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

Jian Chen; Zixuan Yuan

cs.AISep 27, 2026

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

Authors: Jian Chen, Zixuan Yuan

Organizations: The Hong Kong University of Science and Technology (Guangzhou), China

Abstract

Modern learned systems increasingly combine learned components with search, repair, or external solvers. Benchmarks often measure the resulting end-to-end system, while scientific claims may concern only one component, creating an attribution problem: evidence can fail to support the requested component-level claim while still supporting a positive conclusion about the larger system. Existing evidence-to-claim methods primarily calibrate claim strength. We argue that composite systems require a second dimension: scientific subject. We address this problem with subject-typed claim licensing, which separates weaker conclusions about the requested subject from positive but non-substitutive credit about another subject. We instantiate this idea in SCOPE-Routing for preference-conditioned multigraph routing. Non-authors reproducibly apply the declared semantics; held-out review yields fewer reference-relative upward deviations than unstructured review, while the difference from a strong evidence checklist remains unresolved; and a controlled routing study shows that score-optimal and claim-eligible methods can differ while valid hybrid-system credit is preserved. These results motivate treating claim strength and scientific subject as distinct dimensions of evidence-based evaluation.

Figures & tables

Figure 1. When evidence changes the subject. A fallback-heavy router illustrates the difference between strength-only calibration and subject-typed licensing. Strength-only calibration can only move along the requested-subject ladder; subject-typed licensing additionally retains positive evidence about the backup-enabled hybrid without allowing that evidence to support the standalone learned-router claim. A three-panel comparison of a composite learned-routing example, a strength-only claim view, and subject-typed licensing with separate requested-subject support and retained hybrid-system credit.

Level	Licensed learned-primary conclusion	Incremental evidence
$C0$	Scalar decision quality	Primary-only or intervention-free regret against a declared comparator.
$C1$	Feasibility-audited routing	Exact feasibility and invalid-route accounting.
$C2$	Identity-preserving multigraph routing	Edge identity, bundle-aware actions, and collapse checks.
$C3$	Preference-conditioned routing	Preference sampling, coverage/hypervolume, and held-out preference shift.
$C4$	Intervention-audited constrained routing	Constraint masks, fallback/repair taxonomy, and primary versus end-to-end decomposition.
$C5$	Dynamic robust routing	Drift stress, route-change evidence, tail regret, and latency.

Table 1. Learned-primary claim branch used in the routing studies. Each level inherits lower-level requirements; primary-policy attribution and adjudication-record closure are branch-common.

Outcome	Careful review	Evidence checklist	SCOPE-Routing
Agreement with operational reference	24/36	27/36	31/36
Reference-relative upward deviation	8/36	5/36	2/36
Reference-relative downward deviation	1/36	3/36	4/36
Median time per claim	18.1 min	13.4 min	15.8 min

Table 2. Held-out reviewer comparison on 36 expert-resolved claims.

Decision rule	Selected row	Regret	Key evidence	Licensed interpretation
Score only	Fallback-Heavy Safe	0.011091	86.1% backup rescue	Hybrid credit retained; no learned-primary $C5$ .
Target $C5$ , then regret	Shock-Safe Online Repair	0.016189	Dynamic and attribution evidence complete	$C5$ learned-primary.
Coverage-oriented diagnostic	Preference Hypernetwork	0.020865	HV $=0.849$	$C4$ ; $C5$ dynamics incomplete.

Table 3. Selection depends on the scientific decision rule.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Prior line of work	Primary output	Composition role in claim licensing
Argument-based and AI-evaluation validity	Evidence-backed interpretations, uses, and claims.	Supplies validity arguments; the contract operationalizes one bounded row-level conclusion decision.
Evidence-licensed calibration and claim replay	Calibrated assertion scope, maximal licensed frontiers, replayable inference, and semantic stability.	Supplies evidence-relative claim semantics; subject-typed licensing separates requested-subject answers from non-substitutive credit.
Structured assurance cases	Traceable claim–argument–evidence graph.	Supplies explicit support structure that can be encoded as claim obligations.
Claim-evidence retrieval and assessment	Evidence links, support labels, justifications, or graded overstatement scores.	Supplies evidence and support judgments consumed by obligation evaluators.
Datasheets and model cards	Documentation of data, models, intended uses, and limitations.	Supplies inspectable evidence conditions and intended-scope fields.
BenchmarkCards and benchmark documentation	Standardized benchmark metadata and risk reporting.	Supplies benchmark-level evidence records and scope declarations.

