{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp11_mechanism_context_external20_20260804","research_id":"eoa_inverse_innovation_exp11_external_scrutiny_20260804","cell_id":"layer_decay_and_expiration_management__systems_cybernetics","opaque_id":"layer_decay_and_expiration_management__systems_cybernetics__B","search_lanes":{"direct_problem":{"queries":["adaptive control multiple models obsolete models switching recurring operating regimes model set pruning","industrial control obsolete controller model version lifecycle archive restore safety","control model library pruning obsolete models recurring regimes reactivation"],"source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5"],"no_result_note":null},"closest_prior_art":{"queries":["variable structure multiple model adaptive estimation model set pruning merging obsolete models recurring modes","dynamic model set adaptation multiple model control model pruning activation deactivation","model set adaptation multiple model adaptive control adding deleting models"],"source_ids":["SRC1","SRC3","SRC4","SRC5"],"no_result_note":null},"historical_terminology":{"queries":["multiple model adaptive control model library reuse recurring fault modes model management","Improving transient response adaptive control systems multiple models switching Narendra Balakrishnan 1994","likely-model set algorithm deleting unlikely models activating models mode transitions"],"source_ids":["SRC1","SRC2","SRC3"],"no_result_note":null},"products_practices_standards":{"queries":["ISO 10007 configuration management obsolete baseline archive control systems","MathWorks Multiple MPC Controllers different operating regions common approach","MLflow Model Registry archived model versions lifecycle stages restore official","industrial model management lifecycle archive deploy rollback control model product"],"source_ids":["SRC4","SRC5","SRC6","SRC7","SRC8"],"no_result_note":null},"non_english_regional":{"queries":["模型 生命周期 控制系统 旧版本 归档 回滚 模型注册","模型版本生命周期管理 退役 Archive 再利用 删除","gestion du cycle de vie modèles commande adaptative archivage versions modèle obsolète","Lebenszyklus Modelle Regelungstechnik alte Regler Version archivieren Modellverwaltung"],"source_ids":["SRC7"],"no_result_note":null},"composition_subproblems":{"queries":["model registry dependency check before delete model version lineage deployment safe deletion","archive restore test model versions replay rollback validation model registry","retired model inventory ongoing monitoring rollback audit model risk management","control model library pruning obsolete models recurring regimes reactivation"],"source_ids":["SRC3","SRC5","SRC7","SRC8"],"no_result_note":null}},"sources":[{"source_id":"SRC1","title":"Improving Transient Response of Adaptive Control Systems Using Multiple Models and Switching","url":"https://www.researchgate.net/publication/3021995_Improving_transient_response_of_adaptive_control_system_using_multiple_models_and_switching","publisher":"IEEE Transactions on Automatic Control; author-uploaded copy hosted by ResearchGate","date_or_year":"1994","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Multiple-model adaptive control and runtime selection of the model that best approximates the plant were established by 1994.","Adaptive control can exhibit oscillatory, large-amplitude transient tracking errors.","This work does not provide lifecycle states, archival, dependency-aware retirement, or restoration testing."]},{"source_id":"SRC2","title":"Fault Tolerant Flight Control System Design Using a Multiple Model Adaptive Controller","url":"https://journals.sagepub.com/doi/10.1243/09544100JAERO360","publisher":"Institution of Mechanical Engineers / SAGE Publications","date_or_year":"2009 (first published online 2008-12-11)","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Multiple models can represent different flight conditions and aircraft-fault situations.","Selecting and switching to the proper fault model are significant control issues.","Conventional frequent mode switching can make the system fluctuate and deteriorate aircraft performance."]},{"source_id":"SRC3","title":"Efficient Dynamics Estimation with Adaptive Model Sets","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC8098078/","publisher":"IEEE Robotics and Automation Letters; full text archived by PubMed Central","date_or_year":"2021","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Robotic decision, planning, and control systems face changing dynamics, terrains, sensing faults, and actuation faults.","Maintaining a fixed large candidate-model set can be computationally expensive.","The demonstrated adaptive-model-set algorithm removes low-probability models, expands the active set when current models cannot explain measurements, and improved efficiency without compromising measured performance in simulations and a hardware experiment.","Variable Structure Multiple Model algorithms are established terminology for choosing a changing subset from an existing model bank."]