ifc-0192

14.9 Ablations and interaction effects

The base ablation is the full \(2^3\) factorial over CLIC, OPTIC, and RELIC, augmented by controls for the interfaces themselves:

  7. no structural creativity and a matched flat learner;

  8. each component alone;

  9. the three pairwise compositions;

  10. all three components with a fixed workflow order;

  11. all three with learned guarded routing;

  12. the full system without cross-workflow certificate checks;

  13. oracle structure, oracle skills, and oracle regime information as separate ceilings.

These alternatives constitute the registered meta-level control suite \(\mathsf{Ctl}_{\mathrm{AG}}\). It includes the no-change and fixed-order workflows, each component and pairwise composition, a capacity- and budget-matched flat learner, interface-only and certificate-free variants, matched-random acquisition, and the separate oracle ceilings. A nonzero cross-workflow obstruction may route evidence or select among registered graphs, but it licenses \(\Upsilon _{\mathrm{AG}}\) only after the applicable controls fail and a registered representability test shows that no graph in the current workflow grammar realizes the target. Otherwise the correct output is an in-grammar construction, an evidence request, or abstention.

Let \(V(S)\) be preregistered value for the system containing component set \(S\subseteq \{ C,O,R\} \). A descriptive three-way interaction is
\begin{align*} \Delta _{COR} & =V(COR)-V(CO)-V(CR)-V(OR)\\ & \quad +V(C)+V(O)+V(R)-V(\varnothing ). \end{align*}
A positive \(\Delta _{COR}\) suggests complementarity not explained by lower order combinations; a negative value reveals interference. It is not by itself evidence of creativity. The structural record must still show what was constructed beyond each frozen reference space.

Reports should separate task reward from causal coverage, skill construction and execution validity, decision adaptation, false accommodation, abstention, sample and compute cost, catastrophic regression, transport, and persistence. Order sensitivity should be reported directly rather than averaged away.

Experiment: META-DIAL–0. Epistemic status: proposed factorial benchmark; the controlled demonstration reported next tests only a restricted precursor of this specification.
World: finite partially observed tool-use systems with a withheld causal regime, executable probe macro, and decision schema.
Question: can CLIC, OPTIC, and RELIC construct and admit a joint extension that none can complete alone?
Controls: full factorial component ablation, fixed versus guarded workflow order, interface-certificate ablation, flat learner, and three oracle ceilings.
Primary outcomes: joint task success and held-out transport, subject to zero or preregistered-low false accommodation.
Structural outcomes: causal-extension recovery, new-skill validity, new decision-schema validity, interface compatibility, and three-way interaction \(\Delta _{COR}\).
Admission boundary: improved reward without a versioned construction and independent component and interface certificates is optimization, not an admitted creative episode.