Appendix

Table 4. Feature-level positioning of adjacent evaluation artifacts in the claim-licensing pipeline.

Adjudicator	Evidence diagnosis	Claim order	Source semantics	Incomparable credit	Primary output
Regret-only ranking	Scalar score	Absent	Absent	Absent	Score-ordered row.
Missing-field checklist	Structured gap prompts	Reviewer-authored	Prompted or implicit	Reviewer-authored	Gap record plus reviewer-selected claim.
Designated-claim assessment	Evidence links or support scores	Usually absent	Source-bound or implicit	Usually absent	Support label, score, or justification.
Unstructured human review	Contextual	Implicit	Implicit	Implicit	Review prose and judgment.
SCOPE-Routing	Typed obligations	Explicit	Explicit	Explicit $\mathcal{B}$ relation	$(\mathcal{C}^{\star},\mathcal{R}^{\star},F,z)$ .

Appendix

Table 5. Adjudication strategies compared by represented information and primary output.

Architectural role	Routing instantiation	Selection-task replacement
Decision quality	Route regret	Downstream selected-set objective or regret
Feasibility	Valid source–destination path	Budget, cardinality, or packing feasibility
Action identity	Parallel-edge identity	Item identity and duplicate/substitution handling
Coverage	Preference-space hypervolume	Coverage across budgets, preferences, or item strata
Attribution	Fallback solver or repair	Optimizer, repair, abstention, or fallback disclosure
Robustness	Drift, route change, tail regret	Distribution shift, set change, tail decision loss

Appendix

Table 6. Worked design sketch for a predict-then-optimize selection port; empirical cross-domain validation remains a separate study.

Channel	Representative fields	Claim-validity role
Quality	scalarized regret, tail regret, oracle objective, comparator identity	Establishes C0 scoring evidence.
Feasibility	aggregate success, feasible-set-nonempty flag, conditional invalid-route count, hidden-repair flag	Determines whether C1+ feasibility-audited claims can be credited; rounded aggregate rates are not the gate.
Fallback	claim subject, fallback reason/rate/trigger, primary-only and end-to-end quality, fallback-credit flag	Separates learned-component evidence from hybrid rescue and safety-control credit.
Structure	edge identity, bundle id, retained-edge set, collapse-ablation id	Prevents simple-graph evaluation from certifying C2 multigraph claims.
Preference	preference vector, sampler, hypervolume, held-out shift split	Tests whether scalar gains preserve C3 preference-space behavior.
Dynamics	route identity, route-change W/T/L, drift severity, non-static-equivalence flag	Distinguishes dynamic learned behavior from static-equivalent rows.

Appendix

Table 7. Telemetry channels used by the adjudicator.

Evidence source	Instantiated structure	Default admissible claim scope
Controlled multigraph generator	Score, edge identity, preference vectors, hard feasibility, and controlled drift.	Controlled C1–C5 claims under generator assumptions.
Same-contract RouteScope-Lite suite	Multidomain semisynthetic graphs, shared routing contract, 15 policy families, 10 seeds, telemetry-complete rows.	C1–C5 same-contract evidence; C6 additionally requires deployment evidence.
Public shortest-path bridges	Score transfer under a public shortest-path or contextual path contract.	C0 learned-scoring bridge credit; routing C3 additionally requires inherited feasibility and multigraph-identity obligations.
Diagnostic realistic slices	Realistic topology or dynamics with limited slice size and mechanism probes.	Mechanism diagnosis unless the slice is large enough and telemetry-complete enough for confirmatory realistic claims.
Native or external ML4CO adapters	Task-dependent objectives, metrics, and solver-interface behavior.	Compatibility, native-metric, efficiency, or out-of-contract claims; matching route objects and telemetry enable routing claims.
External public-claim corpus	Four blinded non-author annotation streams and executable labels of claims from published evidence snippets.	Contract-relative operational reference for the specified evidence packets.

Appendix

Table 8. Evidence-source scope.

Layer	What it tests	Main anti-failure role
Fixed packets	Whether non-author experts apply the supplied contract consistently.	Conditional applicability of supplied objects.
Object construction	Whether non-authors reconstruct inputs from prespecified sections.	Upstream author dependence.
Reviewer comparison	Whether the method changes agreement, overpromotion, rationales, and time.	Practical benefit and conservative error.
Same-contract suite	Whether target-claim eligibility changes the selected row.	Consequence under shared reporting.
Scope and rule stress	Whether source credit and distinct gates behave as intended.	Transfer discipline and functional distinctness.
Versioned artifact	Whether core and extension-stage evidence can be traced and reanalyzed.	Chronological and computational closure.

