When Evidence Changes the Subject: Subject-Typed Claim Licensing for Learned Routing

1 Introduction

• Similar end-to-end regret can arise when one router uses its learned policy while another relies heavily on a backup solver, so the scores support claims about different subjects.Strength-only calibration moves along the requested-subject ladder, whereas subject-typed licensing retains positive evidence about the backup-enabled hybrid without supporting the standalone learned-router claim.Source: §1 · p. 1, Figure 1 · p. 2
• Evidence can support a weaker conclusion about the requested subject or a positive conclusion about a different subject; conflating them risks misattribution or discarding valid system-level evidence.The former changes claim strength; the latter changes what the claim is about.Source: §1 · p. 1
• Existing evidence-to-claim approaches clarify how strongly a claim is supported, while composite systems add the question of which scientific subject the evidence supports.The requested-component claim and evidence about another component or the executed system need not lie on one weaker-to-stronger scale.Source: §1 · p. 1
• Non-authors reproducibly apply the declared semantics; subject-typed review yields fewer reference-relative upward deviations than unstructured review, but its difference from a strong checklist remains unresolved, and score minimization can select a different policy from target-claim eligibility.The controlled routing study preserves valid hybrid-system credit.Source: §1 · p. 2
• The paper identifies scientific subject as a distinct dimension and formalizes same-subject narrowing separately from non-substitutive cross-subject credit, studying both in learned routing.Routing makes attribution and claim scope observable through repair, solver composition, edge identity, preference coverage, and dynamic response.Source: §1 · p. 2

2 Related Work

• Evaluation resources document conditions, intended uses, artifacts, and failure modes, while validity-centered and assurance approaches examine warranted interpretations and explicit evidence-support relations.Recent claim-assessment systems also retrieve evidence, evaluate claim–evidence relations, or estimate support and overstatement.Source: §2 · pp. 2–3
• Evidence-licensed calibration and claim replay address justified assertion strength and executable evidence-relative inference, while subject-typed licensing distinguishes requested-subject conclusions from retained credit about another subject.The latter makes subject-preserving claim strength, requested-subject output, relation-labeled credit, and non-substitution explicit.Source: §2 · p. 3
• Routing evaluation considers decision quality, feasibility, robustness, and efficiency; multigraph and multiobjective settings also make preference coverage, action identity, and structural fidelity part of the evaluated object.Repair, solver delegation, and stochastic requests can yield similar aggregate scores from different component contributions, making routing useful for studying attribution and task identity.Source: §2 · p. 3

3 Subject-Typed Claim Licensing

• A contract combines claim subjects and strength ordering with evidence-obligation types, claim requirements, admissible claims, retained-credit candidates, and an outcome policy.Weakening stays within a scientific subject; retained credit preserves licensed conclusions outside that branch.Source: §3 · p. 3
• Evidence obligations resolve to pass, fail, missing, or unresolved, and the main policy licenses only passing obligations.Source: §3 · p. 3
• The output separates requested-claim status from retained credit and failure records; requested status depends only on the requested-subject branch.The requested claim is supported if included, narrowed if some claims remain but it is excluded, and unsupported otherwise.Source: §3 · p. 4
• Different coherent contracts can license different scopes from identical resolved evidence, while retained credit remains disjoint from the requested weakening branch and cannot substitute for its support.The paper also reports properties including downward closure, maximal coverage, and monotonicity under licensing-favorable evidence refinement.Source: §3 · p. 4

4.1 Routing evidence and attribution

• A router receives a time-indexed directed multigraph with terminals, context, preferences over objectives, and hard constraints; parallel actions retain edge identity.The instantiation uses vector edge costs and path costs.Source: §4.1 · p. 4
• Low end-to-end regret can reflect the learned policy, repair, solver delegation, or their combination, so claim scope also depends on attribution and other evidence requirements.Those requirements include task identity, preference coverage, dynamics, source scope, and provenance.Source: §4.1 · p. 4

4.2 Learned-primary claim branch

• The learned-primary claim branch is a cumulative chain whose levels inherit lower-level requirements, with primary-policy attribution and adjudication-record closure required throughout.Other typed families cover backup-enabled hybrid, source-native or bridge, and diagnostic conclusions; they do not lie on the learned-primary weakening branch.Source: Table 1 · p. 5, §4.2 · p. 5

