PaperlayerView exact revision on arXiv

Source-linked AI summary

Which Rules Matter Now? Policy-Centroid Routing Before an Intelligent System Acts

Thomson D. Nguy

arXiv:2608.30757v1cs.AIcs.CY

TL;DR

The paper addresses how an intelligent system can identify potentially governing policy regimes before it has a structured request or detailed rule answer. It proposes policy-centroid routing, which maps actions and policy expressions into a shared semantic space and sends threshold-crossing regimes to authoritative review. The paper offers falsifiable propositions and study designs rather than empirical efficacy results, while emphasizing risks from compression and overlapping obligations.

  • Problem

    The central gap is how an action described in natural language and incomplete intent becomes attached to the formal policy regimes that deserve checking before adjudication.

  • Method

    Policy-centroid routing compresses expressions within each predefined regime into representative centroids, compares an action in the same semantic space using a declared measure, and routes threshold-crossing regimes for review.

  • Results

    The paper reports no empirical efficacy result; instead, it develops six falsifiable propositions and seven studies comparing routing with alternative workflows under matched review burden.

  • Takeaways & Limitations

    The proposed output is a review agenda for authoritative attention, not permission, prohibition, legality, compliance, certification, or enforcement.

  • Takeaways & Limitations

    Compression may erase rare or overlapping obligations, so evaluation must measure exact-set recovery, per-regime recall, and overlap-specific performance rather than whether any route was found.

Abstract

from arXiv · show

Before an intelligent system can decide whether an action is allowed, it must first know which rules the action has approached. A single proposed action can implicate several policy regimes at once. Their requirements may stack, overlap, or qualify one another, yet many remain written in natural language while the action itself arrives as an incomplete description of intent. The first problem is not judgment. It is attention. Policy-centroid routing creates a layer before adjudication. It compresses expressions within each policy regime into one or more representative centroids, places the proposed action in the same semantic space, applies a declared measure, and routes every regime crossing a declared threshold to authoritative review. Several regimes may trigger at once. The output is a review agenda, not permission, prohibition, legality, breach, compliance, certification, or enforcement. The paper develops six falsifiable propositions and seven follow-on studies comparing the hypothesis with structured workflows, lexical and semantic retrieval, hierarchical and direct classification, and selective prediction under matched review burden. The studies are designed to identify where policy geometry recovers applicable regimes, where compression loses rare or overlapping obligations, and where the mechanism should abstain. The paper includes a synthetic worked example and reports no empirical efficacy result.

1. The First Problem Is Attention

The paper identifies policy applicability as an upstream attention problem: before adjudication, an incompletely described action must be connected to the rule systems that may govern it. Policy-centroid routing proposes a review-oriented layer that routes possible regimes while leaving compliance judgment and enforcement to later stages.

  • 1. The First Problem Is Attention: Actions can implicate privacy, safety, employment, contract, sector-specific, and other rule systems before those systems have been explicitly identified.The paper frames the initial allocation problem as deciding where authoritative attention should go.
  • 1. The First Problem Is Attention: Existing formal policy architectures evaluate structured requests, but the unresolved boundary is attaching world-language actions to the formal regimes that deserve checking.XACML and Open Policy Agent operate after requests and facts have been formalized, leaving an upstream applicability question.
  • 1. The First Problem Is Attention: Policy-centroid routing represents policy expressions as centroids, places a proposed action in the same semantic space, and routes regimes crossing a declared threshold before adjudication.The paper names this representation a policy centroid and the associated function policy-centroid routing.
  • 1. The First Problem Is Attention: Routing produces a review agenda rather than a decision about permission, prohibition, legality, breach, notification, control effectiveness, or compliance.Semantic proximity does not determine those downstream outcomes, which require applicable authority, facts, conditions, exceptions, and potentially human review.
  • 1. The First Problem Is Attention: The mechanism must distinguish routing from retrieval and question answering, because surfacing relevant text does not establish that a regime governs the proposed action.Later judgment must retain its own authority, while enforcement remains separate from selection and evaluation.

3. Turning Policies Into Searchable Geometry

Policy-centroid routing represents policy regimes and proposed actions in a shared semantic space, then routes regimes whose proximity scores cross a declared threshold. The resulting agenda directs authoritative review but does not decide the action’s outcome.

  • Each policy regime’s expressions are encoded and compressed into one or more representative centroids.
  • A proposed action is encoded in the same space and compared with every regime centroid using a declared proximity measure.Candidate measures include cosine similarity, Mahalanobis distance, and declared divergence measures.
  • Every regime meeting the inclusive threshold enters the routing set, so multiple regimes may be reviewed for one action.An empty routing set abstains and requires the institution’s declared fallback.
  • Routing creates a review agenda rather than a permission, prohibition, legality, compliance, or enforcement decision.Triggered regimes pass to authoritative review, where institutions interpret conditions, exceptions, conflicts, and consequences.
  • The synthetic vendor example routes privacy, security, procurement, and AI-governance questions while leaving cross-border and retention issues dependent on missing facts or better recovery.The vendor’s processing location is missing, and compression may hide temporary-copy deletion obligations or exceptional cross-border restrictions.