Appendix

Table 9. Validation layers and the distinct evidential role of each layer.

Probe	What is removed	Failure mode exposed
Regret-only ranking	Source scope, inherited gates, and record closure.	Selects Fallback-Heavy Safe in RouteScope-Lite because it has the lowest scalar regret; 86.1% fallback assigns the row to rescue-dominated hybrid evidence.
Missing-field checklist	Cumulative claim semantics and narrowing.	Can detect absent fields but cannot decide whether a row should be unsupported , narrowed , or credited at source-native scope.
Source-scope-only rule	Row-level telemetry gates.	Enforces bridge/native/out-of-contract transfer boundaries; fallback attribution, graph identity, dynamic behavior, and latency remain outside this reduced rule.
Leave-one-gate-out rules	One inherited gate at a time.	Each omitted gate enables a distinct false-support mode: fallback-heavy rescue, multigraph collapse, coverage loss, static equivalence, hidden latency, source upgrade, or unlocked artifacts.
Human-majority reference	Deterministic execution and manifest linkage.	Provides external claim-scope evidence, but is not a rerunnable row-level adjudicator.

Appendix

Table 10. Reduced adjudicators and rule ablations. The comparison separates three objects that are easy to conflate: reduced adjudicators, rule-necessity ablations, and the external human reference.

Stratum	Count	Sampling rule
Learned routing and neural routing	24	Public routing papers with empirical route-quality claims.
Decision-focused and predict-then-optimize learning	20	Claims about downstream decision quality, regret, or solver-aware training.
ML4CO benchmarks and native solver suites	22	Claims about benchmark coverage, solver-interface evidence, or native metrics.
Multiobjective optimization and preference learning	18	Claims involving hypervolume, coverage, Pareto quality, or preference conditioning.
Evaluation science, documentation, and stress testing	24	Claims about benchmark validity, intended use, stress tests, or evaluation frameworks.
External bridge tasks and out-of-contract comparators	12	Claims where task scope differs from SCOPE routing but evidence is relevant to source-scope behavior.

Appendix

Table 11. ClaimScope-120 corpus composition.

Stage	Object or count	Reviewer check in the artifact
Candidate collection	Candidate claim records	Candidate IDs, source-family tags, claim location, inclusion/exclusion decision, and exclusion reason.
Corpus sample	120 claims from 40 papers	Stable Claim Cards with anonymized source identifiers and evidence excerpts.
Human annotation	Four blinded human labels per claim	Raw annotation matrix, missing-gate selections, anonymized annotator metadata, blinding/COI fields, and validation report.
Consensus construction	Majority label with tie rule	Scripted majority vote and tie-resolution log.
SCOPE-Routing adjudication	Executable label and missing gates	Adjudicator output linked to evidence source, telemetry/prose evidence, and artifact fields.
Disagreement audit	19 exact disagreements	Disagreement taxonomy and representative anonymized examples.

Appendix

Table 12. Reviewer-facing audit chain for ClaimScope-120 .

Field or outcome	Agreement	Interpretation
Comparator identity	23/24	Comparator extraction is highly reproducible.
Source family	22/24	Residual ambiguity concerns solver-interface credit.
Requested claim family	21/24	Dynamic wording creates the main family ambiguity.
Intended scope	weighted $\kappa=0.78$	High but non-perfect ordinal agreement.
Required/missing gates	mean Jaccard $=0.82$	Most evidentiary obligations are reconstructed similarly.
Final status	20/24 exact; 24/24 within one	Remaining differences are adjacent.

Appendix

Table 13. Section-conditioned claim-evidence-object construction audit.

ID	Claim family	Human majority	SCOPE-Routing	Reason for disagreement
ECC-012	Shortest-path bridge	C1	C0	Humans credited feasibility from the shortest-path task; SCOPE-Routing kept source scope at C0 bridge because multigraph identity and fallback telemetry were absent.
ECC-019	Learned routing	C5	C4	Claim used dynamic language, but route-change and drift-stress telemetry were incomplete.
ECC-027	Multiobjective method	C3	C2	Strong scalarized results were reported, but hypervolume or preference coverage was missing.
ECC-038	Benchmark paper	C6	C5	Humans interpreted broad realistic coverage as deployment-style evidence; SCOPE-Routing required non-author usability and deployment telemetry.
ECC-044	Native ML4CO suite	C3	out-of-contract	Native task instantiated multiobjective metrics but not source-destination multigraph routing.
ECC-061	Decision-focused learning	C0	C1	Human majority credited only scalar scoring; SCOPE-Routing found explicit feasible-decoding evidence in the excerpt.

