{"schema_version":1,"research_id":"eoa_inverse_innovation_exp03_external48_20260801","source_assessment_id":"eoa_inverse_innovation_exp03_opportunity320_20260801","cell_id":"layer_decay_and_expiration_management__systems_cybernetics","selection_stratum":"DEPLOYABLE_PRIORITY","search_queries":["site:nist.gov AI RMF model decommissioning post deployment monitoring drift documentation","site:airc.nist.gov playbook decommission model monitoring drift model inventory","adaptive control multiple model library add remove models online model bank management","\"model set adaptation\" multiple model adaptive control pruning models","site:mlflow.org/docs model registry stages archived transition model version archive official","site:learn.microsoft.com model registry archive restore model version lifecycle","site:nasa.gov software configuration management archive previous versions rollback handbook PDF","NIST SP 800-82 Rev 3 industrial control systems change management test changes configuration official","site:bls.gov/ooh computer information technology software developers 2025 median pay"],"sources":[{"source_id":"S1","title":"AI Risk Management Framework Core","publisher":"National Institute of Standards and Technology","url":"https://airc.nist.gov/airmf-resources/airmf/5-sec-core/","source_class":"OFFICIAL_GUIDANCE","publication_date":"2023","accessed_at":"2026-08-02","claims_supported":["AI-system behavior should be monitored in production and evaluated under conditions similar to deployment.","Post-deployment plans should include override, decommissioning, recovery, and change management.","Limitations beyond development conditions and performance declines in deployment context should be documented."]},{"source_id":"S2","title":"AI RMF Playbook: Manage","publisher":"National Institute of Standards and Technology","url":"https://airc.nist.gov/airmf-resources/playbook/manage/","source_class":"OFFICIAL_GUIDANCE","publication_date":"2023","accessed_at":"2026-08-02","claims_supported":["Organizations should monitor components for drift and decontextualization.","Thresholds and protocols may temporarily or permanently bypass or deactivate components.","The Playbook recommends considering dependencies and downstream consequences, preserving components for review, providing contingency options, and defining criteria for redeployment."]},{"source_id":"S3","title":"Challenges to the Monitoring of Deployed AI Systems","publisher":"National Institute of Standards and Technology","url":"https://www.nist.gov/publications/challenges-monitoring-deployed-ai-systems-center-ai-standards-and-innovation","source_class":"PRIMARY_RESEARCH","publication_date":"2026-03-06","accessed_at":"2026-08-02","claims_supported":["Post-deployment monitoring is needed to detect unexpected behavior under dynamic conditions and verify continued reliability.","NIST's workshops and literature review found stakeholder agreement on monitoring need but fragmented, nascent methods and repeated practitioner requests for guidance."]},{"source_id":"S4","title":"Efficient Dynamics Estimation With Adaptive Model Sets","publisher":"IEEE Robotics and Automation Letters / PubMed","url":"https://pubmed.ncbi.nlm.nih.gov/33969182/","source_class":"PRIMARY_RESEARCH","publication_date":"2021-02-18","accessed_at":"2026-08-02","claims_supported":["Maintaining all candidate dynamics models at every time step can be computationally expensive.","A condition-responsive active subset was tested against fixed-set baselines in simulation and robot hardware.","The tested adaptive subset reduced computation without degrading estimation or tracking performance and sometimes improved mode prediction, but did not study artifact lifecycle governance."]},{"source_id":"S5","title":"NASA Software Engineering Requirements, NPR 7150.2D, Chapter 5","publisher":"National Aeronautics and Space Administration","url":"https://nodis3.gsfc.nasa.gov/npg_img/N_PR_7150_002D_/N_PR_7150_002D__Chapter5.pdf","source_class":"OFFICIAL_GUIDANCE","publication_date":"2022-03-08","accessed_at":"2026-08-02","claims_supported":["NASA requires configuration control of software records, code, data, tools, models, scripts, and versions.","The project manager must define control levels, change authority, configuration status records, audits, storage, release, and maintenance procedures.","Software risks and residual risks must be tracked and communicated."]},{"source_id":"S6","title":"SWE-081 — Identify Software Configuration Management Items","publisher":"NASA Software Engineering Handbook","url":"https://swehb.nasa.gov/spaces/7150/pages/16450453/SWE-081%2B-%2BIdentify%2BSoftware%2BCM%2BItems","source_class":"OFFICIAL_GUIDANCE","publication_date":"2017","accessed_at":"2026-08-02","claims_supported":["Version control permits access to previous versions when a problem is discovered.","Configuration identification supports control, status accounting, verification, and audit.","Incomplete configuration identification can contribute to defective products, delays, and maintenance cost."]},{"source_id":"S7","title":"Model Registry Workflows","publisher":"MLflow","url":"https://mlflow.org/docs/latest/ml/model-registry/workflow","source_class":"OFFICIAL_PRODUCT_DOCUMENTATION","publication_date":"not stated","accessed_at":"2026-08-02","claims_supported":["MLflow supports model versions, status tags, aliases, access-controlled environments, and reassignment of deployment references.","Its deprecated fixed stages included Production and Archived, illustrating established lifecycle-state practice.","MLflow deprecated fixed stages partly because they were too inflexible for mature workflows."]