{"schema_version":1,"research_id":"eoa_inverse_innovation_exp03_external48_20260801","source_assessment_id":"eoa_inverse_innovation_exp03_opportunity320_20260801","cell_id":"invariant_mode_decomposition_design__library_information_science","selection_stratum":"MIDDLE_BAND_AUDIT_Q2","search_queries":["site:dl.acm.org recommender systems feedback loop exposure diversity longitudinal simulation","recommender systems dynamical systems eigenvalue stability feedback loop exposure matrix control","spectral analysis recommender feedback dynamics Jacobian recommendation system","fair recommender dynamic exposure constraints long-term fairness feedback control paper","library recommender system fairness exposure bias study digital library","\"recommender system\" \"Jacobian\" feedback loop eigenvalues","\"recommender systems\" \"linear dynamical system\" control exposure","\"spectral\" \"feedback loop\" recommender systems fairness exposure","\"mode decomposition\" recommender system dynamics","digital library recommendation feedback loop popularity bias longitudinal","\"dynamic mode decomposition\" \"recommendation\" exposure fairness","\"modal control\" recommender system fairness","patent recommender system feedback loop fairness exposure dynamic control eigenmode","site:bls.gov/news.release occupational employment wages data scientists software developers librarians compliance officers May 2025"],"sources":[{"source_id":"S1","title":"Feedback Loop and Bias Amplification in Recommender Systems","publisher":"Proceedings of CIKM 2020 / 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 iterative experiments found that several recommender algorithms amplified popularity bias over repeated feedback cycles.","The simulations found declining catalog coverage, shifting represented preferences, homogenization, and greater effects for the studied minority group.","The authors explicitly distinguish simulated iterations from chronological production evidence."]},{"source_id":"S2","title":"A Study of Position Bias in Digital Library Recommender Systems","publisher":"arXiv / real-world Sowiport and JabRef study","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 digital-library recommendations found that rank position significantly affected clicks independently of relevance.","Top-ranked click-through rates exceeded the paper's hypothetical unbiased rates by 53% in Sowiport and 87% in JabRef."]},{"source_id":"S3","title":"Bias in Book Recommendation: A Case Study on the Danish Public Libraries","publisher":"Centrum Wiskunde & Informatica / ECIR 2026","url":"https://ir.cwi.nl/pub/36328","source_class":"PRIMARY_RESEARCH","publication_date":"2026","accessed_at":"2026-08-02","claims_supported":["A study of Booklens, used in Danish public libraries, evaluated popularity and author-nationality exposure across 10,000 book prompts.","Popularity strongly influenced recommendations, and some configurations produced up to a 40% relative exposure decrease for less-popular nationalities.","The study supports systematic bias analysis in library recommender evaluation but is not longitudinal modal evidence."]},{"source_id":"S4","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":"n.d.","accessed_at":"2026-08-02","claims_supported":["ALA recommends reviewing AI-enhanced discovery, ranking, recommendation, and personalization at implementation, after major changes, and on a schedule for visibility and representation problems.","ALA assigns roles to library workers, administrators, governing bodies, consortia, and parent institutions and calls for local policies, human accountability, vendor review, and regular audits.","The guidance calls for privacy, security, legal, accessibility, data-minimization, retention, consent, disclosure, and human-review safeguards and preservation of non-personalized access paths."]},{"source_id":"S5","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":["The paper models exposure fairness when recommendation models and input data update through ongoing feedback.","It reports that repeatedly applying a static mitigation can fail to achieve long-run exposure fairness.","Its dynamic adaptation improved long-run exposure fairness while maintaining accuracy on a real-world dataset."]},{"source_id":"S6","title":"Deconvolving Feedback Loops in Recommender Systems","publisher":"NeurIPS 2016","url":"https://neurips.cc/virtual/2016/poster/7151","source_class":"PRIMARY_RESEARCH","publication_date":"2016","accessed_at":"2026-08-02","claims_supported":["This spectral-method prior art analyzes iterative recommender feedback and measures recommender influence on a user-item rating matrix.","It attempts to identify affected items and recover intrinsic ratings from a single snapshot rather than estimating cross-cycle exposure modes or controlling them."]},{"source_id":"S7","title":"Dynamic Mode Decomposition with Control","publisher":"arXiv / SIAM Journal on Applied Dynamical Systems","url":"https://arxiv.org/abs/1409.6358","source_class":"PRIMARY_RESEARCH","publication_date":"2014-09-22","accessed_at":"2026-08-02","claims_supported":["DMD with control fits data-driven input-output models from state and actuation snapshots.","It separates estimated unforced dynamics from control effects and supplies a technical analogue for modal sensitivity and control replay.","The paper does not apply the method to recommender fairness or libraries."]