{"schema_version":1,"research_id":"eoa_inverse_innovation_exp04_external_evaluation_20260802","source_assessment_id":"invariant_mode_decomposition_design__library_information_science:SENTINEL_MATCHED:v0","cell_id":"invariant_mode_decomposition_design__library_information_science","search_queries":["site:niso.org library discovery service recommended practice ODI 2020","library recommender systems popularity bias exposure diversity research digital libraries","library discovery systems algorithmic bias diversity official report","spectral analysis feedback loops recommender systems exposure fairness dynamic","recommender systems feedback loop Jacobian eigenvalue stability analysis","dynamic recommender fairness exposure constraints repeated rounds paper","library discovery algorithm diversity equity strategic plan recommender official","ALA privacy library circulation data deidentified logs policy official","\"spectral\" recommender system feedback loop stability exposure","\"Jacobian\" \"recommender systems\" feedback loop","\"eigenvalue\" recommendation feedback loop fairness","site:bls.gov/ooh data scientists median pay 2025","ALA \"Guidance on the Use of Artificial Intelligence in Libraries\" publication date"],"sources":[{"source_id":"S1","title":"Guidance on the Use of Artificial Intelligence in Libraries","publisher":"American Library Association","url":"https://www.ala.org/tools/standards-and-guidelines/guidance-use-artificial-intelligence-libraries","source_class":"OFFICIAL_GUIDANCE","publication_date":"2026-07-16","accessed_at":"2026-08-02","claims_supported":["Library organizations should review AI-enabled discovery, ranking, recommendation, and personalization for patterns that reduce visibility or distort representation.","Library administrators, governing bodies, consortia, and accountable staff are identifiable authorizers for such reviews.","Usage logs should remain under library or public-sector control, and non-public records require formal privacy, security, and where applicable legal review.","Libraries should retain human accountability, explainable non-personalized access paths, auditability, and authority to pause or discontinue unsafe systems."]},{"source_id":"S2","title":"NISO RP-19-2020, Open Discovery Initiative: Promoting Transparency in Discovery","publisher":"National Information Standards Organization","url":"https://www.niso.org/publications/rp-19-2020-odi","source_class":"STANDARD","publication_date":"2020-06-24","accessed_at":"2026-08-02","claims_supported":["Library discovery already has an institutional transparency and conformance framework.","Recognized discovery priorities include library responsibilities, meaningful usage statistics, fair linking, content-coverage disclosure, and record-source identification.","A modal audit would need to complement rather than replace established discovery transparency practices."]},{"source_id":"S3","title":"A Study of Position Bias in Digital Library Recommender Systems","publisher":"arXiv","url":"https://arxiv.org/abs/1802.06565","source_class":"PRIMARY_RESEARCH","publication_date":"2018-02-19","accessed_at":"2026-08-02","claims_supported":["A randomized study covering ten million recommendations in two digital-library-related systems found substantial position effects.","Top-ranked recommendations received 53% and 87% more clicks than the study's respective non-biased expectations, supporting the premise that ranking-generated exposure affects observed engagement."]},{"source_id":"S4","title":"Bias in Book Recommendation: A Case Study on the Danish Public Libraries","publisher":"Centrum Wiskunde & Informatica","url":"https://ir.cwi.nl/pub/36328","source_class":"PRIMARY_RESEARCH","publication_date":"2026-03-24","accessed_at":"2026-08-02","claims_supported":["A recommender used by Danish public libraries exhibited popularity and author-nationality exposure disparities.","Some configurations produced up to a 40% relative exposure decrease for less-popular nationalities.","The authors expressly call for systematic bias analysis as a structural component of library recommender evaluation."]},{"source_id":"S5","title":"Feedback Loop and Bias Amplification in Recommender Systems","publisher":"arXiv","url":"https://arxiv.org/abs/2007.13019","source_class":"PRIMARY_RESEARCH","publication_date":"2020-07-25","accessed_at":"2026-08-02","claims_supported":["Offline repeated-interaction simulation shows that recommender feedback can amplify popularity bias over iterations.","Reported consequences include declining aggregate diversity, preference-representation shifts, and homogenization.","The reported feedback impact was stronger for the modeled minority group, supporting affected-group monitoring while not establishing library-specific prevalence."]