{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp03_full320_20260801","cell_id":"invariant_mode_decomposition_design__library_information_science","trajectory_id":"R","attempt_index":0,"archetype_slug":"invariant_mode_decomposition_design","domain_slug":"library_information_science","decision":"CANDIDATE","problem_id":"feedback_amplified_subject_exposure_imbalance","causal_lever_id":"modal_control_of_discovery_feedback","proposal":{"problem":"Digital-library ranking and recommendation updates can repeatedly amplify coupled patterns of subject, format, branch, or creator underexposure while aggregate engagement remains acceptable. Item-level audits then miss the combinations of exposure deviations that persist or grow across update cycles.","actors_substrate":["digital-library patrons","cataloged resources and represented creators","search and recommendation services","ranking engineers","collection and discovery librarians"],"observable_state":"Per update cycle, a state vector records normalized exposure and engagement deviations across policy-approved subject, format, branch, and creator-group strata; the local transformation is the fitted Jacobian mapping one cycle's deviations into the next.","consequence":"A weakly damped or growing combination of strata can systematically narrow discovery, even when no single stratum or average metric crosses an alert threshold.","affected_objective":"Equitable, diverse, and relevant discovery of library resources without materially degrading retrieval utility.","structural_mapping":[{"archetype_element":"Coupled transformation","domain_realization":"A locally fitted update operator links current exposure and engagement deviations to deviations after the next ranking-model refresh. HYPOTHESIS: this operator is sufficiently stable over a bounded window to support modal analysis.","claim_kind":"HYPOTHESIS"},{"archetype_element":"Invariant directions","domain_realization":"Modes are combinations of subject, format, branch, and creator-group deviations that approximately preserve their shape across update cycles.","claim_kind":"INFERENCE"},{"archetype_element":"Scalar response","domain_realization":"Each eigenvalue estimates whether its exposure-deviation mode grows, decays, persists, or oscillates within the local window.","claim_kind":"INFERENCE"},{"archetype_element":"Action relevance","domain_realization":"Modes are prioritized by persistence, discovery-equity consequence, sensitivity to permissible ranking controls, and reconstruction contribution—not by present magnitude alone.","claim_kind":"INFERENCE"},{"archetype_element":"Residual and drift discipline","domain_realization":"Held-out reconstruction error, residual structure, mode rotation, and spectral-gap changes determine whether the modal account remains usable.","claim_kind":"INFERENCE"}],"component_map":[{"component":"Transformation Scope","status":"adapted","domain_realization":"One query-display-engagement-model-update cycle, represented by a local deviation-update Jacobian."},{"component":"State-Vector Definition","status":"direct","domain_realization":"Normalized exposure and engagement deviations for approved resource strata, with missingness and catalog availability recorded separately."},{"component":"Invariant Mode Basis","status":"direct","domain_realization":"Eigenvectors of the fitted local update operator, traced back to contributing strata."},{"component":"Modal Gain Spectrum","status":"direct","domain_realization":"Eigenvalues, uncertainty intervals, and conditioning diagnostics for all fitted modes."},{"component":"Dominant Mode Selection Rule","status":"adapted","domain_realization":"Retain modes exceeding preregistered persistence, consequence, sensitivity, or reconstruction thresholds."},{"component":"Stable/Unstable Mode Partition","status":"direct","domain_realization":"Classify modes by gain relative to the discrete-cycle stability boundary, with an uncertainty band for marginal cases."},{"component":"Modal Intervention Map","status":"adapted","domain_realization":"Map permissible controls—diversification weight, exploration rate, and candidate-pool coverage—to their estimated effects on retained modes."},{"component":"Reconstruction Residual Check","status":"direct","domain_realization":"Reconstruct held-out exposure deviations and inspect both total error and structured subgroup residuals."},{"component":"Mode Drift Monitor","status":"direct","domain_realization":"Track mode rotation, reordered gains, residual growth, and changes after catalog or ranking updates."},{"component":"Interpretation Scope Contract","status":"direct","domain_realization":"Treat modes as local descriptive dynamics, not patron preferences, creator quality, or globally causal categories."},{"component":"Mode-Coupling Register","status":"direct","domain_realization":"Record near-degenerate, ill-conditioned, or sensitivity-linked modes that cannot be controlled independently."},{"component":"Local Linearization Window","status":"direct","domain_realization":"The traffic mix, catalog state, ranking version, and time interval used for fitting and validation."},{"component":"Spectral Gap Threshold","status":"direct","domain_realization":"A preregistered minimum separation between retained and dropped modes; crossing it suspends reduced-mode decisions."}],"mechanism_dispositions":[{"slug":"eigendecomposition_workflow","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Decompose the explicitly fitted local update operator to obtain the full mode basis and gains.","counterfactual_removal":"Without modes and gains, coupled persistence cannot be distinguished from coordinate-level fluctuation."},{"slug":"modal_sensitivity_sweep","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Perturb only policy-permitted ranking controls in replay and record modal and cross-modal responses.","counterfactual_removal":"The analysis would describe modes but could not identify controllable leverage or harmful cross-effects."},{"slug":"modal_stability_analysis","disposition":"selected_supporting","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Classify gains with uncertainty and bind every verdict to the fitted operating window.","counterfactual_removal":"Persistent or growing modes would lack a consistent escalation rule."},{"slug":"mode_shape_testing","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Empirical excitation is unnecessary for the first test because an operator is estimated directly from logged update cycles; patron-facing excitation is excluded.","counterfactual_removal":"No material change; held-out operator validation supplies the needed empirical check."