},{"source_id":"SRC4","title":"Multiple MPC Controllers","url":"https://www.mathworks.com/help/mpc/ref/multiplempccontrollers.html","publisher":"MathWorks","date_or_year":"Current documentation accessed 2026-08-04","source_type":"FIRST_PARTY_PRODUCT","language":"English","claims_supported":["A commercial control toolbox supports runtime selection of an active controller from candidate MPC controllers designed for different operating points.","Switching among controllers for different operating regions is described as a common control approach.","Inactive controllers remain involved in state estimation for bumpless transfer, and a finite candidate set can be tested thoroughly.","The product documentation does not describe lifecycle validation decay, supersession, archival restoration, or dependency-gated retirement."]},{"source_id":"SRC5","title":"Guideline E-23 – Model Risk Management (2027)","url":"https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/guideline-e-23-model-risk-management-2027","publisher":"Office of the Superintendent of Financial Institutions, Canada","date_or_year":"2025-09-11; effective 2027-05-01","source_type":"OFFICIAL_GUIDANCE","language":"English","claims_supported":["Official guidance requires risk-proportionate model inventories, explicit owners, reviewers and approvers, model versions, dependencies, approved uses, limitations, monitoring status, and review dates.","Monitoring should examine model performance, usage, inputs, external dependencies, scope, drift, thresholds, deterioration, and contingency plans.","Formal decommissioning is prescribed for performance problems or obsolete data or methodology.","A retired model and documentation should be retained for a set period as a benchmark or fallback, with downstream effects monitored.","This package closely anticipates the proposal's governance lever, but it is guidance for regulated financial models rather than regime-specific physical control artifacts."]},{"source_id":"SRC6","title":"ISO 10007:2017 Quality Management — Guidelines for Configuration Management","url":"https://www.iso.org/cms/live/live/en/sites/isoorg/contents/data/standard/07/04/70400.html","publisher":"International Organization for Standardization","date_or_year":"2017; confirmed 2023","source_type":"OFFICIAL_STANDARD","language":"English","claims_supported":["Configuration management across the support lifecycle from concept to disposal is standardized practice.","The public abstract is generic and does not establish the proposal's control-specific stale-selection mechanism or its archive-replay test."]},{"source_id":"SRC7","title":"模型版本生命周期管理 (Model Version Lifecycle Management)","url":"https://support.enos-iot.com/docs/ai-analytics/zh_CN/2.2.0/mi_hub/managing_model_lifecycle.html","publisher":"Envision Digital, EnOS Enterprise Analytics Platform","date_or_year":"Updated 2024-07-18","source_type":"FIRST_PARTY_PRODUCT","language":"Simplified Chinese","claims_supported":["A regional industrial analytics product implements candidate, production, and archived model-version states.","Retirement removes production deployment without deleting the version; archived versions receive no traffic or compute resources.","An archived version can be returned to candidate status and redeployed.","Deletion is limited to candidate or retired versions and is documented as unrecoverable."]},{"source_id":"SRC8","title":"Model Registry Workflows","url":"https://www.mlflow.org/docs/latest/ml/model-registry/workflow/","publisher":"MLflow Project","date_or_year":"Current documentation accessed 2026-08-04","source_type":"FIRST_PARTY_PRODUCT","language":"English","claims_supported":["A widely available model registry manages versions, lineage metadata, aliases, tags, validation status, deployment references, archival, and deletion.","Earlier stage-based workflows moved versions through staging, production, and archived states; current aliases and tags provide more flexible deployment and validation references.","Production versions can be archived before deletion, while registered-model deletion is irrevocable.","The documentation establishes implementation feasibility for lifecycle metadata and callable-versus-archived states, but not control-regime mismatch detection, dependency-safe deletion, or sandbox restoration fidelity."]}],"problem_evidence":{"status":"PARTLY_SUPPORTED","finding":"Public evidence supports the component hazards: multi-regime controllers select among model/controller candidates; poor selection or frequent switching can cause oscillatory or degraded transients; large fixed model sets impose cost; and official model-risk guidance recognizes drift, changed scope or dependencies, deterioration, and obsolete methodology as reasons for monitoring or retirement. The retained sources do not directly document the proposal's complete initiating condition—successive control regimes leaving superseded artifacts callable or operator-visible because lifecycle states are absent—or an incident caused specifically by that condition.