{"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control stale obsolete models switched systems model bank supervisory control outdated model selection","multiple model adaptive control obsolete model selection regime mismatch"],"source_ids":["SRC1","SRC2"],"no_result_note":null},"closest_prior_art":{"queries":["multiple model adaptive control model set reinitialization pruning performance monitoring archive recovery","adaptive control model library lifecycle deactivate archive restore stale model"],"source_ids":["SRC1","SRC2","SRC7"],"no_result_note":null},"historical_terminology":{"queries":["1990 multiple model adaptive control switching supervisory estimator unfalsified models falsification obsolete models","multiple model switching control model falsification hysteresis dormant models"],"source_ids":["SRC2","SRC3"],"no_result_note":null},"products_practices_standards":{"queries":["NIST AI RMF model inventory monitoring decommissioning model lifecycle official","ISO IEC 5338 AI system life cycle operation monitoring retirement model configuration management","MLflow model registry archived stage deprecated model version official documentation","SAS model manager champion challenger retire model governance"],"source_ids":["SRC4","SRC5","SRC6"],"no_result_note":null},"non_english_regional":{"queries":["control adaptatif multi-modèles commutation modèles obsolètes sélection régime","commande adaptative multimodèle commutation banc de modèles reconfiguration","control adaptativo multimodelo conmutación banco de modelos obsoletos"],"source_ids":["SRC7"],"no_result_note":"French/regional searching found a close multimodel switching-control thesis and broader results, but no source showing the full proposed lifecycle package in routine control-system use."},"composition_subproblems":{"queries":["control model registry supersession dependency safe deletion quarantine rollback replay rare operating regime","model bank pruning restore archived controller rare regime dependency audit trail","adaptive controller stale model detector certification archive reactivation sandbox replay"],"source_ids":["SRC4","SRC5","SRC6","SRC8"],"no_result_note":null}},"sources":[{"source_id":"SRC1","title":"Multiple Model Adaptive Control. Part 1: Finite Controller Coverings","url":"https://doi.org/10.1080/002071797224740","publisher":"International Journal of Control / Taylor & Francis","date_or_year":"1997","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Multiple-model adaptive control deliberately maintains a finite family of candidate plant models/controllers and uses a supervisory switching mechanism.","The model family is designed to cover a plant uncertainty set, so retaining multiple regime-specific artifacts is longstanding control practice.","The paper supports the same broad problem context and selection lever, but not artifact-age decay, archival restoration tests, dependency-gated retirement, or quarantine."]},{"source_id":"SRC2","title":"Supervisory Control of Families of Linear Set-Point Controllers—Part 1: Exact Matching","url":"https://doi.org/10.1109/9.661603","publisher":"IEEE","date_or_year":"1998","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Supervisory control evaluates a family of candidate controllers/models using monitoring signals and switches among them.","Hysteresis and dwell-time style logic address harmful switching and stability, establishing close prior art for mismatch detection and eligibility gating.","The focus is safe selection among candidates, not managed retirement, archival tiering, restoration testing, or preservation exceptions."]},{"source_id":"SRC3","title":"Adaptive Control with Multiple Models","url":"https://doi.org/10.1007/978-1-4471-6036-7","publisher":"Springer","date_or_year":"2014","source_type":"SECONDARY_RESEARCH","language":"English","claims_supported":["The control literature recognizes multiple-model adaptive control, switching, mixing, model-set design, and reinitialization as an established research family.","Historical terminology centers on model banks, supervisory switching, performance indices, reinitialization, and model-set adaptation rather than lifecycle or expiration management.","This synthesis makes a direct-phrase miss weak evidence of novelty."]},{"source_id":"SRC4","title":"Artificial Intelligence Risk Management Framework (AI RMF 1.0)","url":"https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf","publisher":"National Institute of Standards and Technology","date_or_year":"2023","source_type":"OFFICIAL_GUIDANCE","language":"English","claims_supported":["NIST calls for documented inventory, roles and responsibilities, ongoing monitoring, change management, incident response, and safe decommissioning of AI systems.","These governance controls support identifiable organizational authorization and lifecycle oversight for model-based decision systems.","The framework is generic AI risk guidance and does not establish the proposed control-memory mechanism or its causal effect."]},{"source_id":"SRC5","title":"ISO/IEC 5338:2023 Information technology—Artificial intelligence—AI system life cycle processes","url":"https://www.iso.org/standard/81118.html","publisher":"International Organization for Standardization","date_or_year":"2023","source_type":"OFFICIAL_STANDARD","language":"English","claims_supported":["ISO/IEC 5338 defines AI-system lifecycle processes and integrates them with software and systems lifecycle processes.","It establishes that operation, maintenance, configuration/change control, and retirement are standards-level concerns for AI systems.","The public standard page does not demonstrate the proposal's detailed combination or efficacy in adaptive control."]