},{"source_id":"S6","title":"Fairness of Exposure in Dynamic Recommendation","publisher":"arXiv","url":"https://arxiv.org/abs/2309.02322","source_class":"PRIMARY_RESEARCH","publication_date":"2023-09-05","accessed_at":"2026-08-02","claims_supported":["Static, round-by-round exposure mitigation can fail to deliver long-run exposure fairness.","A cumulative-exposure-aware dynamic constraint is an implemented comparator that improved long-run exposure fairness while maintaining recommendation accuracy in an offline MovieLens simulation.","The paper supplies a close non-modal rival and demonstrates that repeated-round evaluation is established prior art."]},{"source_id":"S7","title":"Optimal Control Synthesis of Closed-Loop Recommendation Systems over Social Networks","publisher":"arXiv","url":"https://arxiv.org/abs/2603.10275","source_class":"PRIMARY_RESEARCH","publication_date":"2026-03-10","accessed_at":"2026-08-02","claims_supported":["Control-theoretic recommendation with explicit engagement, diversity, deviation, and polarization tradeoffs is existing adjacent research.","The work uses algebraic spectral conditions to distinguish stabilizing from pathological closed-loop behavior.","Its continuous-time opinion-dynamics setting and social-network objectives differ from the candidate's fitted library exposure-stratum Jacobian and retrospective operational audit."]},{"source_id":"S8","title":"Data Scientists","publisher":"U.S. Bureau of Labor Statistics","url":"https://www.bls.gov/ooh/math/data-scientists.htm","source_class":"GOVERNMENT_OR_REGULATOR","publication_date":"2025-08-28","accessed_at":"2026-08-02","claims_supported":["The May 2024 median annual wage for U.S. data scientists was $112,590, with a reported range from below $63,650 at the tenth percentile to above $194,410 at the ninetieth percentile.","Relevant duties include determining data availability, analyzing data, validating models, and presenting stakeholder recommendations.","This wage benchmark supports resource-equivalent labor estimates but not vendor, legal, infrastructure, or local-government procurement prices."]}],"problem_evidence":{"support":"STRONG","rationale":"The exact proposed coupled modal mechanism has not been observed in a library, but the underlying problem is visible and consequential. Randomized digital-library evidence establishes ranking-position effects on engagement; a 2026 public-library case study reports configuration-dependent popularity and nationality exposure disparities; and repeated-round recommender studies show feedback amplification, diversity loss, and long-run failure of static exposure controls. These sources do not establish prevalence across libraries or prove that a reproducible cross-stratum eigenmode exists.","source_ids":["S3","S4","S5","S6"]},"stakeholder_evidence":{"support":"STRONG","rationale":"ALA expressly directs library organizations to review AI-enhanced discovery, ranking, recommendation, and personalization for visibility and representation distortions, while identifying administrators, governing bodies, consortia, and local accountable staff as decision makers. NISO supplies an established discovery-governance context. This is strong expressed need from professional authorities, although no named library has committed logs, staff, funding, or adoption authority to this candidate.","source_ids":["S1","S2","S4"]},"prior_art":{"proximity":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"Systematic bias audit of Booklens in Danish public libraries","similarity":"Same library context, algorithmic recommendation substrate, exposure disparities, configuration comparisons, and institutional equity objective.","remaining_difference":"It audits static configuration-dependent popularity and nationality effects rather than fitting cross-cycle dynamics, decomposing a transition operator, or replaying mode-targeted controls.","source_ids":["S4"]},{"name":"Feedback Loop and Bias Amplification in Recommender Systems","similarity":"Uses offline repeated interaction to measure how recommender updates amplify bias and alter diversity and group outcomes.","remaining_difference":"It simulates algorithm-level feedback effects and does not identify a local cross-stratum Jacobian, invariant modes, eigengaps, or modal control sensitivities.","source_ids":["S5"]},{"name":"Fairness of Exposure in Dynamic Recommendation","similarity":"Targets long-run exposure imbalance under repeated recommendation updates and evaluates a dynamic intervention against static mitigation.","remaining_difference":"Its intervention