Empirically test policy-centroid routing efficacy

Establish whether policy-centroid routing can improve recovery of applicable policy regimes compared with strong feasible alternatives when policy information, input facts, model opportunity, human labor, and review burden are matched.

Background

The paper proposes representing policy regimes with one or more semantic centroids and routing an unstructured proposed action to every regime whose similarity score exceeds a declared threshold. The principal empirical issue is whether this representation provides better applicability recovery than structured workflows, lexical retrieval, learned-sparse retrieval, dense retrieval, multi-vector retrieval, supervised classification, or direct-model alternatives.

The comparison must use matched policy information, facts, model access, human labor, and review burden, and must evaluate both ranking behavior and recovery of independently adjudicated applicable sets. The paper reports no completed experiment resolving this question.

References

Proposition 1: comparative routing under matched burden. Can a declared centroid construction recover applicable policy regimes better than the strongest feasible alternatives when policy information, input facts, model opportunity, human labor, and review burden are matched?

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts  (2608.30757 - Nguy, 31 Aug 2026) in Section 5, Proposition 1: comparative routing under matched burden

Proposition 3: preservation of overlap and the tail. Can the route preserve every applicable regime when actions are multilabel, including narrow or low-prevalence policy families?

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts  (2608.30757 - Nguy, 31 Aug 2026) in Section 5, Proposition 3: preservation of overlap and the tail

Proposition 4: value under burden. Does any recovery advantage survive the cost of irrelevant routes, displaced correct routes, indexing and inference, and hidden expert labor?

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts  (2608.30757 - Nguy, 31 Aug 2026) in Section 5, Proposition 4: value under burden

Proposition 5: augmentation without a strawman. Can a centroid front door strengthen a serious conventional workflow?

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts  (2608.30757 - Nguy, 31 Aug 2026) in Section 5, Proposition 5: augmentation without a strawman

Proposition 6: knowable boundaries. Can the method's limits be described by domain, hierarchy layer, policy version, language, action population, and abstention behavior?

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts  (2608.30757 - Nguy, 31 Aug 2026) in Section 5, Proposition 6: knowable boundaries

Prototype classification and dense, sparse-expansion, and late-interaction retrieval show that learned representations can organize other tasks. They supply plausibility for the geometry, not evidence of policy applicability. That proposition remains untested.

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts  (2608.30757 - Nguy, 31 Aug 2026) in Section 4, Policy Geometry as a Scientific Hypothesis

The empirical ledger for this mechanism is still blank. This paper contains no completed study, pilot result, null result, or operational efficacy claim. The scientific question remains open.

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts  (2608.30757 - Nguy, 31 Aug 2026) in Section 8, What Is Known and What Remains Untested