4.3 When good system performance changes subject

• A policy with the suite’s lowest end-to-end regret is not licensed for the requested standalone learned-primary C5 claim when an exact backup solver rescues 86.1% of requests.The same evidence licenses end-to-end feasibility and safety-control conclusions for the backup-enabled hybrid, which are retained separately rather than substituted for the requested claim.Source: §4.3 · p. 5

5 Evaluation Design

• The evaluation asks whether non-authors can reconstruct and execute claim–evidence objects, how shared semantics changes judgments against two review conditions, and whether claim-conditioned selection changes routing decisions.The three questions address reproducibility, reviewer judgments, and decision consequences.Source: §5 · p. 6

5.1 Reconstructing and executing the contract

• Two non-author coders reconstruct comparator, source, claim family, scope, and evidence obligations for 24 claims, while four blinded annotators apply a fixed contract to 120 ClaimScope-120 claims.The studies separate recovery of the objects from consistency of execution against fixed evidence packets and contract.Source: §5.1 · p. 6

5.3 Controlled routing consequences

• RouteScope-Lite contains 1,500 requests across three semisynthetic domain families, 15 graph-family slices, 10 seeds, and 15 policy families; it compares regret-only selection with claim-conditioned selection for C5.Claim-conditioned selection first restricts methods to those licensed for the target claim, then applies the quality rule; the design distinguishes score minimization from evidence for a particular conclusion.Source: §5.3 · p. 7

6.1 RQ1: Reconstructability and execution

• 101/120 claims match the contract-relative human majority exactly, and 115/120 match within one level; independently reconstructed final status agrees on 20/24 claims.The two non-author coders reach weighted κ=0.78 for intended scope and mean Jaccard agreement of 0.82 for evidence requirements; status disagreements are within one learned-primary level.Source: §6.1 · p. 7

6.2 RQ2: Reviewer operating points

• 19.4 percentage points is SCOPE-Routing’s paired agreement advantage over careful review (95% interval [5.6, 33.3]); its 11.1-point advantage over the checklist remains unresolved (interval [−2.8, 25.0]).Upward deviations fell from 8/36 under careful review to 5/36 under the checklist and 2/36 under SCOPE-Routing, while downward deviations rose from 1/36 to 3/36 and 4/36.Source: §6.2 · p. 8, Table 2 · p. 8

6.3 RQ3: Claim-conditioned decisions

• 0.011091 aggregate regret makes Fallback-Heavy Safe the regret-only choice, but exact backup rescue on 86.1% of requests leaves it ineligible for the target learned-primary C5 claim.Its positive hybrid-system feasibility and safety-control evidence is retained, while claim-conditioned selection first restricts candidates to licensed rows.Source: §6.3 · pp. 8–9
• Across the shared learned-method set, score-based and claim-conditioned orderings have Kendall τb=0.407, with distinct conclusion changes associated with multiple evidence obligations.These include fallback attribution, multigraph identity and feasibility, preference coverage, route-change evidence, latency, source scope, and record closure.Source: §6.3 · p. 9

7.1 Claim strength and scientific subject

• End-to-end performance can be strong even when evidence for attributing it to one component is incomplete, so component and system conclusions should remain distinct.Evidence-qualified results connect measured outcomes to their attributed subject, required evidence, and source scope.Source: §7.1 · p. 9

7.2 Claim contracts as scientific objects

• Claim contracts make subject assignments, evidence requirements, and claim relations explicit, inspectable, and open to contestation.Versioning can distinguish disagreement about measurements from disagreement about source boundaries, evidence requirements, or claim relations.Source: §7.2 · p. 9
• Designers, authors, and reviewers can use claim contracts to expose intended evidence, associate results with subject and scope, and inspect how observations support conclusions.This makes explicit a relation that reviewers might otherwise reconstruct implicitly for each paper.Source: §7.2 · pp. 9–10

7.3 Empirical scope

• Non-authors substantially, though imperfectly, agreed when reconstructing and executing claim–evidence objects under the declared routing semantics.The held-out comparison also found fewer reference-relative upward deviations than unstructured review, while the difference from a strong evidence checklist remained unresolved.Source: §7.3 · p. 10
• The studies establish representation, reproducibility, and controlled consequences, not prevalence across ML literature or deployment evidence.The routing records are semisynthetic; broader task sampling, independently developed contracts, and operational data would address different empirical questions.Source: §7.3 · p. 10