Appendix

Table 14. Representative external-corpus disagreement cases. Source titles are anonymized in the submission version; artifact cards contain stable claim identifiers and source-family labels.

Policy family	Type	Evidentiary role
Dynamic exact oracle	upper bound	Provides oracle quality and feasibility reference.
Static exact anchor	comparator	Strong nonlearned comparator for regret and route-change checks.
Weighted shortest path	heuristic	Classical baseline with explicit weights.
$k$ -shortest-path heuristic	heuristic	Tests nonlearned multigraph candidate generation.
Simple-graph collapse	negative control	Tests C2 multigraph identity gate.
Learned full-exact scorer	learned scoring	Tests feasibility-audited controlled learned-scoring claims.

Appendix

Table 15. Policy-family inventory for RouteScope-Lite .

Policy family	Regret	Fallback	Feas.	HV	Licensed interpretation
Dynamic exact oracle	0.000000	0.000	1.000	0.892	Upper bound; not ranked as learned policy.
Fallback-Heavy Safe	0.011091	0.861	0.997	0.669	Standalone C5 not licensed; hybrid safety-control credit retained in $\mathcal{R}^{\star}$ .
GNN Pruned Top- $k=1$	0.012508	0.505	0.955	0.439	C0; fallback and coverage gates fail.
Static exact anchor	0.014874	0.000	1.000	0.716	Comparator anchor, not learned routing.
Shock-Safe Online Repair	0.016189	0.012	1.000	0.819	C5-eligible under the declared claim-conditioned rule.
Adaptive Pruning Repair	0.016761	0.016	1.000	0.799	C4; repair improves pruning claim level.

Appendix

Table 16. Selected RouteScope-Lite rows. “Feas.” is the reported aggregate rate; C1+ uses exact conditional invalid-route counts.

Selection rule	Selected row	Reason
Regret only	Fallback-Heavy Safe	Lowest aggregate regret, without an attribution eligibility requirement.
Target C5, then regret	Shock-Safe Online Repair	Passes inherited C5 gates; eligible rows are then ordered by regret.
Target C5, then normalized cost	Shock-Safe Online Repair	Remains selected after the declared latency and cost normalization.
Coverage-first diagnostic	Preference Hypernetwork	Highest learned non-oracle HV, but incomplete C5 dynamics.

Appendix

Table 17. Predeclared claim-conditioned selection rules for RouteScope-Lite ; each rule ranks only its eligible policy rows.

Quantity used in claim	Point estimate in main table	Required uncertainty check
Regret order between FallbackHeavySafe and ShockSafeOnlineRepair	0.011091 vs. 0.016189	Request-paired bootstrap interval for regret difference; used to characterize the regret-only ranking before claim-conditioned eligibility.
Fallback attribution	0.861 vs. 0.012	Seed- and request-level fallback intervals plus primary-only/end-to-end decomposition; narrowing depends on the aggregate gain remaining rescue-dominated.
Feasibility boundary	0.997 vs. 1.000	Aggregate interval and exact invalid-route counts conditional on nonempty feasible sets; the exact conditional counts determine C1.
Coverage comparison	0.669 vs. 0.819 HV	Hypervolume interval; C3+ support uses nondegraded preference-space evidence alongside scalar regret.
Route-change and dynamic evidence	artifact-reported W/T/L	Route-change win/tie/loss and non-static-equivalence checks; C5 support uses observed dynamic behavior beyond inherited static-anchor routes.
Latency and cost-conditioned selection	artifact-reported p50/p95/p99	Latency intervals, hardware record, normalization, weights, and tie handling; operational selection depends on these records.

Appendix

Table 18. RouteScope-Lite uncertainty audit. The conclusion change follows from a declared target claim, eligibility evidence, and an explicit within-eligible quality rule.

Requested claim	Observed evidence gap	Evidence-supported effect
Identity-preserving multigraph routing	Evaluation collapses parallel edges to a simple graph	Multigraph claim blocked; lower structural or scalar credit may remain.
Preference-conditioned routing	Scalar quality reported without preference-space coverage	$C3$ blocked; scalar and inherited lower-scope credit retained.
Dynamic robust routing	No route-change or drift evidence	$C5$ blocked; supported static or attribution-audited credit retained.

