{"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:dl.acm.org recommender systems feedback loop exposure bias dynamic fairness ranking","public library discovery platform diversity recommendation equity strategic plan search ranking","site:oclc.org library discovery diversity recommendation algorithm bias","site:ala.org library privacy data ethics personalization recommendation systems","recommender systems dynamic fairness feedback loop exposure allocation repeated interactions paper","recommender feedback loops popularity bias longitudinal exposure empirical study paper","fair ranking amortized exposure constraints repeated rankings paper","spectral analysis recommender systems feedback loop eigenvalue control exposure fairness","\"Jacobian\" recommender system feedback loop exposure fairness","\"spectral\" \"feedback loop\" recommender systems fairness","\"eigenvalue\" recommender system stability feedback loop","library recommender system fairness diversity exposure bias study","dynamic mode decomposition snapshots rank data requirements Jacobian estimation paper","NIST AI RMF recommendation systems fairness privacy monitoring official","NISO privacy principles library systems official standard user data","public library recommendation system vendor analytics logs BiblioCommons privacy"],"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":"2026-06","accessed_at":"2026-08-02","claims_supported":["Automated discovery, ranking, recommendation, and personalization can reduce visibility or distort representation of topics, authors, viewpoints, languages, or communities.","ALA recommends review at implementation, after major changes, and on a schedule.","Libraries should preserve transparent non-personalized access paths, retain human accountability, protect usage logs, and assess vendor audit limitations and privacy."]},{"source_id":"S2","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","accessed_at":"2026-08-02","claims_supported":["A deployed non-personalized recommender used in Danish public libraries was evaluated using 10,000 books.","Popularity strongly affected recommendations, with up to a 40% relative exposure decrease for less-popular nationalities under some configurations.","The authors call for systematic bias analysis as a structural part of library recommender evaluation."]},{"source_id":"S3","title":"Feedback Loop and Bias Amplification in Recommender Systems","publisher":"ACM CIKM / 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 simulation found that recommender feedback can amplify popularity bias, reduce aggregate diversity, and homogenize experience.","The reported effect was stronger for the studied minority user group.","The study used 20 simulated iterations and notes that real-platform testing would be preferable."]},{"source_id":"S4","title":"Fairness of Exposure in Dynamic Recommendation","publisher":"ACM RecSys / arXiv","url":"https://arxiv.org/abs/2309.02322","source_class":"PRIMARY_RESEARCH","publication_date":"2023-09-05","accessed_at":"2026-08-02","claims_supported":["Dynamic exposure bias and long-run feedback amplification are already explicit research problems.","A dynamic adaptation of a static bias-mitigation method outperformed repeated static mitigation for long-term exposure fairness while maintaining recommendation accuracy.","Direct dynamic exposure mitigation is a close functional rival to the candidate."]},{"source_id":"S5","title":"On Dynamic Mode Decomposition: Theory and Applications","publisher":"Journal of Computational Dynamics / arXiv","url":"https://arxiv.org/abs/1312.0041","source_class":"PRIMARY_RESEARCH","publication_date":"2014","accessed_at":"2026-08-02","claims_supported":["Dynamic mode decomposition estimates modes and eigenvalues of an approximating linear operator from time-series snapshots.","The method can characterize low-order dynamics even where underlying dynamics are nonlinear.","Rank-deficient data and failure of linear consistency create documented pitfalls."]},{"source_id":"S6","title":"NISO Consensus Principles on Users’ Digital Privacy in Library, Publisher, and Software-Provider Systems","publisher":"National Information Standards Organization","url":"https://www.niso.org/publications/privacy-principles","source_class":"STANDARD","publication_date":"2015-12-10","accessed_at":"2026-08-02","claims_supported":["Library analytics require shared privacy responsibility, data minimization, security, anonymization, transparency, consent, and accountability.","Privacy obligations extend across libraries, publishers, and software providers."]