{"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 model switching obsolete models model bank recurring regimes archive","concept drift recurring concepts archive old classifiers model repository reuse","stale machine learning models model registry archive deprecate production callable governance"],"source_ids":["S1","S2","S3","S4"],"no_result_note":null},"closest_prior_art":{"queries":["unfalsified control controller archive candidate controllers performance monitoring","concept drift model pool pruning recurring concepts archive restore classifier dependency lifecycle","adaptive controller registry supersession metadata quarantine rollback rare regime model"],"source_ids":["S1","S2","S3","S6","S7"],"no_result_note":null},"historical_terminology":{"queries":["multiple model adaptive control switching supervisory controller bank pdf Narendra Balakrishnan 1997","gain scheduling operating regimes controller bank validity models switching historical","supervisory control multi-estimator switching adaptive control model falsification"],"source_ids":["S3","S4"],"no_result_note":null},"products_practices_standards":{"queries":["site:nist.gov AI model inventory decommission monitoring lifecycle model rollback","site:docs.aws.amazon.com sagemaker model registry model approval status archive rollback lineage","site:nist.gov SP 800-82 configuration management control systems rollback backup restore","site:learn.microsoft.com machine learning model registry archive restore model version shadow deployment rollback"],"source_ids":["S4","S5","S6","S7","S8"],"no_result_note":null},"non_english_regional":{"queries":["\"gestión del ciclo de vida\" modelos control adaptativo versiones obsoletas archivo restauración","\"Lebenszyklusmanagement\" Regelungsmodelle veraltet Archiv Wiederherstellung","適応制御 モデル バンク 古いモデル 再利用 コンセプトドリフト","\"controle adaptativo\" modelos antigos reutilização mudança de conceito recorrente"],"source_ids":["S8"],"no_result_note":"Portuguese, German, Japanese, and Spanish searches found adaptive-control, forgetting-factor, storage-lifecycle, and version-registry material but no closer non-English source integrating regime-aware runtime eligibility with preservation and restoration."},"composition_subproblems":{"queries":["model registry dependency check before deleting model version lineage archive restore test legal hold","control system previous configurations retention test validate changes separate test environment restore sampling","concept drift model pool pruning recurring concepts archive restore classifier dependency lifecycle","adaptive controller registry supersession metadata quarantine rollback rare regime model"],"source_ids":["S2","S5","S6","S7"],"no_result_note":null}},"sources":[{"source_id":"S1","title":"Handling Concept Drift via Model Reuse","url":"https://arxiv.org/abs/1809.02804","publisher":"arXiv; authors from Nanjing University","date_or_year":"2018","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Ignoring distribution change can sharply degrade model performance.","Condor maintains prior models and adapts their weights according to current fit.","Tests on synthetic and real-world streams support context-sensitive reuse of historical models.","Performance ceased improving when the retained model pool became excessively large, supporting managed rather than unlimited accumulation."]},{"source_id":"S2","title":"Recurring Concept Meta-learning for Evolving Data Streams","url":"https://arxiv.org/abs/1905.08848","publisher":"arXiv; authors from the University of Auckland and Télécom Paris","date_or_year":"2019","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Outdated training can bias a current classifier, while older concepts may later recur.","ECPF saves classifiers at drift, tests retained classifiers against current observations, and reuses a suitable one.","ECPF constrains its collection with similarity-based deletion and a fading score based on recency or use.","ECPF preserves an original classifier by adapting a copy, directly addressing recovery of recurring knowledge.","Experiments found accuracy and runtime benefits relative to comparison methods."]},{"source_id":"S3","title":"Multiple Model Adaptive Control. Part 2: Switching","url":"https://www.eng.yale.edu/controls/2001/mmodel2.pdf","publisher":"Yale University / International Journal of Robust and Nonlinear Control","date_or_year":"2001","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Supervisory adaptive control uses a family of candidate controllers, a multi-estimator, monitoring signals, and switching logic.","The controller selected at each instant corresponds to the process model that best fits available data.","Changing operating environments and component failures motivate real-time changes among controllers.","A controller family must retain sufficient variety so every admissible plant can be stabilized by at least one candidate."]},{"source_id":"S4","title":"Gain-Scheduled MPC","url":"https://www.mathworks.com/help/mpc/ug/gain-scheduling-mpc.html","publisher":"MathWorks","date_or_year":"accessed 2026-08-04","source_type":"FIRST_PARTY_PRODUCT","language":"English","claims_supported":["A deployed product supports switching among multiple MPC controllers for different operating points.","The workflow calls for identifying when a nominal controller loses robustness and testing controllers over the full operating range.","Switching without adequate warm-up or far from a controller's design operating point can cause sudden manipulated-variable changes.","Bumpless-transfer and full-range simulation requirements support the claimed transition-safety concern."]},{"source_id":"S5","title":"Register and work with models — Azure Machine Learning","url":"https://learn.microsoft.com/en-us/azure/machine-learning/how-to-manage-models?view=azureml-api-2","publisher":"Microsoft","date_or_year":"2026-01-27","source_type":"FIRST_PARTY_PRODUCT","language":"English","claims_supported":["Azure Machine Learning provides versioned model registration and lifecycle management.","A specific model version can be archived while remaining referenceable and usable in workflows.","Archival therefore reduces ordinary visibility but does not by itself guarantee runtime ineligibility."]