ambiguity in the skill specification rather than conclusively demonstrating that either behavior is correct.

Resolving this question requires domain-level specification of the intended semantics for unavailable or non-identifiable per-application allocations. The answer affects both the expected database values and the classification of agent trajectories as compliant or erroneous.

References

The tool can return host/JVM memory where no defensible per-application split exists, whereas the authored invariant expects zero. Writing the test exposed an implicit policy choice: \emph{what should be stored when a per-application split is impossible?}

— Continuous Process-Level Evaluation for Evolving Enterprise AI Agent Skills  (2610.01833 - Vo et al., 1 Oct 2026) in Appendix, Section "Limitations"