},{"source_id":"S7","title":"The New BiblioCommons Analytics Platform Provides Actionable Insight into the Digital Library Experience","publisher":"BiblioCommons","url":"https://www.bibliocommons.com/news/the-new-bibliocommons-analytics-platform-provides-actionable-insight-into-the-digital-library-experience","source_class":"COMMERCIAL_FIRST_PARTY","publication_date":"2024","accessed_at":"2026-08-02","claims_supported":["A commercial library-discovery platform supplies anonymized digital-journey analytics, dashboards, training, consulting, and custom reports.","A named Marin County Free Library stakeholder expressed a longstanding need for online-patron-journey data beyond basic ILS data.","Analytics infrastructure and organizational demand exist, although the page does not establish access to ranking candidate sets, propensities, or model-update histories."]},{"source_id":"S8","title":"Strategic Plan 2026–2030","publisher":"Brooklyn Public Library","url":"https://www.bklynlibrary.org/strategicplan","source_class":"OFFICIAL_ORGANIZATION_DATA","publication_date":"2026","accessed_at":"2026-08-02","claims_supported":["Brooklyn Public Library plans personalized reading-interest recommendations and continued online-catalog refinement.","It identifies inclusive discovery, underrepresented voices, digital equity, and navigability of millions of resources as institutional priorities.","A large public-library system is therefore an identifiable prospective adopter, although it has not endorsed modal control specifically."]}],"problem_evidence":{"support":"STRONG","rationale":"The general and library-specific problem is visible. A Danish public-library recommender showed configuration-dependent popularity and nationality exposure disparities; longitudinal recommender research demonstrates feedback amplification, diversity loss, and unequal group effects; and current ALA guidance explicitly warns that automated discovery can reduce visibility or distort representation. Evidence does not yet show that a coupled, weakly damped cross-cycle mode exists in any particular target library or that aggregate metrics systematically conceal it.","source_ids":["S1","S2","S3","S4"]},"stakeholder_evidence":{"support":"STRONG","rationale":"ALA calls for recurring audits of ranking and recommendation, BPL plans personalized recommendations while prioritizing inclusive discovery, and a named Marin County librarian reports longstanding demand for better digital-journey analytics. These establish an authorizing professional body, prospective institutional adopter, and operational stakeholder need. None has requested eigendecomposition or modal control specifically.","source_ids":["S1","S7","S8"]},"prior_art":{"proximity":"SUBSTANTIAL_COLLISION","closest_analogues":[{"name":"Fairness of Exposure in Dynamic Recommendation","similarity":"Addresses repeated recommendation updates, long-run exposure amplification, offline evaluation, utility preservation, and dynamic mitigation—the candidate's central functional problem and comparator.","remaining_difference":"Uses adapted exposure-fairness mitigation rather than estimating a cross-stratum Jacobian, decomposing its modes, and damping selected coupled modes subject to conditioning, residual, and spectral-gap gates.","source_ids":["S4"]},{"name":"Feedback Loop and Bias Amplification in Recommender Systems","similarity":"Models iterative recommender-user feedback offline and measures amplification, diversity decline, homogenization, and unequal group effects.","remaining_difference":"Primarily diagnoses simulated aggregate dynamics; it does not identify a local stratum-transition operator or implement modal feedback control.","source_ids":["S3"]},{"name":"Danish Public Library Book-Recommender Bias Audit","similarity":"Direct library-domain evaluation of popularity and represented-nationality exposure under different recommender configurations.","remaining_difference":"It is a configuration-sensitive output audit, not a cross-cycle system-identification and control method.","source_ids":["S2"]},{"name":"Dynamic Mode Decomposition","similarity":"Provides the candidate's core method: fit an approximating linear operator from snapshots, extract modes and eigenvalues, and diagnose dynamics.","remaining_difference":"The source does not apply DMD to library discovery, exposure fairness, policy-authorized controls, or comparison with direct exposure constraints.","source_ids":["S5"]}],"distinctive_claim_remaining":"For a library discovery system with sufficiently many governed, deidentified update-cycle observations, a regularized cross-stratum transition operator will yield reproducible, well-conditioned modes that predict held-out propagation of exposure deviations; controls targeted at consequential modes will reduce persistent coupled underexposure more than aggregate monitoring and direct per-stratum constraints at an equivalent retrieval-utility floor, without structured residual or represented-group harm. Failure to reproduce modes, beat those comparators, or satisfy