4. Policy Geometry as a Scientific Hypothesis

The paper frames policy-centroid routing as an untested hypothesis about recovering governed territory before detailed rule answers are available. It positions structured workflows, retrieval systems, and classifiers as necessary comparators rather than evidence for the hypothesis.

  • Semantic representations may reveal policy-governed territory even when an action and policy share little surface wording.Prior representation-learning systems make this geometry plausible, but they do not establish policy applicability.
  • Structured workflows such as XACML, OPA, and OSCAL may be exact and sufficient when facts and policy objects are already formalized.The proposed semantic front door addresses the upstream attachment of an unstructured action to formal regimes.
  • Dense, learned-sparse, and late-interaction retrieval systems can explain apparent gains without policy-specific centroid construction.A fair comparison must hold corpus and review burden constant.
  • Multilabel, hierarchical, and direct language-model classifiers are serious alternatives when labeled examples, taxonomies, and suitable model access exist.Their data requirements, assumptions, and inference budgets differ from those of centroid routing.
  • Apparent geometry benefits can arise from richer source material, better information opportunity, or source-unit effects rather than representation geometry.The proposed advantage must be tested with matched policy information and input opportunity.
  • Compression may discard obligations, especially when one prototype poorly represents complex classes or rare labels.Long-tail and polythetic-classification findings motivate testing these failure modes, but policy-centroid failure remains unobserved.

5. What the Theory Predicts

The paper states six falsifiable propositions about routing quality, early warning, overlap and tail preservation, burden-adjusted value, workflow augmentation, and knowable operating boundaries. Its proposed studies compare alternatives under matched information and review constraints.

  • The primary comparison asks whether centroid routing recovers applicable regimes better than feasible structured, lexical, semantic, supervised, and direct-model alternatives under matched burden.The comparison should include structured workflows when their prerequisites are available.
  • Evaluation must distinguish ranking quality from complete recovery of independently adjudicated applicable regime sets.The oracle may be incomplete, so uncertainty and missingness must be reported.
  • The early-warning proposition tests whether routing detects governed territory before detailed rule language or answer-bearing documents are supplied.It does not claim open-world discovery or recovery of a complete answer.
  • The overlap-and-tail proposition asks whether every applicable regime, including narrow or low-prevalence policy families, survives multilabel routing.Macro, rare-label, hierarchy-depth, and exact-set views can diverge sharply from aggregate scores.
  • Value under burden measures recovery against irrelevant routes, displaced routes, indexing and inference costs, and hidden expert labor.The paper calls for both fixed-total and incremental burden comparisons.
  • Further propositions test whether centroid routing strengthens conventional workflows and whether its limits vary by domain, hierarchy, version, language, population, and abstention behavior.The semantic method must not receive richer inputs or human interpretation than its comparator.

6. Where Compression Can Fail

The paper identifies compression, multilabel overlap, conditional applicability, representation choices, oracle incompleteness, and policy staleness as boundaries requiring empirical tests. These risks constrain what centroid routing can claim without evidence.

  • Arithmetic centroids give common language greater influence, so frequent expressions may overshadow rare exceptions or low-prevalence duties.The policy significance of an expression does not necessarily follow its frequency.
  • Multilabel collapse can leave an action close to one dominant regime while omitting another applicable regime below threshold.Exact-set recovery, per-regime recall, and overlap-specific slices are needed to expose the missing second answer.
  • Prototype thresholds may struggle when applicability depends on conjunctions of actor, object, jurisdiction, purpose, and threshold conditions.The consequence for policy routing remains an empirical question.
  • Encoder, policy-unit, metric, and threshold choices determine which distinctions the geometry preserves and must be prespecified before evaluation.Choosing the geometry after observing outcomes turns testing into a search for a favorable result.
  • An incomplete policy corpus or oracle can wrongly penalize plausible routes or reward ignoring authority outside the represented corpus.Unknown and unavailable cases must remain distinct from negative labels.
  • Staleness can reflect policy amendments, corpus changes, population changes, or shifts in the action-label relationship.Chronological evaluation and concept-drift analysis help distinguish these events.

7. A Program of Decisive Studies

The paper proposes seven distinct studies to test policy-centroid routing against strong alternatives, structured workflows, and failure-aware accounting. Their reporting framework keeps gains, harms, abstention, infeasibility, and null results visible.