Appendix

Table 19. Representative interpretation changes in held-out public-paper claims.

Quantity	Value	Interpretation
Claims / papers / domains	120 / 40 / 6	Public corpus outside the internal SCOPE ledger.
Annotators	4 blinded non-author experts	ML4CO/routing, evaluation/benchmarking, OR/optimization, and ML systems/reproducibility perspectives.
Mean pairwise exact agreement	0.761	Experts often agree exactly on nontrivial claim-scope labels.
Mean pairwise within-one-level agreement	0.931	Most disagreements occur at adjacent scope levels.
Weighted Cohen’s $\kappa$ (mean pairwise)	0.703	Substantial ordinal agreement.
Krippendorff’s $\alpha$ (ordinal)	0.686	Conservative multi-annotator reliability estimate.

Appendix

Table 20. Conditional fixed-packet agreement using blinded non-author annotation. Human majority provides the operational reference under the specified guide and evidence packets.

Majority group	C0/bridge	C1–C2	C3–C4	C5–C6	Off/insufficient
C0/bridge	25	2	0	0	1
C1–C2	3	21	2	0	1
C3–C4	1	2	32	2	0
C5–C6	0	0	3	15	2
Off/insufficient	0	0	0	0	8

Appendix

Table 21. Collapsed confusion matrix for SCOPE-Routing vs majority human label on ClaimScope-120 . Rows are majority labels; columns are SCOPE-Routing labels.

Disagreement type	Count	Typical pattern
Ambiguous deployment wording	6	The paper uses realistic or practical language without C6-scale evidence.
Source-scope strictness	5	Humans credit bridge or native-task evidence one level higher than the source-scope rule allows.
Artifact availability	4	Humans infer reproducibility from prose, while SCOPE-Routing requires manifest-linked objects.
Preference coverage boundary	2	Scalar preference results are strong but coverage evidence is partial.
Latency/online-effect boundary	2	Dynamic or online language appears, but route-change or latency telemetry is incomplete.

Appendix

Table 22. Disagreement taxonomy for the 19 exact disagreements between SCOPE-Routing and the majority human label.

Public source	Native/source-supported credit	SCOPE-transfer scope	Boundary reason
PyEPO ( Tang and Khalil, 2022 )	Predict-then-optimize library evidence across shortest path, knapsack, TSP, and LP/IP tasks.	Bridge/native decision-quality evidence only.	Does not instantiate the SCOPE dynamic multigraph route object or its telemetry contract.
DataSP ( Lahoud et al., 2024 )	Contextual cost learning and differentiable all-to-all shortest-path/path-prediction evidence.	Shortest-path bridge evidence with coverage caveat.	Parallel-edge identity, SCOPE fallback taxonomy, preference-shift coverage, and dynamic route-change telemetry are absent.
FrontierCO ( Feng et al., 2026 )	Native large-scale CO solver benchmark/readiness evidence; local TSP diagnostics are executable.	Native diagnostic/readiness, not same-contract SCOPE routing.	No SCOPE source-destination route-producing policy and no SCOPE route telemetry.
CO-Bench ( Sun et al., 2026 )	Native LLM-agent algorithm-search benchmark; RCSP native solver-interface audit solves or certifies local source-scope review instances and derived records.	Solver-interface/source-scope evidence; no learned multigraph-routing escalation.	Evaluated object is algorithm search or native RCSP solving, not a SCOPE route-producing policy.
PredOpt-SP	Reproduced shortest-path decision-focused bridge comparators.	C0 shortest-path bridge audit.	Not strict same-walltime evidence and not dynamic preference-conditioned multigraph routing.
PredOpt-Knapsack	Native constrained-optimization closure over local selector and official comparator rows.	Off-thesis positive evidence, not routing evidence.	Knapsack is not a source-destination route contract.

Appendix

Table 23. Public Scope Firewall. Native/source-supported credit is preserved, while SCOPE-transfer scope remains bounded by the dynamic preference-conditioned multigraph routing contract.