},{"source_id":"S6","title":"NIST AI Risk Management Framework Playbook","url":"https://airc.nist.gov/docs/AI_RMF_Playbook.pdf","publisher":"National Institute of Standards and Technology","date_or_year":"2023","source_type":"OFFICIAL_GUIDANCE","language":"English","claims_supported":["NIST recommends continual monitoring for drift and decontextualization.","It recommends assigned responsibility and documented thresholds for superseding, bypassing, deactivating, or decommissioning systems and components.","It calls for upstream/downstream impact analysis, contingency options, preservation for forensic, regulatory, and legal review, and criteria for redeployment.","The guidance identifies organizational personnel responsible for maintenance, reverification, monitoring, updating, and decommissioning as authorizers or accountable adopters."]},{"source_id":"S7","title":"Guide to Industrial Control Systems (ICS) Security, NIST SP 800-82 Revision 2","url":"https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-82r2.pdf","publisher":"National Institute of Standards and Technology","date_or_year":"2015","source_type":"OFFICIAL_GUIDANCE","language":"English","claims_supported":["ICS configuration management should maintain, monitor, and document controlled changes and retain previous configurations.","The control overlay includes testing, validation, documentation, separate test environments, impact analysis, inventories, and restricted change authority.","Backup controls include reliability testing and sampled restoration; recovery guidance warns that restoring physical-system state may disrupt ongoing processes.","A senior organizational official accepts residual risk and authorizes system operation, establishing an identifiable authority path.","Laboratory testing is recommended where testing the operational ICS could be harmful."]},{"source_id":"S8","title":"Control de versiones de modelos con Model Registry","url":"https://docs.cloud.google.com/vertex-ai/docs/model-registry/versioning?hl=es-419","publisher":"Google Cloud","date_or_year":"2025-10-19","source_type":"FIRST_PARTY_PRODUCT","language":"Spanish","claims_supported":["Vertex AI Model Registry maintains multiple model versions with identifiers, aliases, status, descriptions, labels, and performance details.","A designated default version is preselected for prediction, while other versions remain selectable.","Individual versions can be evaluated, tested, batch-run, deployed, and compared, supporting a bounded replay or shadow assessment."]}],"problem_evidence":{"status":"PARTLY_SUPPORTED","finding":"External evidence supports the mechanism: changing regimes can invalidate models, selecting a controller outside its validated operating region can cause abrupt manipulated-variable changes, and irrelevant historical learning can degrade predictions. Research systems already detect drift and gate reuse of retained models. The search did not locate incident-frequency evidence showing that unmanaged callable historical control artifacts are a prevalent cause of real industrial oscillation or unsafe transitions, so the proposal's operational prevalence and effect magnitude remain hypotheses.","source_ids":["S1","S2","S3","S4","S5"],"uncertainty":"Most empirical evidence concerns streaming classifiers rather than safety-critical plant controllers; the control sources establish the hazard and selection architecture but not the frequency of stale-artifact incidents."},"adopter_evidence":{"status":"SUPPORTED","finding":"The adopter and authorizer are identifiable as the control-system asset owner or operating organization, with a senior authorizing or safety official accepting residual risk and personnel explicitly assigned to monitoring, reverification, change control, and decommissioning.","source_ids":["S6","S7"],"uncertainty":"Exact titles and separation-of-duty requirements vary by sector and jurisdiction."},"implementation_evidence":{"status":"PARTLY_SUPPORTED","finding":"The component capabilities are demonstrably implementable: model registries provide identity, version chronology, status, comparison, default selection, and archiving; recurring-concept systems provide fit-based reuse and collection pruning; and ICS guidance supplies retained baselines, separate test environments, change authorization, impact review, backup testing, and sampled restoration. No retained source demonstrates the complete proposed package—including dependency-aware retirement, rare-regime exception holds, reversible quarantine, and semantic restore testing—inside one live multi-regime control system.","source_ids":["S1","S2","S5","S6","S7","S8"],"uncertainty":"Registry archival may hide a model without preventing explicit use, and technical restoration does not prove that dependencies, assumptions, or safety cases remain valid."},"prior_art":{"disposition":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"Enhanced Concept Profiling Framework (ECPF)","source_ids":["S2"],"same_problem":true,"same_causal_lever":true,"overlap":"Detects drift, saves historical classifiers, evaluates them against current observations, reuses a fitting prior classifier, preserves originals through copying, and manages collection size through fading and similarity-based deletion.","remaining_difference":"It is a streaming-classification algorithm, not an accountable control-artifact lifecycle with callable/standby/archive states, dependency and assurance holds, quarantine, restoration tests, or actuator-safety outcomes."},{"name":"Multiple-model supervisory adaptive control","source_ids":["S3","S4"],"same_problem":true,"same_causal_lever":true,"overlap":"Maintains regime-specific models or controllers, measures current fit, switches eligibility at runtime, and tests switching over the operating envelope.","remaining_difference":"It assumes a designed candidate family and switching logic rather than managing accumulated versions, supersession, retention, archival, deletion, provenance, and restoration over a long operational lifecycle."