ifc-0151

10.12 How DIAL-X sharpens an experiment

Once the target card is fixed, the ablation is almost forced. A credible DIAL-X evaluation separates:

  1. fixed-declaration learning;

  2. one typed side;

  3. a capacity-matched but untyped two-operator learner;

  4. two typed sides without the mixed witness;

  5. mixed diagnosis without structural repair;

  6. local repair without a finite-realization observer;

  7. realized repair without an explicit package comparison;

  8. package comparison without independent admission; and

  9. the complete DIAL-X workflow.

Negative controls shuffle the mixed witness, corrupt one side, change the presentation, suppress a required observer, and supply an exact witness as a diagnostic ceiling. Resource matching includes interventions, validation samples, memory, and compute. Otherwise DIAL can appear to win merely because it receives a richer experiment or maintains more learner states.

A creativity experiment requires one further contrast: the same DIAL controller with a fixed proposal vocabulary versus an expandable one. The evaluation must report how often each system produces a genuinely new candidate relative to the frozen reference space, how far that candidate extends or transforms the space, and how often independent admission retains it. This separates the ability to detect disequilibrium from the ability to construct a valuable response to it.

Experiment: Minimum DIAL-X experimental record. Epistemic status: reporting template; the fields below are required for a DIAL-X claim but do not themselves constitute experimental evidence.
Target: publish \((Y_X,\mathbb S_X,\mathcal C_X)\).
Creativity: when claimed, publish \(\mathsf{Cr}_X\), freeze the reference space, label the Boden mode, and score novelty separately from value.
Sides: type \(H_X\) and \(V_X\), test each coherence law, and report its uncertainty separately.
Interaction: estimate \(\Theta _X\) only after both side gates pass; test localization and presentation invariance.
Integration: register the finite path, rollout, or persistence observer \(\gamma _X\); report local success separately from finite coherence.
Action: report assimilation, proto-accommodation proposals, admitted accommodations, interventions, and abstentions separately.
Proposal: include current and no-change controls; record the exact language from which repairs were drawn, the registered control suite, the package comparison for every declaration change, and its presentation map when applicable.
Admission: use evidence not consumed by proposal generation or ranking.
Persistence: measure transport, regression, and reuse rather than counting storage alone.

This discipline turns a vague application claim into a falsifiable algorithm. For DIAL-GIRL, it exposed the difference between recognizing an old regime and benefiting from an old state. For DIAL-SKFM, it separates detecting nonclosure from identifying a causal explanation. For DIAL-ALLORA, it prevents reduced order sensitivity from standing in for a useful adapter composite.