Evidence block	Coverage	Claim boundary
PredOpt-SP	Latest local official-code extension plus preference-function bridge rows; reproduced SPO+ bridge remains stronger than the local selector row on the audited bridge slice.	Shortest-path bridge audit; blocks all-baseline or state-of-the-art routing claims.
PredOpt-Knapsack	Local selector rows and official comparator rows across method files; the selector-family mean is better than the best comparator-family mean on the native constrained-optimization audit.	Full native PredOpt knapsack closure, but out-of-contract for routing.
DataSP	Three-seed SCOPE bridge over fixed synthetic episodes plus SPO+ bridge baseline.	C0 shortest-path bridge with coverage caveat, not native full DataSP routing validation.
FrontierCO TSP	Native TSP/LKH diagnostics and seed portfolio; no SCOPE-native tour-producing policy.	Native diagnostic/readiness evidence only under unmatched task contract.
CO-Bench RCSP bridge	Subset of raw RCSP instances converted through a single-resource SCOPE bridge; the subset is saturated by static/exact baselines.	RCSP bridge coverage audit; no learned-routing win claimed.
CO-Bench RCSP native audit	All local source-scope review instances are audited; each instance is solved optimally or certified infeasible under the native solver-interface check, without redistributing upstream raw assets whose reuse terms are unclear.	Native solver-interface/source-scope evidence; C2–C6 SCOPE routing promotion remains blocked.

Appendix

Table 24. Executable external benchmark audit. These rows support native, bridge, out-of-contract, or diagnostic claims; they are not promoted into same-contract SCOPE routing unless the SCOPE route contract is instantiated.

Claim family	Regret-only interpretation	Status	Deciding evidence and supported scope
Aggressive pruning	Top- $k=1$ appears attractive because it reaches regret 0.009552.	unsupported	Fallback rises to 40.94% and hypervolume falls to 227.985; no positive pruning or operational routing claim.
Simple-graph collapse	Feasible-only regret could suggest that collapse is acceptable.	unsupported	Feasibility falls to 90.6% on the controlled slice and 50.0% on the diagnostic slice; C2+ requires edge identity.
Realistic diagnostic learned win	Equality with a strong static anchor could be read as realistic stability.	unsupported	Equality is mediated by 100% fallback and route-static behavior; supported claim is diagnostic failure analysis.
DataSP bridge	Scalar regret improves from 0.21546 to 0.03352.	narrowed	Hypervolume decreases from 0.32903 to 0.30314 and the source is not multigraph; supported claim is learned-scoring bridge evidence with a coverage caveat.
FrontierCO diagnostics	External compatibility could be read as native routing success.	unsupported	The requested routing branch is empty; $\mathcal{R}^{\star}$ retains compatibility and solver-interface efficiency diagnostics.
Public Scope Firewall	Real public benchmark evidence could be overread as same-contract routing validation.	narrowed	Public rows retain native/source-supported credit while SCOPE-transfer is blocked when the dynamic multigraph route contract is absent.

Appendix

Table 25. Closed internal registry used for the finite demonstration; population-rate estimation requires probability sampling.

Row	Regret	$\Delta$ vs static	Feas.	Fallback	Adjudicated claim
Static exact anchor	0.014154	–	100%	0.000000	Static exact comparator.
Learned full-exact scorer	0.009878	-30.2%	100%	0.000000	supported to C0–C4: feasibility-audited controlled learned scoring.
Top- $k=4$ retained decode	0.009931	-29.9%	100%	0.000000	narrowed to top- $k$ -specific retained-decode evidence.
Collapsed exact control	n/a	–	90.6%	0.000000	unsupported for C2+ multigraph claims.

Appendix

Table 26. Positive calibration on the controlled multigraph slice. The evaluator preserves valid positive credit while blocking unsupported escalation.

Evidence block	Highest licensed level	Missing C6 evidence	Interpretation
Controlled internal audit	C4	Multi-domain realistic support, dynamic robustness, non-author usability.	Valid controlled learned-scoring evidence only.
RouteScope-Lite semisynthetic suite	C5	Deployed topology, live workload, non-author reproduction, intended-use validation.	Strong same-contract stress evidence, not deployment.
Diagnostic realistic slice	diagnostic only	Sample size, fallback-free learned win, multi-domain realism.	Mechanism evidence for fallback/static failures.
External public rows	Scoped public credit	Same routing contract, telemetry completeness, route identity.	External scope firewall evidence.
Public Scope Firewall	Scoped source credit	Dynamic preference-conditioned multigraph routing contract and telemetry.	Real public sources receive native, bridge, solver-interface, or out-of-contract credit, not deployment or multigraph-routing escalation.
Artifact dry run	Record closure	Live external use, public non-author adapters, and deployment evidence.	Reviewer-runnable analysis, not community adoption.

Appendix

Table 27. Deployment boundary: available evidence licenses controlled claims through C5, while C6 requires deployment evidence.