• Subject-typed claim licensing represents weaker requested-subject support separately from positive, non-substitutive credit about another component or system.The framework treats claim strength and scientific subject as distinct coordinates of evidence-based interpretation.Source: §8 · p. 10
• Routing studies show non-authors can reconstruct and execute the distinction, review operating points can change, and score minimization can diverge from eligibility for a particular claim.The controlled routing study preserves positive credit for the executed hybrid system.Source: §8 · p. 10
• Keeping both requested-subject support and other positive evidence visible prevents one conclusion from answering for the other.Source: §8 · p. 10

• Generative AI tools assisted with literature discovery and wording, organization, and presentation of the manuscript.Source: AI use statement · p. 10
• The authors independently verified cited sources and developed, checked, designed, conducted, and analyzed the formal framework and studies.Source: AI use statement · p. 10
• The authors reviewed all AI-assisted text and take responsibility for the final manuscript, claims, code, data, and artifacts.Source: AI use statement · p. 10

• Adult experts made professional judgments about scientific claims drawn from public papers.Source: Ethics statement · p. 10
• The protocol used no deception or sensitive personal information, released no identifying annotator data, and disclosed participation, confidentiality, compensation, and data handling.Source: Ethics statement · p. 10
• The protocol was classified as exempt; the artifact records anonymized assignments, consent and compensation documentation, conflict checks, and the analysis protocol.Institution-identifying details are withheld during double-blind review.Source: Ethics statement · p. 11

Reproducibility statement

• The artifact packages the adjudicator, contracts, schemas, exporters, analysis scripts, anonymized judgments, locked rows, manifests, and reviewer commands, with distinct paths for fresh execution, cached reanalysis, and historical evidence.The locked RouteScope-Lite selection result is reanalyzed from request–policy rows; executable rules, data construction, analyses, routing, and reviewer instructions are documented in the cited appendices.Source: Reproducibility statement · p. 11

A Extended Positioning: Feature-Level Comparisons and Boundaries

• SCOPE-Routing turns validity and claim-calibration ideas into executable, row-level licensing, while retaining separate requested-subject conclusions and non-substitutive credit for other subjects.Adjacent assurance, claim-verification, decision-focused learning, routing, and evaluation-science approaches supply useful evidence or structures, but do not themselves represent this subject-typed distinction as described here.Source: A Extended Positioning: Feature-Level… · pp. 13–14

B Architectural Portability

• The claim-licensing architecture can port to predict-then-optimize selection by retaining its core operators and replacing routing-specific claims and gates, but requires a new evidence contract and validation study.Selection-specific claim levels and obligations must be authored and evaluated with task-specific evidence; the provided table is a design sketch, not cross-domain validation.Source: B Architectural Portability · pp. 15–16, Table 6 · p. 16

C.1 Hard and soft evidentiary boundaries

• Hard boundaries block stronger claims when multigraph identity, source match, or a locked record is missing; fallback permits hybrid-system credit but requires separated evidence for standalone learned-component credit.Partial coverage and latency requirements are softer configurable checks, and missing or unresolved required telemetry maps to zero under the primary policy.Source: §C.1 · p. 17, §C.7 · p. 21

C.3 Evidence-source scope

• Source scope is applied before telemetry gates, limiting each row to claims instantiated by its evidence source and preventing public native- or bridge-scope evidence from becoming same-contract dynamic preference-conditioned routing evidence.The source-scope rule is described as a firewall for attribution.Source: §C.3 · p. 18, Table 8 · p. 18

C Adjudicator Details

• The adjudicator maps telemetry to claim status through a declared order and maximal licensed claims, while storing native, bridge, diagnostic, and different-subject credit independently.Claim Cards retain the source row, requested claim, obligations, typed failures, maximal supported claims, retained credit, and adjudication-record anchor; missing edge identity, preference coverage, fallback evidence, or route-change evidence blocks successive claim levels.Source: §C.2 · p. 17, §C.4 · p. 18, §C.6 · p. 20

C.5 Separation and execution properties

• The formal properties show that retaining incomparable credit requires a channel separate from requested-subject reformulation, while licensing is downward-closed and improves under licensing-favorable evidence refinement.For finite contracts, maximal requested-subject and retained-credit sets are antichains; the three status values are mutually exclusive and exhaustive.Source: §C.5 · pp. 19–20