},{"source_id":"S8","title":"Azure CLI: az ml model","publisher":"Microsoft","url":"https://learn.microsoft.com/en-us/cli/azure/ml/model?view=azure-cli-latest","source_class":"OFFICIAL_PRODUCT_DOCUMENTATION","publication_date":"not stated","accessed_at":"2026-08-02","claims_supported":["Azure Machine Learning can archive and restore model containers or individual versions.","Archived models are hidden from default listings but remain referenceable and usable, showing that archival alone is not an execution-safety gate.","Restoration granularity depends on whether a container or individual version was archived."]}],"problem_evidence":{"support":"MODERATE","rationale":"Official guidance recognizes that deployment context changes can invalidate assumptions, calls for drift and decontextualization monitoring, and contemplates bypass or decommissioning. Primary robotics research also shows that keeping every candidate model active can impose computational and selection costs. However, the bounded search found no study measuring how often superseded control laws, thresholds, or causal maps remain callable or directly linking such stale eligibility to oscillation, mistuning, or unsafe transitions. The precise prevalence and consequence chain therefore remain unverified.","source_ids":["S1","S2","S3","S4"]},"stakeholder_evidence":{"support":"MODERATE","rationale":"NIST's 2026 practitioner work reports agreement on the need for post-deployment monitoring and repeated requests for methodological guidance. NASA requirements also assign concrete configuration-control duties and authority to project managers. This establishes governance pull, but not purchasing demand, incident frequency, or urgency specifically for stale regime-controller memory.","source_ids":["S3","S5","S6"]},"prior_art":{"proximity":"SUBSTANTIAL_COLLISION","closest_analogues":[{"name":"NIST AI RMF monitoring, bypass, preservation, and redeployment guidance","similarity":"It already combines drift or decontextualization detection, decision thresholds, temporary or permanent deactivation, dependency-aware change management, preserved components, contingency options, and redeployment criteria.","remaining_difference":"It is voluntary cross-domain guidance rather than an implemented registry for regime-specific plant models, control laws, thresholds, and causal maps; it does not test rare-regime recovery against latest-certified-only control.","source_ids":["S1","S2"]},{"name":"MLflow and Azure Machine Learning model registries","similarity":"Both provide versioned model records and lifecycle metadata; MLflow supports production/archive-style states, tags, aliases, and controlled environments, while Azure supports reversible archive and restore.","remaining_difference":"Neither cited product documentation derives eligibility from regime validity, assumptions, dependencies, or exceptions. Azure archival also does not prevent continued use, so an additional callable-eligibility gate would be necessary for the candidate's safety claim.","source_ids":["S7","S8"]},{"name":"Adaptive Model Set estimator for changing robot dynamics","similarity":"It chooses a small active subset from a larger candidate-model universe using observed explanatory adequacy and re-expands the set when active models no longer explain measurements, directly paralleling regime-sensitive eligibility and recovery.","remaining_difference":"It addresses online estimation efficiency and accuracy, not lifecycle metadata, operator visibility, audit holds, semantic restoration, quarantine, or authorization of control artifacts.","source_ids":["S4"]},{"name":"NASA software configuration management","similarity":"It requires identification and status accounting for models and versions, named change authority, audits, maintained records, and recoverable historical versions.","remaining_difference":"It governs configuration integrity but does not automatically score regime mismatch or compare reversible quarantine with latest-certified-only controller selection.","source_ids":["S5","S6"]}],"distinctive_claim_remaining":"The defensible remaining claim is empirical and system-specific, not a broad architecture novelty claim: for a heterogeneous control-memory registry, eligibility gates based jointly on current-regime fit, validation assumptions, dependencies, holds, and exceptions may reduce adjudicated stale selections relative to latest-certified-only control while preserving valid rare-regime recovery within a predeclared noninferiority margin. No reviewed source demonstrated that full comparison.","confidence":"HIGH"},"implementation_evidence":{"support":"MODERATE","rationale":"Versioning, status metadata, aliases, archives, restoration, access control, audits, and adaptive active-model subsets are independently implemented or experimentally demonstrated. The unverified engineering work is joining these components to controller-specific dependencies and assurance records, ensuring that archive status actually blocks selection, and proving semantic—not merely file-level—restoration for old regimes.","source_ids":["S2","S4","S5","S7","S8"]},"scores":{"meaningful_impact":{"score":3,"rationale":"Context drift and changing dynamics are credible, and incorrect model eligibility could affect control performance. No retrieved evidence establishes incident prevalence or the magnitude of safety and stability improvements from lifecycle gating, so impact remains plausible rather than demonstrated.","source_ids":["S1","S2","S3","S4"]},"stakeholder_pull":{"score":3,"rationale":"Official requirements