Gate removed or stressed	Overcredit events	False-support mode blocked by full rule
Fallback	4	Fallback-heavy pruning, fallback-heavy same-contract winners, or diagnostic equality would be credited as learned routing.
Multigraph identity / feasibility	3	Simple-graph collapse would be credited as valid multigraph routing.
Preference coverage	3	Scalar regret would be credited as preference-conditioned coverage despite HV loss.
Route change	3	Static-equivalent or no-effect online rows would be credited as dynamic behavior.
Latency	2	Operational claims would ignore slower or impractical paths.
Source scope	5	Bridge, native, or out-of-contract evidence would be silently upgraded to full routing.

Appendix

Table 28. Gate ablations. Each removed gate enables a distinct false-support mode.

Policy	Missing-evidence treatment	Boundary results expected to remain unchanged
Conservative SCOPE-Routing	Any inherited missing gate blocks support at the affected level; lower supported levels may remain.	Main paper rule.
Intermediate SCOPE-Routing	Record, source-scope, feasibility, fallback-attribution, and multigraph-identity requirements remain hard; latency and partial coverage gaps usually narrow.	Rescue-dominated rows still cannot become standalone learned-routing evidence; semisynthetic rows still cannot become C6.
Permissive SCOPE-Routing	Missing non-core telemetry narrows by one or more levels when source scope is compatible and no hard requirement fails.	Bridge/out-of-contract evidence still cannot become full multigraph routing; unlocked records still cannot support adjudicated claims.

Appendix

Table 29. Sensitivity policies for missing telemetry. The manuscript’s main results use the conservative policy; the artifact can report intermediate and permissive variants as a robustness check.

Evidence object	Core snapshot	Extension snapshot	Execution mode
Adjudicator, schema, Claim Cards, registry	yes	unchanged/core	Fresh validation and table/card generation.
ClaimScope-120 packets and four-expert labels	yes	extended reporting	Fresh reanalysis from released labels and tie logs.
RouteScope-Lite request and policy rows	yes	unchanged/core	Cached reanalysis from locked rows; not model retraining.
Public source-scope audits	yes	extended summaries	Fresh local checks or derived source-scoped summaries as documented.
40-claim reviewer comparison	no	added	Fresh analysis from anonymized raw judgments, assignments, and times.
24-object construction audit	no	added	Fresh analysis from independent encodings and resolution logs.

Appendix

Table 30. Artifact chronology and execution mode. “Fresh analysis” recomputes results from released raw records; “cached reanalysis” recomputes summaries from locked experimental outputs; “derived-only” preserves historical summaries without silently rerunning external or heavy experiments.

Asset	Asset type	Source-provided status	Restricted use in this submission
PyEPO	Code/library	MIT License	Credited to the original repository and paper; used for source-scope and bridge/native decision-quality comparisons. Upstream notices are preserved when the package is installed or referenced.
PredOpt-family benchmark assets	Code/benchmark repository	MIT License for the public benchmark repository checked in the audit	Used for shortest-path and knapsack source-scope comparisons; upstream notices are preserved and no stronger SCOPE-routing claim is inferred from native PredOpt evidence.
FrontierCO dataset card	Dataset/benchmark card	Dataset-card license marker recorded in the manifest	Used as source-scoped external benchmark evidence; no raw dataset is relicensed by this submission and no deployment-style or same-contract SCOPE routing claim is inferred.
FrontierCO code repository	Code repository	No clear reusable software license identified in the repository front matter during the audit	Not vendored, mirrored, or required by the default reviewer path; only paper/repository citation and independently authored diagnostic wrappers or aggregate audit summaries are included.
CO-Bench dataset card	Dataset card	License marked unknown on the dataset card during the audit	Treated conservatively as source-bound; raw upstream data are not redistributed, automatically downloaded, or relicensed. The artifact includes only source citations, Claim Cards, and derived aggregate source-scope records.
CO-Bench code repository	Code repository	No clear reusable software license identified in repository front matter during the audit	Not vendored, mirrored, or required by the default reviewer path; solver-interface checks use independently authored review code and derived summaries rather than copied upstream code.

Appendix

Table 31. License and terms-of-use status for external assets touched by the source-scope audit, as checked on the official project pages on May 4, 2026. “Not redistributed” means that the artifact does not vendor, mirror, or relicense the upstream code or raw data; only independently authored scripts, source citations, Claim Cards, and derived aggregate audit records are included.