directly adjusts cumulative item-exposure constraints; it does not discover or damp coupled latent propagation modes. It is the most important operational comparator.","source_ids":["S6"]},{"name":"Optimal Control Synthesis of Closed-Loop Recommendation Systems over Social Networks","similarity":"Frames recommendation as a controlled feedback system and uses spectral conditions to assess stabilizing versus pathological behavior under engagement and diversity tradeoffs.","remaining_difference":"It is a theoretical continuous-time opinion-dynamics model for social and commerce networks, not an empirically fitted, deidentified library stratum-transition Jacobian with residual, conditioning, representation, and policy-authority gates.","source_ids":["S7"]}],"distinctive_claim_remaining":"On held-out library update cycles, a regularized Jacobian over policy-approved exposure and engagement strata will reveal reproducible coupled modes that are missed by aggregate and one-stratum audits, and offline controls targeted to consequential modes will improve the exposure-disparity–retrieval-utility frontier relative to both ordinary monitoring and direct cumulative-exposure constraints, without breaching preregistered residual, conditioning, eigengap, privacy, or worst-stratum error limits.","confidence":"MODERATE"},"implementation_evidence":{"support":"MODERATE","rationale":"The component methods—offline feedback simulation, repeated-round exposure measurement, dynamic direct constraints, spectral stability analysis, and policy-controlled auditing—are credible. The candidate's core estimation step is not yet demonstrated: eight aggregate update cycles provide at most seven transitions, and fitting on six cycles may be underidentified for a multistratum Jacobian unless each cycle contains defensible independent transition units or strong regularization. Vendor audit access, catalog/query covariates, stratum definitions, deidentification, legal review, mode conditioning, and out-of-sample reproducibility remain unresolved. No patron-facing experiment is needed for the first test.","source_ids":["S1","S4","S5","S6","S7"]},"scores":{"meaningful_impact":{"score":4,"rationale":"Library evidence shows potentially material exposure disparities, while general recommender evidence connects feedback with diversity and minority-group harms; cross-library prevalence and realized impact remain unmeasured.","source_ids":["S3","S4","S5"]},"stakeholder_pull":{"score":4,"rationale":"ALA now explicitly calls for scheduled review of discovery and recommendation visibility distortions, and the Danish study calls for structural bias evaluation. A committed institutional partner or funder is still absent.","source_ids":["S1","S4"]},"incremental_advantage":{"score":3,"rationale":"Coupled modes could reveal interactions missed by per-stratum controls, but dynamic cumulative-exposure mitigation already addresses long-run imbalance and must be beaten empirically.","source_ids":["S5","S6","S7"]},"distinctiveness_plausibility":{"score":3,"rationale":"No exact library-specific fitted-Jacobian modal-control workflow was found, but dynamic feedback mitigation, spectral feedback analysis, eigensystem-based ranking, and closed-loop recommendation control make the proposal an adjacent composition rather than a wholly new method class.","source_ids":["S6","S7"]},"technical_implementability":{"score":2,"rationale":"Standard numerical tools are available, but the proposed six-cycle fit may be statistically underidentified, and nonstationarity, ill-conditioning, small eigengaps, catalog changes, and query-mix confounding can invalidate modal interpretation.","source_ids":["S5","S6","S7"]},"adoption_authority_feasibility":{"score":3,"rationale":"Library administrators, governing bodies, consortia, and accountable discovery owners are credible authorizers, but actual vendor audit rights, local policy approval, privacy review, and control over usage logs must be secured.","source_ids":["S1","S2"]},"evidence_readiness":{"score":2,"rationale":"The evaluation can be bounded and offline, but decisive evidence requires proprietary longitudinal logs, catalog and query covariates, vendor configuration history, and a named library partner.","source_ids":["S1","S4"]},"safety_net_benefit":{"score":4,"rationale":"A no-deployment retrospective replay, explicit residual and utility stops, retained non-personalized access paths, human authority, and rollback to the existing ranker provide a strong safety net if governance requirements are followed.","source_ids":["S1"]},"scalability":{"score":3,"rationale":"The analytical workflow is computationally modest and potentially reusable, but