},{"slug":"network_spectral_centrality_analysis","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Node-importance ranking answers a different question from cross-cycle amplification of exposure deviations.","counterfactual_removal":"No change to the causal chain."},{"slug":"power_iteration_probe","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Dominant-only recovery would hide near-dominant and equity-relevant low-magnitude modes; the bounded state permits full decomposition.","counterfactual_removal":"No change; complete decomposition is retained."},{"slug":"principal_component_analysis","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Variance directions need not describe the update transformation or persistence across cycles.","counterfactual_removal":"No change; covariance structure is not the causal lever."},{"slug":"reduced_order_model","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"A runnable surrogate is premature; the first test only estimates modes and simulates bounded control perturbations.","counterfactual_removal":"No change to diagnosis or the first test."},{"slug":"residual_reconstruction_test","disposition":"selected_load_bearing","contribution_type":"TEST_DESIGN","adaptation_or_rejection":"Select mode count using held-out reconstruction error plus subgroup-structured residual checks.","counterfactual_removal":"Compression could discard a small but consequential underexposure pattern without detection."},{"slug":"singular_value_decomposition","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Singular directions measure input-output amplification but not invariant repeated-cycle dynamics; ill-conditioned non-normal fits instead trigger suspension.","counterfactual_removal":"No change within the stated admissibility boundary."},{"slug":"spectral_decomposition_report","disposition":"selected_supporting","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Document traceability, uncertainty, prohibited interpretations, coupling, residuals, and validity window for reviewers.","counterfactual_removal":"Decision-makers could mistake local descriptive modes for stable patron traits or causal categories."},{"slug":"spectral_gap_monitor","disposition":"selected_load_bearing","contribution_type":"OPERATIONAL","adaptation_or_rejection":"Re-estimate gaps and mode alignment after each relevant model or catalog update.","counterfactual_removal":"A once-valid retained basis could continue governing decisions after its separation or identity disappears."}],"causal_chain":["Repeated ranking updates couple exposure and engagement deviations across resource strata.","The local update operator may contain combinations that decay slowly or grow although individual metrics look ordinary.","Full decomposition exposes those combinations and their cycle-to-cycle gains.","Sensitivity replay identifies permissible controls that damp consequential modes and reveals cross-mode effects.","A bounded control policy is considered only if held-out residual, utility, and affected-group constraints pass.","Gap, drift, and residual monitoring suspend the policy when the modal approximation loses validity."],"baseline":"Ordinary practice monitors aggregate retrieval utility and item- or category-level exposure, then adjusts boosts, quotas, or collections when a visible metric crosses a threshold.","nearest_rival":"Direct per-stratum exposure constraints or diversification re-ranking tuned without modeling cross-cycle coupled dynamics.","authority_safety":{"affected_parties":["patrons whose discovery options may change","creators and communities represented in cataloged resources","librarians responsible for discovery and collection policy","ranking-service operators"],"decision_authority":"The library's accountable discovery-policy owner, with privacy, accessibility, and collection-governance review; the analyst may recommend but not deploy.","authorized_first_step":"Run a preregistered offline test on deidentified logs from eight update cycles: fit on six, validate on two, compare modal control replay with the ordinary baseline and direct-constraint rival, and make no patron-facing changes.","excluded_actions":["inferring protected traits from behavior","assigning mode scores to individual patrons","suppressing lawful resources or viewpoints","live personalization experiments without separate approval and consent review","deploying controls when residual, utility, conditioning, or representation limits fail"],"halt_rollback":"Halt if held-out residual or subgroup error exceeds budget, modes are ill-conditioned, the retained spectral gap falls below threshold, utility breaches its floor, or a protected-harm review flags risk. Discard the proposed control and retain the existing ranking configuration; any later pilot must be independently authorized."}},"negative_tests":{"strongest_counterevidence":"Controlled replay may show that observed underexposure is fully explained by catalog availability and contemporaneous query intent, with no reproducible cross-cycle mode beyond noise.","analogy_break":"Library discovery is nonlinear, partly human-driven, and changed by catalog and policy interventions; eigenmodes of a local fitted operator are not intrinsic user preferences and may have no stable meaning outside the validation window.","failure_condition":"The operator is too nonstationary or ill-conditioned, no consequential mode is reproducible out of sample, or residual structure remains above the declared budget.","problem_falsifier":"After controlling for catalog availability and query mix, exposure deviations do not propagate across update cycles, and one-step stratum metrics explain held-out behavior as well as the modal model.","intervention_falsifier":"Reproducible modes exist, but mode-targeted controls fail to damp consequential modes, underperform direct constraints on held-out equity and utility, or create unacceptable cross-mode residual harms.","risks":["Reifying statistical modes as patron or community traits","Masking catalog omissions behind a ranking model","Noisy mode swapping near a small spectral gap","Trading retrieval relevance for apparent modal improvement","Privacy leakage from overly granular state strata","Feedback created by a later live intervention invalidating the fitted operator"]},"null_rationale":null,"classification":{"candidate_kind":"MECHANISM_COMPOSITION","prior_art_status":"UNSEARCHED","evidence_maturity":"HYPOTHESIS"},"revision_change_log":{"revision_kind":"ORIGINAL","prior_problem_id":null,"prior_causal_lever_id":null,"problem_changed":false,"causal_lever_changed":false,"conceptual_changes":[],"operational_changes":[],"repairs_addressed":[]},"confidence":0.78,"generator_notes":"Closed-book structural synthesis. Domain-specific propagation, stability, and intervention effects are hypotheses pending the bounded offline test."}