},{"source_id":"S8","title":"On Dynamic Mode Decomposition: Theory and Applications","publisher":"arXiv / Journal of Computational Dynamics","url":"https://arxiv.org/abs/1312.0041","source_class":"PRIMARY_RESEARCH","publication_date":"2013-11-29","accessed_at":"2026-08-02","claims_supported":["DMD can be defined as eigendecomposition of an approximating linear operator fitted from snapshots.","The paper identifies linear-consistency and rank-deficiency pitfalls and connects DMD to system-identification methods.","These limitations are directly relevant to fitting a multivariate operator from very few cycle transitions."]}],"problem_evidence":{"support":"MODERATE","rationale":"Longitudinal recommender research demonstrates feedback-driven popularity amplification and loss of diversity, while two real digital-library or public-library studies demonstrate position and book-exposure biases [S1,S2,S3]. This supports the broader underexposure-feedback problem. It does not establish the candidate's specific claim that consequential coupled subject, format, branch, and creator-group modes occur across update cycles in a library or evade coordinate-level audits.","source_ids":["S1","S2","S3"]},"stakeholder_evidence":{"support":"MODERATE","rationale":"ALA guidance explicitly calls for recurring review of library discovery, ranking, and recommendation for patterns that reduce topic, author, viewpoint, language, or community visibility, and identifies accountable institutional participants [S4]. The Danish-library study independently argues for systematic bias evaluation [S3]. No library was found requesting modal analysis, committing logs, or authorizing this candidate.","source_ids":["S3","S4"]},"prior_art":{"proximity":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"Fairness of Exposure in Dynamic Recommendation","similarity":"Addresses exposure unfairness over repeated model and feedback updates, tests a dynamic intervention against static mitigation, and evaluates fairness-accuracy tradeoffs.","remaining_difference":"It adapts exposure-fair re-ranking directly; it does not fit a cross-stratum transition operator, decompose growth or decay modes, or target controls by modal sensitivity.","source_ids":["S5"]},{"name":"Deconvolving Feedback Loops in Recommender Systems","similarity":"Uses spectral methods to detect recommender influence created by iterative feedback and analyzes real and synthetic recommendation data.","remaining_difference":"It deconvolves a single user-item snapshot and ranks affected items; it does not estimate cycle-to-cycle stratum dynamics or perform mode-targeted control replay.","source_ids":["S6"]},{"name":"WO2024249880A1, Implementing and Maintaining Feedback Loops in Recommendation Systems","similarity":"Describes offline metrics, predictive models, repeated feedback-loop simulation, bias-harm prediction, model comparison, and mitigation for recommendation systems.","remaining_difference":"The searched disclosure predicts harms from offline metrics and simulations but does not disclose the candidate's policy-approved stratum state, fitted Jacobian eigendecomposition, spectral-gap validity checks, or modal control comparison.","source_ids":["S1"]},{"name":"Dynamic Mode Decomposition with Control","similarity":"Provides the general data-driven operator, modal decomposition, actuation separation, and input-output control machinery that most closely matches the candidate's mathematical mechanism.","remaining_difference":"It is domain-general systems methodology, with no library discovery, exposure-fairness objective, direct-constraint rival, or library governance layer.","source_ids":["S7","S8"]}],"distinctive_claim_remaining":"The bounded difference is whether a low-rank, cross-cycle operator over policy-approved library-resource strata yields reproducible modes whose targeted offline control replay predicts better held-out exposure-equity and retrieval-utility outcomes than both coordinate monitoring and direct per-stratum constraints. The components are established separately; their library-specific composition and incremental value were not found in this search.","confidence":"HIGH"},"implementation_evidence":{"support":"MODERATE","rationale":"DMD and DMD-with-control establish that snapshot data can support an approximating operator, modal analysis, and input-output modeling [S7,S8], while dynamic exposure-control research supplies a recommender-specific comparator [S5]. However, the proposed six fitting cycles yield only five one-step transition pairs. As a mathematical inference from the rank limitations documented in [S8], a full state operator is not identifiable when the effective state dimension exceeds five unless additional independent transitions, strong structural restrictions, or regularization are introduced.","source_ids":["S5","S7","S8"]},"scores":{"meaningful_impact":{"score":3,"rationale":"Digital-library studies show material position and book-exposure disparities, and longitudinal