12.7 The open construction problem

The unresolved test is construction rather than routing alone.

1. Withheld macro. Supply primitive skills but withhold a rollback-aware compare–repair–confirm macro required by some worlds. Test whether OPTIC constructs the macro after persistent bracket interference.

2. Withheld interface. Give two individually valid skills whose types do not compose. Permit a bounded mediator object and require the system to infer its interface from execution counter-witnesses.

3. Order and sufficiency. Compare active bracket-based probes with random paired executions at equal cost, scoring equivalence classes of decision-sufficient probes rather than exact labels.

4. Transport. Freeze the constructed skill theory and test it on disjoint task families, tool versions, and workflow graphs. Measure retained old competence separately from new competence.

5. Baselines. Compare ordinary SkillOpt, fixed-geometry LASKO, flat prompt search, OPTIC without the mixed witness, OPTIC without explicit factorization, and the complete system under matched execution budgets.

The decisive outcome is not terminal reward alone. Reports should include construction accuracy, type validity, false extension, regression, execution cost, abstention, transfer, and persistence. A successful new macro should remain available as a versioned artifact whose future uses and failures can be audited.