},{"source_id":"SRC6","title":"ML Model Registry","url":"https://mlflow.org/docs/latest/ml/model-registry/","publisher":"MLflow","date_or_year":"2026","source_type":"FIRST_PARTY_PRODUCT","language":"English","claims_supported":["A deployed product supports model versioning, aliases, tags, lineage, descriptions, and controlled promotion/deployment workflows.","Registry stages are deprecated in favor of aliases and environment-specific registered models, showing current product practice for separating identity, status, and deployment eligibility.","It offers adjacent identity and lifecycle machinery, but not dependency-safe control-law deletion, rare-regime restore validation, or proof of reduced instability."]},{"source_id":"SRC7","title":"Commande adaptative multimodèle : une approche par modèles locaux","url":"https://theses.hal.science/tel-00011338","publisher":"HAL theses / Université de Bordeaux 1","date_or_year":"2005","source_type":"PRIMARY_RESEARCH","language":"French","claims_supported":["French-language research uses the terms commande adaptative multimodèle and modèles locaux for adaptive control across different operating zones.","Local models and switching or fusion across operating conditions corroborate the multi-regime model-bank substrate outside English terminology.","The retained record did not show the full inventory-decay-quarantine-archive-restore package."]},{"source_id":"SRC8","title":"ModelDB: A System for Machine Learning Model Management","url":"https://doi.org/10.1145/2939502.2939516","publisher":"ACM","date_or_year":"2016","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Model-management research identifies versioning, lineage, comparison, deployment, and lifecycle handling as distinct operational problems.","The work supports registries and provenance as practical components for accumulated model artifacts.","It does not establish control-regime mismatch gating plus reversible retirement and rare-regime restoration as a tested package."]}],"problem_evidence":{"status":"PARTLY_SUPPORTED","finding":"The substrate and hazard mechanism are credible: multiple-model adaptive and supervisory control systems maintain candidate models/controllers and select among them using performance or mismatch signals, while governance and registry sources recognize inventories, status, lineage, monitoring, and retirement. However, the eight retained sources do not directly document incidents or measured instability caused specifically by superseded artifacts remaining callable after regime succession. The exact problem therefore remains a testable hypothesis rather than an established empirical failure mode.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC6","SRC7"],"uncertainty":"It is unclear how often real adaptive-control deployments expose uncertified superseded artifacts to live selectors or operators, and whether existing certification and supervisory logic already prevents the alleged stale selection."},"adopter_evidence":{"status":"SUPPORTED","finding":"An identifiable authorizer exists: the accountable owner/operator of the controlled system, with safety or assurance approval where artifacts affect safety. Standards and official guidance assign lifecycle, monitoring, change, risk, and retirement responsibilities to organizations, and product registries provide implementable custody points.","source_ids":["SRC4","SRC5","SRC6"],"uncertainty":"Precise legal authority and required independent approval vary by sector and jurisdiction; no particular adopter organization is evidenced."},"implementation_evidence":{"status":"PARTLY_SUPPORTED","finding":"Most ingredients are implementable using established model banks, supervisor monitoring, registries, lineage, lifecycle status, and decommissioning controls. A read-only registry inventory and offline replay is bounded and reversible. Evidence is missing for semantic restoration fidelity, dependency completeness, scoring calibration, and the claimed reduction in stale selections or instability relative to latest-certified-only control.","source_ids":["SRC1","SRC2","SRC4","SRC5","SRC6","SRC8"],"uncertainty":"The sources do not validate the complete integrated workflow in a live or shadow adaptive-control setting."},"prior_art":{"disposition":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"Multiple-model adaptive and supervisory switching control","source_ids":["SRC1","SRC2","SRC3","SRC7"],"same_problem":true,"same_causal_lever":true,"overlap":"Maintains regime-covering model/controller families, monitors candidate performance or mismatch, and gates which controller guides the plant; stability-aware switching and model-set adaptation directly address inappropriate selection.","remaining_difference":"It does not, in the retained sources, combine explicit supersession state, age/provenance scoring, dependency and assurance holds, active/standby/archive tiers, reversible quarantine, deletion markers, and tested restoration for rare regimes."},{"name":"AI lifecycle governance and decommissioning standards","source_ids":["SRC4","SRC5"],"same_problem":false,"same_causal_lever":true,"overlap":"Requires or recommends inventories, accountable roles, monitoring, change management, lifecycle processes, risk controls, and retirement/decommissioning.","remaining_difference":"Generic system governance does not specify stale regime-controller memory, performance-based control eligibility, replay restoration, or a causal comparison with a latest-certified-only policy."