conditioning, residual, privacy, and utility thresholds falsifies the claim.","confidence":"HIGH"},"implementation_evidence":{"support":"MODERATE","rationale":"The component methods are implementable: modal decomposition is established, library analytics platforms can provide anonymized journey data, and current guidance supplies governance and monitoring expectations. The proposed dataset is the major defect: eight cycle-level state vectors provide only seven transitions, with six training cycles yielding at most five fitted transition pairs. A full Jacobian across numerous subject, format, branch, and creator strata is therefore underidentified or rank-deficient absent drastic prespecification, pooling, or regularization. Two held-out cycles cannot credibly establish stability across regimes. Counterfactual replay also requires candidate sets, positions, availability, query mix, model/configuration histories, and defensible response assumptions, none of which are verified as available.","source_ids":["S1","S5","S6","S7"]},"scores":{"meaningful_impact":{"score":4,"rationale":"If reproducible, the method could expose cumulative discovery narrowing hidden by aggregate utility and support earlier intervention in a public-information setting. Library-specific measured disparities and ALA concern establish consequence, but realized benefit is unmeasured.","source_ids":["S1","S2","S3"]},"stakeholder_pull":{"score":4,"rationale":"Libraries and their professional body express demand for inclusive discovery, personalized recommendations, recurring audits, and better analytics. Pull is for the outcome and governance, not the proposed modal method.","source_ids":["S1","S7","S8"]},"incremental_advantage":{"score":3,"rationale":"Coupled-mode detection could outperform independent stratum thresholds when deviations propagate jointly, but dynamic exposure mitigation already addresses the long-run problem and no comparative result supports an advantage.","source_ids":["S4","S5"]},"distinctiveness_plausibility":{"score":3,"rationale":"The exact composition of library-governed cross-stratum Jacobian estimation, modal sensitivity control, residual checks, and gap/drift suspension was not found. Its functional objective substantially overlaps established dynamic exposure-fairness work.","source_ids":["S2","S4","S5"]},"technical_implementability":{"score":2,"rationale":"Standard numerical tools can perform the decomposition, but the stated eight-cycle design is generally inadequate for a multistratum Jacobian and is especially vulnerable to rank deficiency, noise, mode swapping, and nonstationarity.","source_ids":["S5"]},"adoption_authority_feasibility":{"score":3,"rationale":"A library discovery-policy owner can authorize an offline audit, and ALA guidance supports human oversight and recurring review. Vendor control of logs, model internals, and configurations may prevent meaningful replay or control.","source_ids":["S1","S7","S8"]},"evidence_readiness":{"score":2,"rationale":"Problem and adjacent-method evidence are ready, but the incremental claim requires proprietary longitudinal logs and controlled replay. Required event fields, cycle count, data quality, intervention degrees of freedom, and counterfactual validity are unverified.","source_ids":["S2","S4","S7"]},"safety_net_benefit":{"score":4,"rationale":"The candidate's offline-only first step, utility floor, residual and conditioning gates, privacy review, suspension thresholds, and rollback to the existing ranker are useful safeguards aligned with library guidance. They remain design commitments rather than tested controls.","source_ids":["S1","S6"]},"scalability":{"score":3,"rationale":"Once log schemas and governance are standardized, offline matrix estimation and monitoring could be reused across systems. Scaling is constrained by vendor-specific telemetry, changing metadata, sparse strata, local policy definitions, and differing confidentiality law and practice.","source_ids":["S1","S6","S7"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"50K_TO_250K","scope":"A bounded partner-library data audit and offline prototype: governed state-vector specification, extraction of ranking and availability logs, privacy review, identifiability diagnostics, regularized operator fitting, bootstrap stability analysis, direct-constraint and aggregate baselines, replay, and a decision report.","confidence":"MODERATE","assumptions":["Existing infrastructure retains the necessary deidentified event, candidate-set, position, catalog-availability, query-mix, and configuration records.","Approximately 3–6 resource-equivalent person-months are required across data engineering, recommendation science, librarianship, privacy, and governance.","No patron-facing experiment, new production telemetry, or vendor contract amendment is