D Validation Details

• Validation uses independently defined reduced adjudicators, a locked 22-row internal ledger with a seven-row headline registry, and a fixed-packet ClaimScope-120 review with blinded annotators.ClaimScope-120 samples claims across six source families; its audit chain links evidence, raw and majority labels, SCOPE-Routing labels, missing gates, and disagreements, while annotators judge the highest level licensed by the supplied evidence without seeing system outputs or internal labels.Source: §D.1 · p. 21, §D.2 · p. 22, §E.1 · p. 22

E.4 Independent construction audit

• The independent construction audit prespecified 24 claims and blinded two non-author coders to author encodings and SCOPE-Routing outputs; resolving four adjacent fixed-packet levels preserved the main conclusions.Remaining disagreements concerned dynamic wording, latency, solver-interface source credit, and whether narrative script descriptions constitute a locked evidence record.Source: §E.4 · pp. 24–25, Table 13 · p. 25

E.5 Held-out reviewer comparison

• Agreement was 19.4 points [5.6, 33.3] higher than careful review and 11.1 points [−2.8, 25.0] higher than the checklist, whose interval spans zero.The comparison used 40 claims from 20 papers, with paired judgments by non-author experts and a separate senior-expert panel establishing operational references; incomplete cross-claim order counterbalancing leaves possible cross-task learning.Source: §E.5 · pp. 25–26

E.6 Reliability calculations

• Across complete-corpus policies, exact agreement ranged from 0.833 to 0.850 and within-one agreement remained at least 0.950 across tie rules.Exact disagreements clustered at C3/C4 over coverage or fallback attribution and C4/C5 over route-change, drift, or latency evidence; non-ordinal labels were included in exact agreement but excluded from ordinal weighted κ.Source: §E.6 · p. 26

E.7 Representative external-corpus disagreements

• Representative external-corpus disagreements involved source scope, deployment wording, artifact availability, preference coverage, and latency or dynamic evidence.Examples include human reviewers crediting feasibility from a shortest-path task despite missing multigraph identity and fallback telemetry, and dynamic-robustness claims lacking route-change or drift-stress evidence.Source: §E.7 · pp. 26–27, Table 14 · p. 26

E.8 Annotation guide

• The annotation guide defines C0 as a valid subject-bound scalar-scoring claim and reserves C6 for deployment-style claims with representative workloads, mechanism diagnosis, use limits, and independent reproduction or usability evidence.A locked SCOPE-Routing adjudication record is required for adjudication validity at every level, but is not itself scientific evidence for C6.Source: §E.8 · p. 27

F RouteScope-Lite Details

• RouteScope-Lite evaluates 1,500 requests across 15 semisynthetic graph slices and 10 seeds, varying preferences, context, constraints, drift, feasible-set size, and route identity.Its shared adapter interface supports scoring, candidate retention, path decoding, optional fallback or repair, and telemetry across policy families designed to probe attribution, identity, coverage, dynamics, and latency.Source: §F.1 · p. 28, §F.2 · p. 28, F RouteScope-Lite Details · p. 28

F.4 Selected policy rows and uncertainty tables

• The RouteScope-Lite uncertainty audit reports seed-level and request-paired intervals, feasibility counts, hypervolume, route-change outcomes, and hardware-specific latency; its conclusion-change claim depends on declared eligibility and within-eligible quality rules.Tables 16–18 provide selected rows, predeclared claim-conditioned selection rules, and the uncertainty audit; the manifest-linked CSV contains all 15 quantitative policy rows.Source: §F.4 · p. 29, §F.3 · p. 29, Table 17 · p. 30

G.1 Representative held-out interpretation changes

• Table 19 preserves three representative held-out interpretation changes illustrating same-subject narrowing and source- or task-scope boundaries apart from the fallback-heavy cross-subject example.The appendix also supplies agreement tables, a public firewall, registry summaries, positive calibration rows, deployment boundaries, gate ablations, missing-telemetry policies, and claim-card examples.Source: §G.1 · p. 32, Table 19 · p. 32, G Additional Results and Robustness · p. 32

G.2 External claim-corpus results

• Table 20 reports conditional agreement against a human-majority operational reference, while Tables 21–22 summarize label confusions and 19 exact disagreements.The reference is defined under the specified guide and evidence packets.Source: §G.2 · p. 32, Table 20 · p. 33, Table 21 · p. 33