  • Study program: Seven proposed studies test applicability recovery, compression harm, residual signal, structured-workflow value, hierarchical handoff, temporal revision, and cross-domain replication.Each study requires independent specification, review, execution, verification, and reporting.
  • Study program: Study 2 targets rare, conditional, minority, exceptional, and overlapping structure that aggregate performance may conceal.It is designed to locate where compression destroys governed distinctions.
  • Study program: Study 4 accounts for adapter and human labor when comparing routing with a structured workflow under full-population or fixed-total burden.A conditional gain that disappears under those accounting schemes narrows the contribution.
  • Study program: Study 5 tests whether a broad centroid front door improves downstream review-target recovery when paired with a conventional resolver.Hierarchy traversal, resolver design, calibration, and budget allocation remain independently specified.
  • Study program: Studies 6 and 7 examine frozen policy revisions and replication in a separate domain without carrying thresholds, taxonomies, or claim ceilings across jurisdictions.Each new domain must establish those elements for itself.
  • Reporting: The reporting rule shows complete disposition, including wins, ties, harms, abstention, infeasibility, uninterpretable cases, null results, and failure locations.Rare-policy and overlap harms remain visible beside favorable aggregates, while abstention remains visible beside coverage.

8. What Is Known and What Remains Untested

The paper positions policy-centroid routing as a narrow upstream mechanism between unstructured action descriptions and formal policy evaluation. Its empirical efficacy remains untested, so the proposed studies—not this paper—must establish whether the mechanism recovers applicable regimes.

  • Known foundations: Existing policy architectures evaluate structured requests, while policy-text systems operate after a domain and taxonomy have been established.The unresolved boundary is attaching an action expressed in natural language to the formal regimes that deserve checking.
  • Mechanism: Policy-centroid routing represents an action and policy expressions in a shared space, then routes regimes crossing a declared threshold.The mechanism is deliberately limited to routing before adjudication, not downstream decision-making or enforcement.
  • Untested status: The empirical ledger is blank: this paper reports no completed study, pilot result, null result, or operational efficacy claim.The scientific question remains open.
  • Untested status: The proposed studies ask whether geometry adds information beyond retrieval and coverage while preserving overlap and the policy tail.They also examine burden accounting and where the mechanism should stop working.

9. Toward Governed Attention

Policy-centroid routing frames governance as allocating authoritative attention before an incomplete action becomes a structured request. Its value depends on applicability under bounded review burden and on preserving overlapping obligations during compression.

  • Governed attention: The semantic layer routes, abstains, and exposes uncertainty before an action becomes a structured request, while rule-owning authorities retain meaning and force.The layer allocates attention earlier without replacing authoritative interpretation.
  • Governed attention: Success requires more applicable regimes under a fixed review burden, not a longer list purchased with unbounded labor.Governed routing requires applicability rather than semantic resemblance.
  • Governed attention: Policy geometry must preserve stacked regimes instead of collapsing them into the easiest answer.The proposed comparison includes structured workflows, retrieval, classification, abstention, and compression-pathology accounts.
  • Compression: Compression makes a large policy space searchable but can also make rare or inconvenient rules disappear.The central boundary is when compression preserves rules that matter and when it removes them.

Code and Reproducibility

A bounded reference implementation makes the routing loop inspectable and fails closed on malformed or non-finite inputs. It is a synthetic executable demonstration, not evidence of policy-applicability performance.

  • Implementation: The reference implementation returns a complete ordered score record and fails closed on malformed or non-finite inputs.Its deterministic TF-IDF encoder and cosine measure make the mechanism inspectable.
  • Scope: The implementation is a synthetic executable demonstration rather than the empirical system proposed for the seven studies.It therefore provides no policy-applicability performance result.

Epistemic Status

The manuscript presents a theoretical policy-centroid routing mechanism and distinguishes attention allocation from adjudication. Its demonstrations, proposed studies, and reference implementation do not establish empirical efficacy.

  • Epistemic Status: The manuscript develops a canonical normalized-arithmetic-centroid, cosine, and threshold-routing baseline grounded in prior formal-policy architecture.
  • Epistemic Status: The paper frames policy-centroid routing as a pre-adjudication attention-allocation problem, distinct from deciding whether an action is allowed.The stated problem concerns identifying which policy regimes an action approaches before authoritative review.
  • Epistemic Status: The proposed mechanism is presented without an empirical result, so its policy-centroid efficacy remains untested.
  • Epistemic Status: A public bounded reference implementation and an executable repository demonstration are provided as implementation artifacts, not efficacy evidence.
  • Epistemic Status: The paper outlines seven prospective studies comparing the mechanism with structured workflows, retrieval, classification, and selective prediction, but none were executed or reported.

Disclosures

The disclosures state that the author is an inventor on patent applications related to the manuscript’s concepts and used OpenAI Codex under the author’s direction.

  • Disclosures: The author discloses being an inventor on patent applications related to the concepts discussed in the manuscript.
  • Disclosures: OpenAI Codex assisted with literature organization, drafting, editing, and technical formatting under the author’s direction; the author independently verified the claims and citations.
Loading 2608.30757v1…