},{"name":"NIST AI RMF supersede/deactivate/preserve workflow","source_ids":["S6"],"same_problem":true,"same_causal_lever":true,"overlap":"Monitors drift and decontextualization, assigns authority, uses thresholds to supersede or deactivate components, analyzes downstream effects, preserves evidence, and defines redeployment criteria.","remaining_difference":"It is general governance guidance and does not specify recurring-regime model selection, model-pool scoring, tested historical recovery, or comparative control outcomes."},{"name":"NIST ICS configuration-management and recovery controls","source_ids":["S7"],"same_problem":false,"same_causal_lever":false,"overlap":"Provides inventories, retained previous configurations, controlled and audited changes, impact analysis, isolated tests, backup verification, sampled restoration, recovery, and explicit authorization.","remaining_difference":"The controls govern system configurations and recovery rather than context-conditioned runtime eligibility of accumulated regime models or control laws."},{"name":"Cloud model registries","source_ids":["S5","S8"],"same_problem":false,"same_causal_lever":false,"overlap":"Provide stable identities, versions, metadata, status/default references, evaluation, comparison, and archival.","remaining_difference":"Azure archival remains explicitly callable, and neither product source establishes regime-mismatch gating, dependency-safe retirement, preservation exceptions, quarantine, or safe rare-regime restoration."}],"contrastive_claim_remaining":"For a multi-regime controller with multiple historical artifacts still callable, adding explicit regime-fit eligibility plus reversible quarantine, dependency/authority gates, and verified restoration will reduce prespecified stale-artifact selections and instability proxies relative to both current practice and latest-certified-only selection, without materially increasing failure to recover held-out recurring regimes.","contrastive_claim_falsifier":"In preregistered shadow replay, the lifecycle-gated policy fails to reduce stale selections or actuator/oscillation proxies versus the nearest rival, or it significantly increases missed valid selections or failed restoration in recurring rare regimes.","confidence":"HIGH","search_limitations":"The bounded search cannot reveal confidential industrial deployments or incident reports, and some standards and publisher archives may be incompletely indexed. The search establishes close mechanisms and component practice, not world novelty, patentability, freedom to operate, routine adoption, market size, or realized impact."},"researchability_gates":{"externally_supported_problem":{"status":"PASS","rationale":"Primary research and first-party control guidance support drift-induced obsolescence, context-sensitive model selection, and hazardous switching outside validated operating regions, although prevalence in deployed industrial controllers is not established.","source_ids":["S1","S2","S3","S4"]},"identifiable_adopter_or_authorizer":{"status":"PASS","rationale":"ICS and AI-risk guidance identify the operating organization, assigned lifecycle personnel, safety or risk officials, and senior authorizing official as concrete adopters and decision authorities.","source_ids":["S6","S7"]},"distinct_testable_incremental_claim":{"status":"PASS","rationale":"Close art separately covers fit-based historical-model reuse and governed lifecycle/recovery controls, but the control-specific combination retains a falsifiable comparative claim about stale selection, stability proxies, and rare-regime recovery.","source_ids":["S2","S3","S6","S7"]},"bounded_next_evidence_step":{"status":"PASS","rationale":"A read-only inventory and replay on one noncritical subsystem can compare policies without changing live eligibility or outputs; registries, model comparison, isolated testing, and sampled restoration are established capabilities.","source_ids":["S2","S7","S8"]},"no_unresolved_safety_or_authority_stop":{"status":"PASS","rationale":"The first step can remain read-only and isolated, with an immutable snapshot, no actuator changes, explicit owner and safety approval, and halt criteria for hidden dependencies or failed restoration. Safety considerations constrain later deployment but do not stop the proposed evidence-gathering step.","source_ids":["S6","S7"]},"adequate_search_evidence":{"status":"PASS","rationale":"The bounded adversarial search covered all six required lanes and retained exactly eight opened direct sources from seven publisher organizations, including three primary studies, two official NIST publications, and three first-party product sources, with one retained Spanish regional source.","source_ids":["S1","S2","S3","S4","S5","S6","S7","S8"]}},"strict_success":true,"screen_survival":true,"remaining_research_value":"MODERATE","recommended_next_step":"Preregister and run the proposed 30-day read-only test on one noncritical subsystem. Snapshot the registry; inventory every callable artifact and dependency; label validation regime, supersession, owner, current fit, and recovery value; then replay identical logged transitions under current selection, latest-certified-only, and lifecycle-gated policies. Primary endpoints should be stale-selection rate, conflicting recommendations, actuator-delta and oscillation proxies, false rejection of valid models, and successful sandbox restoration on held-out recurring regimes. Require independent safety review and stop on any undiscovered live dependency, non-faithful restoration, or rejection of a validated rare-regime model.","world_novelty_boundary":"This result is a bounded public-web prior-art and researchability assessment. It does not establish world novelty, patentability, freedom to operate, market size, routine adoption, or realized safety and performance impact."}