included."],"source_ids":["S1","S5","S7"]},"initial_deployment_startup":{"band_2026_usd":"250K_TO_1M","scope":"After successful evidence, build a shadow-mode pipeline, governed stratum registry, secure feature and log processing, versioned replay environment, monitoring dashboards, vendor integration, accessibility/privacy assessment, and independently authorized limited pilot controls.","confidence":"LOW","assumptions":["A discovery vendor supplies documented APIs or exports and permits audit and configuration control.","Approximately 1.5–4 resource-equivalent staff-years plus security, legal, procurement, and integration effort.","No replacement of the integrated library system or discovery platform is included."],"source_ids":["S1","S6","S7"]},"operational_launch":{"band_2026_usd":"250K_TO_1M","scope":"Production hardening and institution-wide launch: parallel validation, incident and rollback procedures, staff training, documentation, community/governance review, accessibility testing, alert ownership, and staged activation across branches or collections.","confidence":"LOW","assumptions":["The pilot passes preregistered equity, utility, stability, and privacy gates.","Existing ranking infrastructure supports reversible bounded controls.","Launch covers one large library system rather than a national shared service."],"source_ids":["S1","S6","S8"]},"annual_recurring":{"band_2026_usd":"50K_TO_250K","scope":"Recurring data-quality checks, spectrum/residual/drift monitoring, quarterly or major-change audits, librarian review of represented strata, privacy and vendor review, incident response, retraining, and annual reporting.","confidence":"MODERATE","assumptions":["Approximately 0.5–1.5 resource-equivalent full-time staff plus modest compute and vendor support.","Major platform replacement, litigation, and collection-acquisition costs are excluded.","Monitoring frequency increases after model, metadata, topology, or policy changes."],"source_ids":["S1","S6","S7"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"Primary research documents recommender feedback amplification, and a 2026 public-library case measured substantial configuration-dependent exposure disparities; ALA explicitly recognizes visibility and representation risk in library discovery.","source_ids":["S1","S2","S3","S4"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"BPL is an identifiable prospective adopter planning personalized recommendations and inclusive discovery; ALA identifies library administrators, governing bodies, and workers as accountable reviewers; Marin County staff express analytics need.","source_ids":["S1","S7","S8"]},"distinct_testable_incremental_claim":{"status":"YES","reason":"The remaining claim is contrastive and falsifiable against aggregate monitoring and direct dynamic exposure constraints using held-out propagation, equity, utility, conditioning, residual, and stability outcomes.","source_ids":["S4","S5"]},"bounded_next_evidence_step":{"status":"YES","reason":"A preregistered, offline, no-patron-change identifiability-and-replay study can be time-boxed and has explicit comparators and stop criteria. It requires proprietary partner data.","source_ids":["S1","S4","S5"]},"no_unresolved_safety_or_authority_stop":{"status":"UNCERTAIN","reason":"Offline analysis is plausibly authorizable, but actual authority over vendor logs, lawful retention and reuse, anonymization adequacy, represented-stratum definitions, and configuration control has not been verified. Analysis must stop if those approvals or a non-personalized access path are unavailable.","source_ids":["S1","S6","S7"]},"credible_cost_scope_and_range":{"status":"YES","reason":"Broad resource-equivalent bands are scoped by distinct evidence, startup, launch, and recurring work packages with explicit labor and infrastructure assumptions. Confidence is limited because vendor pricing and log-access terms are unavailable.","source_ids":["S1","S6","S7"]}},"next_evidence_step":"Secure one accountable library and vendor/data-controller agreement, then run a six-week, offline, preregistered feasibility gate before any control efficacy claim. First inventory at least the stated eight update cycles and the candidate sets, positions, queries, catalog availability, engagement, model versions, and policy changes. On the candidate's six-cycle fit/two-cycle validation split, compare (A) aggregate and one-step per-stratum monitoring, (B) direct dynamic exposure constraints modeled on existing prior art, and (C) regularized modal control. Before replay, require that the training snapshot matrix supports the prespecified state dimension, that modes remain aligned under bootstrap/noise perturbations, and that conditioning, spectral-gap, and held-out residual thresholds pass. Falsify the proposed eight-cycle design if its state dimension