Check	Snapshot / mode	Output artifact
Manifest and snapshot validation	Both / fresh	Hash and chronology report.
Registry and main-table generation	Reviewed / fresh	Registry, exclusions, tables, and CSV mirrors.
Claim Cards, schema, adapters	Reviewed / fresh	Cards plus schema and 15-adapter compliance reports.
RouteScope-Lite analysis	Reviewed / cached reanalysis	22,500 locked policy-row records, full 15-row table, uncertainty and selection reports.
ClaimScope-120 analysis	Reviewed / fresh	120 cards, raw-label matrix, tie policies, clustered intervals, and disagreement taxonomy.
Reviewer comparison	Extension / fresh	240 judgments, 36 resolved claim summaries, timing and clustered analyses.

Appendix

Table 32. Versioned artifact checks and execution modes.

Explore similar work

Sep 30, 2026cs.AI

PathAnchor: Path-Structured Evidence for Scientific Agents

Scientific agents can retrieve relevant passages yet still lose functional order, mix evidence across sources, or state conclusions that exceed the retrieved record. We introduce PathAnchor, a bounded scientific reasoning system built on path-structured evidence workspaces. Instead of treating passages or extracted concepts as independent units, the system retrieves source-linked Material-Sensor-Signal-System trajectories that preserve role, direction, and the evidence supporting each transition. A controller uses three read-only tools to search paper-specific trajectories, trace paths across candidate sources, and open exact evidence before producing a claim-cited answer and an explicit evidence boundary. On 120 single- and cross-paper flexible-sensor questions, PathAnchor scores 82.6% and leads six evaluated systems. Under a matched controller, corpus, and six-call budget, replacing unordered concept graphs with path-structured records raises source recall from 61.3% to 82.9%, increases answers whose claims all cite opened evidence from 69.2% to 90.0%, and reduces tool calls. These results show that evidence organization affects retrieval and citation completeness under fixed agent resources.

Qiuhui Chen, Jiafan Lu, Shuaimin Tang +5

School of Information Science and Engineering, East China University of Science and Technology · College of Intelligent Robotics and Advanced Manufacturing, Fudan University · Tencent

Citations Scientific Discovery Evidence-Grounded Questions Graph-Based Retrieval Scientific Agents

Ask PDF

Jul 29, 2026cs.AI

Evidence-Ledger Adjudication for Claim-Evidence Traceability

AI agents can draft claims faster than authors can check whether the cited or retrieved evidence supports them. We study evidence-ledger adjudication: a claim-evidence traceability workflow that pairs each claim with an evidence packet, assigns a support relation, and routes unsupported, contradicted, or mixed-evidence claims back to the author. The empirical core is a 2,335-row blind benchmark built from independent external labels in AVeriTeC, CLIMATE-FEVER, and SciFact. Gold relations and source evidence labels are hidden during prediction and joined only for scoring. On this benchmark, the agent evidence-ledger condition achieves 0.676 relation accuracy and 0.601 macro-F1, compared with 0.383 accuracy and 0.303 macro-F1 for the best non-agent baseline. It also routes 1270/1435 claims whose gold labels indicate contradiction, missing evidence, or mixed evidence, while routing 295/900 supported claims. These results show that evidence-ledger adjudication can turn heterogeneous evidence packets into an auditable traceability layer for AI-assisted writing.

Gengyu Chen, Yongjie Yu, Weiling Wang

Carnegie Mellon University · Syracuse University

Category-Aware Atomic Claims Ai-Assisted Writing

Ask PDF

May 26, 2026cs.AI

DeepSciVerify: Verifying Scientific Claim--Citation Alignment via LLM-Driven Evidence Escalation

Misalignment between claims and their cited evidence is a common failure mode in reports generated by large language models, limiting their reliability in scientific and other high-stakes settings. We present DeepSciVerify, a two-stage pipeline for scientific claim-citation verification that combines abstract-level reasoning with selective escalation to passage-level evidence. The system first verifies claims using the abstract and defers uncertain cases, retrieving and analyzing full-text passages only when necessary. This design leverages complementary behaviors across LLMs, as some models are more conservative while others are more decisive under uncertainty. On the SCitance benchmark, DeepSciVerify achieves 86.7 Micro-F1, outperforming strong abstract-only baselines by +4.5 points while resolving 67% of instances without full-text retrieval. These results suggest that selective evidence escalation improves both accuracy and efficiency in claim-citation verification.

Shaghayegh Sadeghi, Khashayar Khajavi, Rise Adhikari +1

FirstPrinciples · School of Computing Science, Simon Fraser University

Category-Aware Atomic Claims Citations Scientific Discovery Large Language Model Reliability Peer Review Reasoning Skills

Ask PDF