11.5 Explicit alternative posteriors

DIAL-SKFM–1D makes the unsupported alternatives first-class inferential hypotheses without making them admissible repairs. Its posterior contains the four supported structures, the registered two-latent and noncentral alternatives, and an open “other unsupported” outcome. Unsupported posterior risk can trigger a probe even after an apparently confident supported decision. Admission is asymmetric: a previously abstained case needs positive joint confirmation before becoming a repair, whereas an initially admissible repair is retained unless the new evidence strongly contradicts it.

This change nearly eliminates false accommodation. Unsupported abstention is 99.8%, compared with 84.8% for the legacy posterior, while supported selective accuracy is 97.2% and the H0 false-extension rate remains zero. Active probing recovers substantial usefulness lost by the safer posterior: supported coverage rises from 41.5% with explicit alternatives alone to 65.1%, and overall accuracy rises from 60.7% to 75.4%. At the same 83.5% probe rate, random selection reaches only 58.3% coverage and 70.9% accuracy. The active selector therefore contributes information beyond merely gathering more samples.

The study still fails its overall-accuracy gate and does not open confirmation. Its principal error is no longer a false causal construction. Rather, the flat posterior treats the null closure hypothesis as just another template and rejects too many genuinely closed worlds: only 23.3% of H0 worlds are retained, although none is falsely extended. By contrast, noncentral worlds are rejected perfectly and two-latent worlds at 99.6%. This supports a hierarchical model: first perform a calibrated closure-versus-nonclosure decision, then conditionally infer a repair family. The null needs its own evidence-accumulation and stopping rule rather than symmetric competition with every nonzero structural template.

Experiment: DIAL-SKFM–1D. Posterior: four supported structures, two explicit unsupported alternatives, and an open unsupported outcome.
Admission: unsupported hypotheses can request evidence and veto a repair but can never themselves become repairs.
Result: 99.8% unsupported abstention, 97.2% supported selective accuracy, 65.1% supported coverage, and zero H0 false extensions; overall accuracy is 75.4%, so the registered decision is abstain.
Lesson: explicit alternatives nearly eliminate observed unsafe accommodation under this registered contract, but a flat posterior is excessively skeptical of genuine closure. Separate null detection from conditional structural diagnosis.