exceeds supported rank, held-out modes swap or collapse, residuals retain structured stratum/query effects, or catalog availability and contemporaneous query mix explain propagation as well as the operator. Falsify incremental advantage if modal control does not improve prespecified persistent-exposure error over comparator B at the same NDCG or other retrieval-utility floor, or if any approved-stratum error, privacy review, or intellectual-freedom review breaches its limit. With only two held-out cycles, treat any pass as permission to collect or recover a longer longitudinal series, not as deployment evidence.","blocking_evidence":["Whether any willing library and discovery vendor can lawfully provide deidentified cycle-level logs with candidate sets, rank positions, query mix, catalog availability, engagement, and model/configuration histories.","Whether eight update cycles contain enough independent transitions to identify the prespecified multistratum operator; the stated six-cycle fit is likely rank-deficient for anything beyond a very small state.","Whether consequential modes reproduce out of sample and remain stable under bootstrap, regularization, metadata changes, query shifts, and near-degenerate eigenvalues.","Whether offline replay has defensible counterfactual response assumptions or randomized/exploration data sufficient to compare interventions.","Whether exposure deviations persist after controlling for catalog holdings, acquisition and withdrawal changes, language/metadata quality, branch availability, seasonality, and query intent.","Whether policy-approved creator and subject strata can be defined without inferring protected traits, reifying communities, or producing privacy leakage through small cells.","Whether the library—not merely its vendor—has authority to inspect and alter ranking behavior while preserving transparent non-personalized discovery paths.","No field evidence yet shows that modal control beats direct dynamic exposure constraints, preserves retrieval utility, or avoids cross-mode harms."],"research_disposition":"PARTNERED_RESEARCH_PROGRAM","world_novelty_boundary":"This bounded search does not measure world novelty, patentability, freedom to operate, market size, or realized impact. It found substantial functional prior art for dynamic exposure-fairness mitigation, established offline feedback-loop simulation, direct public-library bias auditing, and established modal decomposition. It did not find an exact published match for library-governed cross-stratum Jacobian decomposition with modal intervention, residual, spectral-gap, and drift suspension; absence from these searches is not evidence of world novelty.","arm":"SENTINEL","candidate_version":0,"controller_recommendation":{"action":"STOP_EMPIRICAL_RESEARCH_NEEDED","repairable":true,"material_progress_observed":true,"progress_targets":["Obtain written participation and data-use authority from a library discovery-policy owner, privacy officer, and relevant vendor or municipal/campus controller.","Produce a field-level log inventory proving availability and lawful use of candidate sets, positions, catalog availability, query mix, engagement, and configuration histories.","Resolve the eight-cycle identifiability defect by demonstrating a prespecified low-dimensional state supported by the training rank or by obtaining a substantially longer cycle history; do not infer a full multistratum Jacobian from five training transitions.","Preregister comparators, including aggregate/per-stratum monitoring and a direct dynamic exposure-constraint method, plus utility, equity, conditioning, residual, spectral-gap, drift, and privacy thresholds.","Demonstrate held-out mode reproducibility with bootstrap and negative-control tests, including controls for catalog availability, metadata changes, query composition, and seasonality.","Validate counterfactual replay assumptions or obtain exploration/randomization evidence adequate for intervention comparison.","Document librarian and affected-community governance of strata and meanings, prohibit individual patron scoring and protected-trait inference, and preserve a non-personalized discovery path.","Return only if modal control exceeds direct constraints on persistent-exposure outcomes at an equivalent utility floor without structured residual, subgroup, privacy, accessibility, or intellectual-freedom harm."],"reason":"Web evidence verifies the problem, stakeholder need, operational context, close prior art, and an exact falsifiable contrast. The decisive remaining questions require proprietary longitudinal logs, institutional authority, and offline testing. The stated eight-cycle design is likely underidentified for the proposed multistratum Jacobian, so further bounded web research cannot establish feasibility or advantage."}}