recommender experiments show amplification and diversity loss [S1,S2,S3]. The prevalence and magnitude of coupled cross-cycle modes in actual libraries remain unknown.","source_ids":["S1","S2","S3"]},"stakeholder_pull":{"score":3,"rationale":"ALA's explicit call for recurring discovery and recommendation audits establishes professional demand for visibility-risk oversight [S4], but no prospective adopter or demand for modal control was found.","source_ids":["S4"]},"incremental_advantage":{"score":3,"rationale":"Cross-stratum modes could reveal combinations missed by coordinate monitoring, but dynamic exposure control and offline feedback-loop analysis already exist [S5,S6]. Only the proposed held-out rival comparison can establish added value.","source_ids":["S5","S6","S7"]},"distinctiveness_plausibility":{"score":3,"rationale":"No searched source combined library strata, fitted cross-cycle modes, modal sensitivity replay, direct-constraint comparison, and library governance. Closely adjacent dynamic-fairness, spectral-feedback, and DMD-control art substantially narrows the remaining claim.","source_ids":["S5","S6","S7","S8"]},"technical_implementability":{"score":2,"rationale":"The underlying methods are implementable, but five training transitions are generally inadequate for an unrestricted multistratum Jacobian and stable eigensystem. Rank, conditioning, spectral separation, uncertainty, and nonstationarity must be resolved before the specified analysis is credible [S7,S8].","source_ids":["S7","S8"]},"adoption_authority_feasibility":{"score":3,"rationale":"ALA recognizes library administrators, governing bodies, consortia, staff experts, and parent institutions as relevant decision and review actors [S4]. Feasibility depends on local platform control, vendor audit access, log access, and governance capacity, none of which is confirmed.","source_ids":["S4"]},"evidence_readiness":{"score":2,"rationale":"A safe comparison is conceptually clear, but there is no data partner, no fixed decision thresholds, and the eight-cycle design may be structurally underidentified. Existing studies demonstrate suitable metrics and offline comparisons but do not validate this dataset design [S1,S5,S8].","source_ids":["S1","S5","S8"]},"safety_net_benefit":{"score":4,"rationale":"A deidentified offline-only test with no patron-facing change, predeclared halts, retained baseline access, and accountable review aligns well with ALA privacy, human-oversight, audit, and non-personalized-access guidance [S4].","source_ids":["S4"]},"scalability":{"score":3,"rationale":"The fitted state model can be computationally small and reusable, and related dynamic exposure-control methods operate on recommendation data [S5,S7]. Scaling across libraries still requires local strata, metadata mapping, privacy review, platform access, and repeated validation.","source_ids":["S4","S5","S7"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"50K_TO_250K","scope":"Conduct an identifiability and data-quality audit; privacy-review and prepare eight historical cycles; define policy-approved strata; control catalog availability and query mix; preregister rank, residual, conditioning, utility, representation, and uncertainty rules; compare coordinate, low-rank modal, and direct-constraint models; and report held-out results without deployment.","confidence":"LOW","assumptions":["Existing logs and catalog snapshots are accessible without new patron data collection.","The work uses a compact team comprising data-science, ranking-engineering, discovery-librarian, privacy/compliance, and evaluation effort.","The cost includes loaded labor, governance coordination, secure computation, software, documentation, and independent review.","If effective state rank exceeds available transitions, the work stops at the identifiability finding rather than forcing a model."],"source_ids":["S4","S8"]},"initial_deployment_startup":{"band_2026_usd":"250K_TO_1M","scope":"Only after favorable offline evidence and separate authorization, engineer bounded controls and monitoring in a staging or production-adjacent environment; integrate logging, privacy and accessibility controls, audit records, rollback, utility tests, and collection-policy review.","confidence":"LOW","assumptions":["The library or consortium can modify its discovery service or obtain vendor cooperation.","Existing deployment, observability, identity-access, security, and rollback infrastructure can be extended.","The band includes engineering, data, software, security, privacy, accessibility, librarian, governance, vendor-coordination, and evaluation labor.","It excludes replacement of the underlying discovery platform."],"source_ids":["S4","S7"]},"operational_launch":{"band_2026_usd":"250K_TO_1M","scope":"Run a separately authorized, limited launch only after staging validation, with prospective utility, exposure, residual, drift, privacy, accessibility, subgroup-error, incident-response, and rollback monitoring plus independent evaluation.","confidence":"LOW","assumptions":["Live deployment is not part of the next evidence step.","A later launch would retain non-personalized access paths and prohibit individual modal scoring or inferred