heterogeneous vendors, metadata, update schedules, policies, catalog composition, and log-retention rules limit transfer without local re-estimation.","source_ids":["S1","S2","S8"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"10K_TO_50K","scope":"A six-to-eight-week partnered data-feasibility and retrospective analysis: data inventory, privacy review, stratum specification, catalog/query controls, identifiability check, baseline and direct-constraint implementation, regularized modal fit if estimable, held-out replay, and decision memo.","confidence":"MODERATE","assumptions":["Existing deidentified logs and update snapshots can be accessed without purchasing new data.","Approximately 0.15 FTE-year of data-science labor plus limited librarian, engineer, privacy, and legal review is sufficient.","The BLS May 2024 data-scientist median is treated as a labor anchor and adjusted conceptually for 2026 benefits and overhead.","If only seven aggregate transitions exist, work stops at the identifiability assessment rather than forcing a modal estimate."],"source_ids":["S1","S8"]},"initial_deployment_startup":{"band_2026_usd":"50K_TO_250K","scope":"Production-grade data pipeline, repeatable decomposition and residual tests, access controls, documentation, dashboards, vendor integration, governance review, staff training, and rollback tooling for one library system.","confidence":"LOW","assumptions":["The library controls ranking configuration or has cooperative vendor support.","No replacement discovery platform or major data warehouse is required.","Roughly 0.5 to 1.5 resource-equivalent FTE-years are distributed across data science, engineering, librarianship, privacy, security, and project governance."],"source_ids":["S1","S2","S8"]},"operational_launch":{"band_2026_usd":"250K_TO_1M","scope":"A separately authorized, time-limited patron-facing pilot across multiple branches or discovery surfaces, including community consultation, accessibility and legal reviews, prospective monitoring, incident response, independent evaluation, vendor change work, and rollback coverage.","confidence":"LOW","assumptions":["A live pilot proceeds only after the offline claim passes.","Costs include multidisciplinary staff time and vendor/integration work but not procurement of an entirely new discovery service.","Public notice, human oversight, alternative access paths, and enhanced monitoring are maintained throughout launch."],"source_ids":["S1","S2","S8"]},"annual_recurring":{"band_2026_usd":"50K_TO_250K","scope":"Scheduled re-estimation, gap/drift/residual monitoring, data-quality checks, privacy and access reviews, librarian interpretation, vendor coordination, incident response, reporting, and periodic independent audit for one system.","confidence":"LOW","assumptions":["Monitoring consumes roughly 0.3 to 1.0 resource-equivalent FTE-year across several roles.","Compute is modest relative to labor because the state is aggregated by approved strata.","Material vendor fees, litigation, or platform replacement are excluded."],"source_ids":["S1","S8"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"Direct digital-library and public-library studies show ranking-position effects and material exposure disparities, while repeated-round recommender research shows amplification and long-run exposure failures.","source_ids":["S3","S4","S5","S6"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"ALA identifies library administrators, governing bodies, consortia, and accountable local staff as responsible for reviewing AI-enabled discovery and recommendation; NISO establishes library discovery responsibilities. No specific partner commitment has yet been obtained.","source_ids":["S1","S2"]},"distinct_testable_incremental_claim":{"status":"YES","reason":"The candidate can be tested against ordinary aggregate/per-stratum auditing and dynamic direct-exposure constraints using held-out prediction, disparity, utility, residual, conditioning, eigengap, and worst-stratum metrics.","source_ids":["S5","S6","S7"]},"bounded_next_evidence_step":{"status":"YES","reason":"A no-deployment retrospective study over the specified eight cycles is bounded, provided it begins with an identifiability gate and stops rather than fitting an unsupported multistratum Jacobian.","source_ids":["S1","S5","S6"]},"no_unresolved_safety_or_authority_stop":{"status":"UNCERTAIN","reason":"The design excludes individual profiling and live deployment, but no named library has confirmed lawful log access, privacy and security approval, vendor audit rights, accountable policy ownership, or acceptable stratum definitions. ALA guidance makes these prerequisites rather