},{"name":"Operational ML model registries and model management","source_ids":["SRC6","SRC8"],"same_problem":false,"same_causal_lever":true,"overlap":"Provides stable model identities, versions, lineage, metadata, aliases/status, comparison, and controlled deployment or archival workflows.","remaining_difference":"These systems do not establish dependency-safe retirement of control laws, preservation of rare-regime controllability, safety-assurance exceptions, or lower instability from lifecycle gating."}],"contrastive_claim_remaining":"For an adaptive multi-regime controller whose historical artifacts can remain eligible, adding explicit supersession and regime-mismatch status plus dependency/assurance gates and reversible quarantine—while admitting archive restoration only after sandbox replay—will reduce selections of invalid-for-current-regime artifacts versus a latest-certified-only-plus-undifferentiated-archive rival, without materially reducing successful recovery when a previously seen rare regime recurs.","contrastive_claim_falsifier":"In a preregistered shadow comparison, the package fails to lower invalid selections or instability proxies relative to the nearest rival, or causes a material prespecified decrease in correct rare-regime recovery or restoration fidelity; an audit showing historical artifacts are already never eligible would also falsify the motivating problem locally.","confidence":"MODERATE","search_limitations":"Bounded public-web review of exactly eight sources; paywalled abstracts and public landing pages may omit implementation details. Search cannot establish world novelty, patentability, freedom to operate, market size, routine prevalence, or realized impact. Product documentation is mutable, and no sector-specific regulator or deployed incident dataset was retained."},"researchability_gates":{"externally_supported_problem":{"status":"PASS","rationale":"Primary control research establishes the multi-model substrate and that supervisory monitoring/gating is needed to prevent unsuitable candidate selection; registry and governance sources establish accumulated-version lifecycle concerns. Direct prevalence and incident evidence is limited, so support is mechanistic rather than epidemiological.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC6","SRC7"]},"identifiable_adopter_or_authorizer":{"status":"PASS","rationale":"The controlled-system owner/operator is identifiable, with safety/assurance approval for safety-relevant retirement; official lifecycle guidance and standards support accountable organizational ownership.","source_ids":["SRC4","SRC5"]},"distinct_testable_incremental_claim":{"status":"PASS","rationale":"Closest control prior art covers monitoring and selection, and lifecycle prior art covers registries and retirement, but the retained sources do not test the integrated reversible-retirement and validated-restore package against the stated rival. Its joint selection and recovery outcomes are falsifiable.","source_ids":["SRC1","SRC2","SRC4","SRC5","SRC6","SRC8"]},"bounded_next_evidence_step":{"status":"PASS","rationale":"A 30-day read-only inventory and offline replay on one noncritical subsystem can label supersession/regime fit, simulate both policies from logged data, and test archival restoration without changing live outputs.","source_ids":["SRC2","SRC4","SRC6","SRC8"]},"no_unresolved_safety_or_authority_stop":{"status":"PASS","rationale":"The first step is observational and sandboxed, preserves the registry snapshot, excludes live gain/actuator changes and hard deletion, respects incident and assurance holds, and has explicit halt conditions. Sector-specific approvals still must be obtained before live gating.","source_ids":["SRC4","SRC5"]},"adequate_search_evidence":{"status":"PASS","rationale":"All six lanes were searched adversarially using direct, historical, standards/product, French/regional, and component-combination terminology. Exactly eight retained direct sources span six publishers and include four primary studies, one official standard, one official guidance document, and one first-party product source.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5","SRC6","SRC7","SRC8"]}},"strict_success":true,"screen_survival":true,"remaining_research_value":"MODERATE","recommended_next_step":"Preregister and run the authorized 30-day shadow study: freeze a registry snapshot; enumerate every model, control law, threshold, causal map, owner, successor, dependency, certification basis, and known regime; score regime fit without changing eligibility; replay logged normal and rare-regime episodes under baseline, latest-certified-only, and proposed gating; sample archived artifacts for sandbox restoration; and compare invalid-selection rate, disagreement, switching/oscillation proxies, restoration fidelity, rare-regime recovery, reviewer time, and false-staleness rate. Halt on any hidden live dependency, failed restore, or rejection of a valid rare-regime artifact.","world_novelty_boundary":"This bounded search found adjacent prior art in multiple-model supervisory control, AI lifecycle standards, and model registries, but did not find the full package tested as stated. That is only a contrastive research gap within the retained search, not evidence of world novelty, patentability, freedom to operate, market size, or realized impact."}