protected traits.","Costs include launch engineering, service operations, monitoring, governance, communications, evaluation, and contingency capacity.","No major hardware purchase is required."],"source_ids":["S4"]},"annual_recurring":{"band_2026_usd":"50K_TO_250K","scope":"Maintain secure pipelines and dashboards; investigate alerts; periodically refit and validate the operator; reassess strata and catalog effects; perform privacy, accessibility, collection-policy, and vendor reviews; and test rollback after material changes.","confidence":"LOW","assumptions":["Monitoring is integrated with an existing discovery service.","Recurring effort is dominated by specialist labor and coordination rather than dedicated hardware.","Material ranking, catalog, vendor, or policy changes trigger additional validation.","The band reflects resource-equivalent cost rather than a quoted product price."],"source_ids":["S4"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"Independent longitudinal recommender research and real library studies support feedback-amplified exposure or ranking bias, although the exact coupled modal form remains unverified [S1,S2,S3].","source_ids":["S1","S2","S3"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"ALA guidance makes library administrators, governing bodies, staff experts, consortia, and parent institutions credible authorization and review actors for discovery-system audits [S4]. No specific institution has committed.","source_ids":["S4"]},"distinct_testable_incremental_claim":{"status":"YES","reason":"The remaining claim is explicitly falsifiable by comparing held-out coordinate prediction, direct exposure constraints, and low-rank modal prediction/control under identical equity and utility criteria; adjacent art defines credible rivals [S5,S6,S7].","source_ids":["S5","S6","S7"]},"bounded_next_evidence_step":{"status":"YES","reason":"A retrospective, deidentified, no-deployment test can first falsify identifiability and propagation, then compare models and offline control replay while leaving two cycles held out [S1,S8].","source_ids":["S1","S8"]},"no_unresolved_safety_or_authority_stop":{"status":"YES","reason":"The first step can remain offline, preserve the existing service, minimize and deidentify logs, require accountable privacy and collection review, and stop on representation or privacy failures, consistent with ALA guidance [S4].","source_ids":["S4"]},"credible_cost_scope_and_range":{"status":"YES","reason":"The broad bands enumerate labor, data, software, secure computing, compliance, accessibility, coordination, engineering, monitoring, and evaluation. They remain low-confidence resource estimates because no institution-specific architecture, staffing, vendor, or data-condition evidence was available.","source_ids":["S4","S7","S8"]}},"next_evidence_step":"Run a preregistered, zero-deployment retrospective feasibility and prevalence test on the eight deidentified cycles. Before fitting, calculate the effective state dimension and stop as not estimable if it exceeds the information in the five training transitions unless a predeclared low-rank restriction is justified. Fit on cycles 1–6: (a) a coordinate lag-one baseline controlling catalog availability and query mix and (b) the restricted modal operator; keep cycles 7–8 sealed. Falsify the problem claim if the modal model does not improve predeclared held-out exposure prediction or if modes fail conditioning, residual, alignment, and spectral-separation rules. Only after that comparison, replay identical permissible controls against direct per-stratum constraints; falsify the intervention claim if modal targeting does not improve held-out equity without breaching the utility floor or produces larger residual harms. Make no patron-facing change.","blocking_evidence":["Whether an actual library has enough consistently defined historical update cycles and independent transitions to identify the proposed operator.","Whether exposure propagation remains after catalog availability, query mix, seasonal demand, and policy changes are controlled.","Whether any consequential mode is stable under resampling, alternative state definitions, and the two held-out cycles.","Whether modal targeting beats direct constraints on the same held-out equity, utility, and residual-harm criteria.","Whether approved strata have adequate support without privacy leakage, trait reification, or masking catalog omissions.","Whether a library or vendor will provide logs, platform documentation, engineering access, and accountable review.","Whether preregistered residual, conditioning, spectral-gap, alignment, utility, privacy, and representation thresholds can be agreed before analysis."],"research_disposition":"PROBLEM_PREVALENCE_STUDY","world_novelty_boundary":"This was a bounded web search through 2026-08-02 across recommender feedback loops, dynamic exposure fairness, spectral feedback analysis, DMD/control, digital-library bias, library governance, and patents. It found substantial adjacent prior art but no searched source implementing the exact library-specific composition. That absence is not a world-novelty, patentability, or freedom-to-operate conclusion."}