and practitioner workshops show concrete demand for monitoring, configuration control, and clearer methods. Evidence of budgets or expressed demand for this exact control-memory intervention was not found.","source_ids":["S3","S5","S6"]},"incremental_advantage":{"score":3,"rationale":"Compared with latest-certified-only control, the proposal preserves reusable old regimes and adds dependency and exception gates. Adaptive model-set research supports the value of condition-responsive subsets, but the incremental result for lifecycle-governed control artifacts has not been tested.","source_ids":["S2","S4","S8"]},"distinctiveness_plausibility":{"score":2,"rationale":"Most architectural elements already appear in NIST lifecycle guidance, model registries, adaptive model-set selection, and configuration management. The remaining distinction is the combined control-specific comparison and its measured outcome, not the general idea of lifecycle states or reversible archiving.","source_ids":["S2","S4","S5","S7","S8"]},"technical_implementability":{"score":4,"rationale":"Commercial and open documentation demonstrates registries, metadata, aliases, archives, restore operations, and access controls; research demonstrates observation-driven active model subsets. Controller integration and semantic dependency restoration remain substantial but bounded engineering tasks.","source_ids":["S4","S7","S8"]},"adoption_authority_feasibility":{"score":4,"rationale":"NASA guidance supplies a credible precedent for project-manager responsibility, named change authority, audits, and configuration records. A control-system owner can authorize a read-only study, although safety-relevant eligibility changes would require separate assurance approval.","source_ids":["S5","S6"]},"evidence_readiness":{"score":4,"rationale":"The candidate can be evaluated offline with registry records and historical replay, using fixed and adaptive model-set precedents to define comparators and metrics. The main limitation is whether the target retains adequate regime labels, selection logs, dependencies, and rare-regime episodes.","source_ids":["S3","S4","S5"]},"safety_net_benefit":{"score":4,"rationale":"Preservation, contingency options, redeployment criteria, version recovery, archives, and restoration are all established mechanisms. Read-only shadowing avoids live-control risk, though semantic restoration and correct exception handling must still be demonstrated.","source_ids":["S2","S6","S8"]},"scalability":{"score":3,"rationale":"Registry primitives scale across model versions and environments, but MLflow's deprecation of inflexible fixed stages warns that lifecycle semantics must be tailored. Regime definitions, dependency graphs, certification evidence, and restoration tests will vary materially among control systems.","source_ids":["S5","S7","S8"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"50K_TO_250K","scope":"A 30-day, read-only inventory and offline replay for one noncritical subsystem, including data extraction, artifact and dependency classification, blinded validity adjudication, three-arm comparison, restoration sampling, safety review, and analysis. Includes control engineering, software/data work, operator time, assurance coordination, existing software, and evaluation; excludes live changes and new plant hardware.","confidence":"LOW","assumptions":["Usable registry snapshots, selection logs, historical telemetry, and a replay or simulator already exist.","A small multidisciplinary team can complete the work in roughly one to two calendar months.","The broad band allows for loaded labor above published base wages and for assurance and coordination overhead.","No regulated recertification or new physical testbed is required."],"source_ids":["S3","S4","S5"]},"initial_deployment_startup":{"band_2026_usd":"250K_TO_1M","scope":"Design and integrate lifecycle metadata, callable-eligibility enforcement, mismatch indicators, dependency and hold checks, exception workflow, audit logging, operator interface, restoration tests, cybersecurity review, and assurance documentation for one production subsystem. Includes labor, software, data migration, coordination, compliance, and evaluation; excludes major controller replacement.","confidence":"LOW","assumptions":["The controller and registry expose version and selection interfaces.","Four to eight specialist person-years or equivalent mixed internal and contractor effort fit within this resource band depending on loading and duration.","Existing compute, telemetry, and replay infrastructure can be extended.","Safety approval is subsystem-scoped rather than full-plant recertification."],"source_ids":["S2","S5","S7","S8"]},"operational_launch":{"band_2026_usd":"1M_TO_5M","scope":"Staged multi-subsystem launch with historical-artifact migration, dependency mapping, validation and regression campaigns, cybersecurity and safety assurance, operator training, incident and rollback procedures, monitoring, software integration, and program coordination. Includes limited equipment or test-environment upgrades but not wholesale plant redesign.","confidence":"LOW","assumptions":["Several heterogeneous subsystems and artifact classes are included.","Independent safety and cybersecurity reviewers participate.","Launch uses shadow and staged authorization before any live eligibility enforcement.","Existing control platforms remain serviceable and integration-accessible."],"source_ids":["S2","S5","S7","S8"]},"annual_recurring":{"band_2026_usd":"250K_TO_1M","scope":"Ongoing monitoring, lifecycle and exception reviews, dependency