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5"],"uncertainty":"The exact prevalence and consequence of stale callable control artifacts remain unknown without an audit or incident study in an operating multi-regime control installation."},"adopter_evidence":{"status":"SUPPORTED","finding":"Identifiable adopters and authorizers exist at the role and organization-class level. MathWorks targets owners of multi-operating-point MPC systems; OSFI assigns responsibilities to regulated institutions and explicit model owners, reviewers, users, and approvers; EnOS and MLflow expose lifecycle operations to model-platform operators. For safety-relevant physical control, final retirement authority would still need to be assigned by the installation's accountable owner and safety-assurance process.","source_ids":["SRC4","SRC5","SRC7","SRC8"],"uncertainty":"No retained source identifies a named industrial control installation that has agreed to run this exact intervention."},"implementation_evidence":{"status":"PARTLY_SUPPORTED","finding":"The main primitives are implementable and separately demonstrated: regime detection and controller switching, likelihood-gated active model subsets, version registries, validation tags, archive states, reversible redeployment, owner/approver metadata, dependency records, and disciplined decommissioning. No retained source demonstrates the full control-specific composition, especially semantic archive-restore replay, rare-regime exception handling, dependency-safe quarantine, and a stale-selection outcome comparison.","source_ids":["SRC3","SRC4","SRC5","SRC7","SRC8"],"uncertainty":"Technical restoration of a stored artifact does not prove that its data, dependencies, assumptions, and control semantics can be reconstructed faithfully."},"prior_art":{"disposition":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"Variable Structure Multiple Model and Adaptive Model Set estimation","source_ids":["SRC3"],"same_problem":true,"same_causal_lever":true,"overlap":"Adapts the active model bank under changing dynamics, removes low-probability models, and expands it when active models no longer explain observations; this directly addresses irrelevant-model competition and regime mismatch.","remaining_difference":"It manages the online estimation subset, not a governed history of model and control-law versions with provenance, supersession, retention exceptions, dependency gates, quarantine, audit, and tested restoration."},{"name":"OSFI risk-based model lifecycle and decommissioning guidance","source_ids":["SRC5"],"same_problem":true,"same_causal_lever":true,"overlap":"Requires inventory, risk rating, owner and approver identity, approved-use and dependency metadata, drift and scope monitoring, review triggers, formal retirement, fallback retention, and downstream-effect checks.","remaining_difference":"It governs financial-institution models generally and neither controls a plant nor tests regime-specific eligibility, switching stability, or rare-regime recovery."},{"name":"Model-registry lifecycle states in EnOS and MLflow","source_ids":["SRC7","SRC8"],"same_problem":false,"same_causal_lever":true,"overlap":"Implements version identity, production or deployment references, validation metadata, archival, later reuse, and guarded separation between retirement and irreversible deletion.","remaining_difference":"These registries do not determine whether a controller artifact matches the observed plant regime, check live control dependencies, or verify sandbox replay before reactivation."},{"name":"Multiple-model adaptive control and Multiple MPC Controllers","source_ids":["SRC1","SRC4"],"same_problem":false,"same_causal_lever":false,"overlap":"Maintains candidate regime models or controllers and selects an active controller using plant fit, measurements, or an operating-region switch; candidate controllers can be tested offline.","remaining_difference":"This is runtime selection rather than lifecycle governance and does not distinguish current, superseded, quarantined, or recoverably archived artifacts."