than optional details.","source_ids":["S1"]},"credible_cost_scope_and_range":{"status":"YES","reason":"The four ranges are scoped by stage, labor intensity, and explicit assumptions, using an official data-scientist wage benchmark. Confidence is limited because local vendor, legal, engineering, and public-sector overhead prices were not obtained.","source_ids":["S1","S8"]}},"next_evidence_step":"Conduct a partnered, preregistered, no-deployment retrospective study using the proposed eight update cycles. First determine whether the six fitting cycles contain enough defensible independent transition observations to estimate the declared state dimension; if they provide only a handful of aggregate transitions, classify the Jacobian as non-identifiable and stop. Otherwise fit a regularized operator on six cycles and validate on two, with bootstrap mode stability and catalog-availability and query-mix covariates. Compare: (A) existing aggregate and one-stratum monitoring, (B) a direct cumulative-exposure constraint modeled on dynamic discrepancy minimization, (C) the modal replay, and (D) a negative-control explanation based only on catalog availability and contemporaneous query intent. Preregister held-out transition error and residual structure, cumulative exposure disparity, nDCG or recall, worst-stratum error, condition number, eigengap, mode-angle stability, and privacy/representation review outcomes. Falsify the candidate if no consequential mode reproduces out of sample; the negative control explains propagation; ordinary or direct-constraint methods match or beat the modal equity–utility frontier; residuals or conditioning exceed budget; modes swap materially; or governance approval fails.","blocking_evidence":["No named library, consortium, or vendor has committed the required longitudinal logs, catalog snapshots, query-mix covariates, configuration history, or staff time.","The proposed eight-cycle design may yield too few independent transitions to identify a multivariable Jacobian.","Library-specific cross-cycle propagation beyond catalog availability, seasonal demand, and query intent has not been demonstrated.","No evidence yet shows modal replay outperforms dynamic direct-exposure constraints on held-out equity and retrieval utility.","Local privacy law, retention policy, deidentification adequacy, vendor audit rights, and accountable authorizer approval are unresolved.","No accepted thresholds exist yet for residual error, conditioning, eigengap, mode drift, utility loss, or worst-stratum harm."],"research_disposition":"PARTNERED_RESEARCH_PROGRAM","world_novelty_boundary":"This bounded search found library bias audits, repeated-round exposure mitigation, feedback-loop simulations, and spectral closed-loop recommendation control, but did not establish an exact prior implementation of the candidate's library-stratum Jacobian decomposition and mode-targeted replay. The search is not a systematic review and does not measure world novelty, patentability, freedom to operate, market size, cross-library prevalence, or realized impact.","arm":"SENTINEL","candidate_version":0,"controller_recommendation":{"action":"STOP_EMPIRICAL_RESEARCH_NEEDED","repairable":false,"material_progress_observed":true,"progress_targets":["Secure a named library or consortium partner, accountable discovery-policy owner, vendor cooperation, and documented privacy, security, accessibility, legal, and community-governance approval for retrospective log analysis.","Demonstrate that the available update history supplies enough independent transitions for the chosen state dimension, or falsify the design as underidentified before eigendecomposition.","Preregister strata, catalog and query controls, comparators, utility and disparity endpoints, residual and conditioning budgets, eigengap and mode-stability thresholds, and halt rules.","Run the held-out comparison against ordinary monitoring, dynamic direct-exposure constraints, and catalog/query negative controls; continue only if the modal method yields reproducible incremental advantage without unacceptable utility, privacy, representation, or worst-stratum harm."],"reason":"Web evidence supports the problem, stakeholder need, adjacent methods, governance requirements, and a bounded study design, but the candidate's decisive claims require proprietary longitudinal library data and live partner authorization. Whether a cross-cycle operator is identifiable, whether its modes reproduce, and whether modal control beats direct dynamic constraints cannot be resolved by further bounded web search. This is therefore an empirical stop, and repairable is false as required for a stop recommendation."}}