maintenance, archive-integrity and semantic-restoration tests, replay-suite upkeep, audit evidence, software and storage, incident response exercises, operator training, and periodic safety reassessment.","confidence":"LOW","assumptions":["New regime-specific artifacts continue to be produced.","Human review remains necessary for safety holds and ambiguous regime matches.","The estimate covers a bounded organizational deployment rather than global fleet operations.","Major controller replacement and extraordinary incident response are excluded."],"source_ids":["S2","S3","S5","S7","S8"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"External sources support the underlying problems of deployment-context drift, decontextualization, changing dynamics, and the burden or instability of indiscriminately active candidate sets. The narrower prevalence claim—that superseded control artifacts commonly remain callable and cause material incidents—remains unmeasured and is retained as blocking evidence.","source_ids":["S1","S2","S3","S4"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"NASA requirements demonstrate a credible governance pattern in which a project manager defines configuration-control responsibilities and named change authorities. For the proposed shadow study, the analogous accountable control-system owner is a credible authorizer.","source_ids":["S5","S6"]},"distinct_testable_incremental_claim":{"status":"YES","reason":"Despite substantial prior-art overlap, a distinct comparison remains: dependency- and regime-aware eligibility gating versus latest-certified-only control, with both stale-selection reduction and rare-regime recovery measured. None of the reviewed sources reports that combined test.","source_ids":["S2","S4","S7","S8"]},"bounded_next_evidence_step":{"status":"YES","reason":"A read-only inventory and offline historical replay can compare fixed, latest-only, and lifecycle-gated eligibility without changing actuators. Research precedent supplies measurable mode prediction and tracking outcomes, while official guidance supports testing in deployment-relevant conditions.","source_ids":["S1","S3","S4"]},"no_unresolved_safety_or_authority_stop":{"status":"YES","reason":"For the evidence step only, live gains, actuators, model eligibility, and registry contents remain untouched; the accountable owner authorizes access, and safety or incident holds are labels rather than targets for change. This does not authorize later production deployment.","source_ids":["S2","S5"]},"credible_cost_scope_and_range":{"status":"YES","reason":"The four scopes separately cover evidence, one-subsystem integration, multi-subsystem launch, and recurring operations, and explicitly include engineering, software, data, assurance, coordination, compliance, and evaluation. Bands remain low-confidence because target architecture and certification burden are unknown.","source_ids":["S3","S5","S7","S8"]}},"next_evidence_step":"Run a 30-day read-only inventory and offline historical replay on one noncritical subsystem. Pre-register three arms: current ungated eligibility, latest-certified-only, and the proposed regime/dependency/exception-gated lifecycle policy. Have domain experts blinded to arm adjudicate whether each selected artifact was valid for the logged regime. Compare stale-selection rate, tracking error or residual, overshoot, constraint-violation proxies, selector churn, and successful semantic restoration of held-out rare-regime artifacts. Use an untouched registry snapshot and pre-register a five-percentage-point noninferiority margin for rare-regime recovery. Falsify advancement if gating does not lower adjudicated stale-selection rate versus latest-certified-only, fails to improve any prespecified instability proxy, reduces rare-regime recovery by more than the margin, or hides any currently used or safety-held artifact.","blocking_evidence":["Target-system prevalence: number, age, visibility, and actual callability of superseded models, control laws, thresholds, and causal maps.","Incident or replay evidence connecting regime-mismatched selections to material instability, delay, or safety consequences.","Reliable labels or observable signals for regime fit, validation assumptions, dependencies, and exceptions.","Adequate historical coverage of rare recurring regimes; absence of such episodes can make recovery estimates uninformative.","Semantic restorability of archived artifacts together with their software, data, parameters, assumptions, interfaces, and assurance context.","False-positive and false-negative rates of stale-artifact classification, especially for safety-held or low-frequency models.","Target-specific certification, retention, incident-hold, cybersecurity, and mandatory-destruction constraints.","System-specific labor, data-cleaning, simulator, integration, and recertification requirements needed to narrow cost bands."],"research_disposition":"PARTNERED_RESEARCH_PROGRAM","world_novelty_boundary":"This bounded search found substantial collision with NIST lifecycle guidance, established model-registry archive and restore functions, adaptive active-model-set research, and software configuration management. It did not find a published evaluation of the exact heterogeneous control-memory policy against latest-certified-only selection with dependency gates and a rare-regime recovery noninferiority test. That remaining difference is a testable empirical claim, not evidence of world novelty; patent literature, proprietary control platforms, IEC/IEEE standards behind paywalls, and non-English literature were not exhaustively searched."}