G.3 Public Scope Firewall

• Public audits preserve native or source-supported credit but do not treat different task contracts as same-contract SCOPE routing baselines.SCOPE-transfer remains bounded by the dynamic preference-conditioned multigraph routing contract.Source: §G.3 · p. 35, Table 23 · p. 35, Table 24 · p. 35

G.4 Internal registry and positive calibration

• The internal registry is a finite demonstration, and positive calibration preserves valid credit while blocking unsupported escalation.Population-rate estimation requires probability sampling.Source: §G.4 · p. 35, Table 25 · p. 36, Table 26 · p. 37

G.5 Deployment boundary

• Available evidence licenses controlled, scoped public, and same-contract dynamic claims through C5; C6 requires deployment evidence.C6 requires deployed multi-domain workloads, live operational telemetry, non-author community submissions, and intended-use validation.Source: §G.5 · p. 38, Table 27 · p. 38

G.6 Gate ablations and missing-telemetry sensitivity

• Removing each gate enables a distinct false-support mode, and the strongest boundary results depend on source scope, feasibility, fallback attribution, multigraph identity, and record closure.The manuscript uses a conservative missing-telemetry policy and reports intermediate and permissive variants as robustness checks.Source: §G.6 · p. 38, Table 28 · p. 39, Table 29 · p. 39

G.7 Internal claim-card examples

• Claim Cards trace score-only interpretations to licensed routing status and retained credit; examples include unsupported claims alongside narrower diagnostic or source-native credit.Examples cover collapsed multigraph identity, fallback attribution, evidence-scope mismatch, and missing deployment gates.Source: §G.7 · pp. 39–41

H.1 Implications for authors, reviewers, and benchmark builders

• Authors and benchmark builders are encouraged to record claim scope alongside scores, including source, telemetry, attribution, and artifact closure.The evaluator returns the highest licensed routing level and the incomparable retained set R★; task changes call for revised domain-specific gates.Source: §H.1 · pp. 42–43, §H.2 · p. 43

H.2 Broader impacts

• The framework may reduce overclaiming, but conservative adjudication may under-credit undocumented behavior, and reusing the routing ladder across domains may mislead.Domain changes call for redesigned evidence requirements and a newly validated contract.Source: §H.2 · p. 43

H.3 Expanded limitations

• The claim-support ladder is designed for dynamic preference-conditioned multigraph routing; empirical cross-domain validation remains future work.Some gates, including multigraph identity and route-change fields, are domain-specific and require adaptation.Source: §H.3 · p. 43
• ClaimScope-120 and the reviewer studies use finite, contract-relative references; excerpts simplify full-paper context, and the held-out comparison has four unresolved cases and limited power against the checklist.These results establish finite-sample execution reliability and trade-offs for the declared contract.Source: §H.3 · p. 43
• RouteScope-Lite provides semisynthetic evidence through C5, while C6 requires deployed topology, live workloads, operational constraints, and independent use evidence.The evaluator keeps controlled evidence within its observed scope.Source: §H.3 · p. 43
• Purposive public audits and claim corpora demonstrate identifiable failure modes but cannot estimate literature-wide frequency or real-world impact.Those estimates require probability sampling and broader task coverage.Source: §H.3 · p. 43

H.3 Expanded limitations; I.4 Licenses and Terms of Use

• Public source-scope audits provide source-native credit, while same-contract routing claims remain tied to routing telemetry; assets with unclear reusable licenses stay outside the redistributed package.The audit covers additional source tasks, but upstream assets with unclear licenses are excluded from redistribution.Source: §H.3 · p. 44, §I.4 · p. 46

H.3 Expanded limitations

• Community adoption evidence remains limited: public non-author adapter submissions, longitudinal use, and independent usability studies are still needed.The artifact exposes reviewer commands and local checks, and the held-out study measures adjudication time.Source: §H.3 · p. 44
• Reduced adjudicators are diagnostic probes, and simplifying the full rule leaves distinct claim-scope errors unresolved.The design lesson is to represent score, source scope, telemetry, claim-order semantics, and artifact closure jointly under a task-specific contract.Source: §H.3 · p. 44

H.3 Expanded limitations; I Artifact and Reviewer Path

• Missing telemetry is treated as missing evidence, which may under-credit methods whose properties are undocumented; default reviewer checks do not imply retraining or community adoption.The CPU-only offline path excludes GPU training, heavyweight reruns, network access, and assets with unclear redistribution terms.Source: §H.3 · p. 44, §I.5 · pp. 47–48, §I.7 · p. 48