}],"contrastive_claim_remaining":"For an operating multi-regime controller that already performs runtime best-fit or operating-region selection, adding explicit lifecycle metadata plus regime-fit, dependency, exception, and replay-restoration gates will reduce selections or recommendations from superseded or context-invalid artifacts relative to fixed-bank, adaptive-model-set, and latest-certified-only comparators, without materially increasing failure to recover valid recurring rare-regime behavior.","contrastive_claim_falsifier":"A preregistered shadow evaluation falsifies the increment if lifecycle gating does not reduce stale-eligibility selections, conflicting recommendations, switching or instability proxies, or review burden versus the comparators, or if it materially increases false rejection or failed restoration of valid rare-regime artifacts.","confidence":"MODERATE","search_limitations":"This was an exactly eight-source, six-lane public-web review. It included older control terminology, primary control research, official guidance, a current ISO standard page, products, Simplified Chinese documentation, and component combinations. It was not an exhaustive patent, source-code, procurement, incident-database, or paywalled-standards review; the full text of ISO 10007 was unavailable on its public landing page. Absence of an exact integrated source is therefore not world-novelty evidence."},"researchability_gates":{"externally_supported_problem":{"status":"INDETERMINATE","rationale":"The literature directly supports changing-regime selection problems, oscillatory switching consequences, irrelevant-model-set cost, model drift, and disciplined retirement. It does not directly establish that real adaptive control installations commonly accumulate superseded callable artifacts because lifecycle states are missing, which is the proposal's specific initiating problem.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5"]},"identifiable_adopter_or_authorizer":{"status":"PASS","rationale":"Control-system owners and platform operators are identifiable adopter classes, while official guidance explicitly identifies model owners, reviewers, users, and approvers. A concrete installation would still have to name its accountable safety authority before any live retirement.","source_ids":["SRC4","SRC5","SRC7","SRC8"]},"distinct_testable_incremental_claim":{"status":"PASS","rationale":"Despite close component prior art, the retained sources do not test the integrated claim that governance-grade lifecycle eligibility and replay-tested recoverability improve stale-selection outcomes in a multi-regime physical controller while preserving rare-regime recovery.","source_ids":["SRC3","SRC4","SRC5","SRC7","SRC8"]},"bounded_next_evidence_step":{"status":"PASS","rationale":"A read-only inventory and historical or sandbox replay on one noncritical subsystem can compare fixed-bank, adaptive-set, latest-certified-only, and proposed lifecycle policies without modifying eligibility, gains, actuators, or production outputs. Existing controller and registry primitives make that step technically bounded.","source_ids":["SRC3","SRC4","SRC5","SRC7","SRC8"]},"no_unresolved_safety_or_authority_stop":{"status":"PASS","rationale":"The proposed first step is observational and reversible. No hard deletion or live control change is required; later decommissioning must use named owners and approvers, retain fallback artifacts, inspect dependencies and downstream effects, and remain subject to independent safety approval.","source_ids":["SRC5","SRC7","SRC8"]},"adequate_search_evidence":{"status":"PASS","rationale":"All six required lanes were searched adversarially, including historical control terms, products and a standard, Simplified Chinese and European-language queries, and combinations involving dependency-safe deletion, archive restoration, revalidation, and recurring regimes. Exactly eight retained sources were opened; they span seven publisher or project organizations and include primary research, official guidance, an official standard, and first-party products.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5","SRC6","SRC7","SRC8"]}},"strict_success":false,"screen_survival":false,"remaining_research_value":"MODERATE","recommended_next_step":"With the accountable owner and safety reviewer, preregister a 30-day read-only shadow study on one noncritical subsystem. Snapshot the registry; inventory artifact identity, version, age, validation regime, assumptions, dependencies, current callability, successor, owner, holds, and rare-regime value; replay sampled current and archived artifacts in a sandbox; and compare fixed-bank or current practice, adaptive-model-set gating, latest-certified-only, and the proposed lifecycle policy. Primary outcomes should be stale-eligible selections, conflicting recommendations, mode-switch count and instability proxies, restoration success, rare-regime false rejection, dependency violations, and reviewer effort. Halt without changing live state if a used artifact is hidden, reconstruction fails, or a valid rare-regime artifact is rejected.","world_novelty_boundary":"This bounded search supports only a research-screen judgment. It cannot establish world novelty, patentability, freedom to operate, market size, routine adoption across industries, realized safety or performance impact, or the absence of undiscovered prior art."}