{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp05_complete_proposal_portfolio20_20260803","cell_id":"layer_decay_and_expiration_management__cognitive_science","arm":"COMPLETE_PROPOSAL_PORTFOLIO","candidate_id":"cand_cognitive_evidence_lease_manager_02","proposal_index":2,"version":0,"title":"Cognitive Evidence Lease Manager for Longitudinal Adaptive Training","problem":"A longitudinal adaptive cognitive-training system accumulates trial-derived evidence about a user's working-memory performance across sessions. Evidence produced under older ability levels, task parameters, scoring models, or contexts can remain active in the current user model by default. Those deposits can continue influencing difficulty selection after their relevance has changed, while simply deleting them would erase the history needed to interpret trajectories, reproduce prior adaptations, or investigate model behavior.","actors":["Person completing the cognitive-training sessions","Cognitive scientist defining the measured construct and task","Model developer maintaining the user-state estimator","Program operator reviewing adaptive recommendations","Data steward governing retention and participant permissions","Researcher auditing or reconstructing an earlier adaptation"],"observable_state":"For one user's task history, the live estimator consumes timestamped evidence layers from multiple sessions and task or scoring versions, but some layers lack an inference-active, review-due, archived, quarantined, or preserved state. Counterfactual replay can identify whether older layers materially alter the current state estimate or next-block difficulty recommendation after newer observations or a documented context change. The system can also be tested for whether an earlier recommendation remains reconstructable from retained evidence and model versions.","consequence":"A current recommendation can reflect an unresolved mixture of present performance and superseded evidence, producing task difficulty or feedback that is mismatched to the state the system claims to estimate. Aggressive history deletion can instead prevent longitudinal interpretation, audit, rollback, or reconstruction of earlier adaptations.","affected_objective":"Keep online estimates and adaptations responsive to currently relevant cognitive evidence while preserving the longitudinal record required for accountable interpretation and reconstruction.","intervention":"Give every trial-derived evidence layer an explicit inference lease at creation, defined by evidence class, task version, scoring version, context signature, and review horizon. Lease expiry removes the layer from ordinary current-state inference but does not destroy it: the layer moves to an archival tier with its provenance and prior model bindings intact. New corroborating evidence can renew a lease, while detected task, scoring, or context changes can trigger early review. An age-weighted value score ranks review cases using recency, provenance, current predictive contribution, rare-condition coverage, reconstruction value, and sensitivity, but cannot authorize disposition. Deletion requires retention-policy clearance, a trace of dependencies from derived scores and prior recommendations, reversible quarantine, and a tombstone. Sampled archive-restore drills verify that earlier estimates can still be reconstructed. Initial use is offline shadow replay only.","structural_mapping":[{"archetype_element":"Sequentially accumulated layers","domain_realization":"Timestamped trial results, session summaries, inferred ability contributions, context annotations, and model-version bindings accumulated across training sessions."},{"archetype_element":"Changed active usefulness","domain_realization":"Evidence may cease to represent the user's current task performance after learning, interruption, context change, task revision, or scoring-model replacement."},{"archetype_element":"Stale layers remaining active by default","domain_realization":"Older observations continue contributing to the current latent-state estimate without an explicit decision that they remain inferentially applicable."},{"archetype_element":"Age and decay rule","domain_realization":"Each evidence class receives a review horizon and decay profile governing its standing in current-state inference rather than its physical existence."},{"archetype_element":"Expiration trigger","domain_realization":"Lease elapsed, incompatible task or scoring version, documented context transition, or contradiction by newer validated evidence triggers review or archival from active inference."},{"archetype_element":"Differentiated disposition paths","domain_realization":"An evidence layer can remain active, be renewed, be demoted to historical archive, enter quarantine, be preserved by exception, or be destroyed after safeguards."},{"archetype_element":"Dependency and reconstruction protection","domain_realization":"Derived scores, prior recommendations, reports, and model audits are traced before an underlying evidence layer can be removed."},{"archetype_element":"Preservation exceptions","domain_realization":"Named research, consent, records, safety-investigation, or rare-condition coverage needs can suspend ordinary expiry or destruction."},{"archetype_element":"Reversible cleanup and surviving deletion evidence","domain_realization":"Quarantine permits restoration, while a tombstone retains identity, provenance boundary, disposition reason, and successor information."},{"archetype_element":"Revalidation loop","domain_realization":"Offline replays periodically test lease assumptions, exception validity, archive readability, and reconstruction of sampled historical adaptations."}],"mechanism_mapping":[{"mechanism_slug":"time_to_live_ttl_policy","role":"Attaches a predeclared inference lease to each evidence layer. Expiry means exclusion from ordinary current-state estimation or mandatory review, not automatic destruction.","counterfactual_removal":"Without a creation-time lease, old evidence remains active until someone notices it, recreating indefinite accumulation in the live estimator."},{"mechanism_slug":"stale_layer_detection_dashboard","role":"Inventories evidence layers and surfaces incompatibilities involving task versions, scoring versions, context signatures, contradictory newer evidence, or missing ownership.","counterfactual_removal":"Without semantic staleness detection, the system can expire evidence by age but miss recent evidence rendered inapplicable by a task or context change."},{"mechanism_slug":"age_weighted_value_score","role":"Ranks expired or contested layers for renewal, archive review, or preservation using age together with provenance, predictive contribution, rare-condition coverage, sensitivity, and reconstruction value.","counterfactual_removal":"Without the score, stewards face an unranked review queue; safety remains intact because the score only orders attention and never disposes evidence."},{"mechanism_slug":"lifecycle_storage_tiering_policy","role":"Separates the small active inference set from historical evidence retained in warm or cold tiers for trajectory analysis and reconstruction.","counterfactual_removal":"Without tiering, expiration tends either to leave evidence active or to equate inferential retirement with physical deletion."},{"mechanism_slug":"retention_schedule","role":"Defines minimum and maximum retention conditions for evidence classes and registers participant-permission, research, records, and investigation overrides.","counterfactual_removal":"Without it, inferential usefulness could be mistaken for authority to retain or destroy participant-linked evidence."},{"mechanism_slug":"dependency_safe_delete_check","role":"Blocks destruction while a derived score, historical recommendation, analysis, report, or reconstruction path still depends on the evidence layer.","counterfactual_removal":"Without it, apparently stale evidence could be removed even though it remains load-bearing for interpreting past system behavior."},{"mechanism_slug":"soft_delete_quarantine_window","role":"Makes an approved destruction reversible for a bounded period, with the window based on reconstruction impact and sensitivity constraints.","counterfactual_removal":"Without quarantine, an erroneous dependency judgment or bulk disposition becomes immediately irreversible."},{"mechanism_slug":"tombstone_or_deletion_marker","role":"Leaves a durable marker recording that an evidence layer existed, why it was removed, when removal occurred, and which aggregate or successor now resolves its identity.","counterfactual_removal":"Without a tombstone, missing evidence becomes indistinguishable from evidence never collected, undermining interpretation of historical model states."},{"mechanism_slug":"archive_restore_test","role":"Samples archived evidence across task versions and age bands and reconstructs the corresponding historical state estimate and recommendation in a sandbox.","counterfactual_removal":"Without an end-to-end restore test, archival preserves nominal history without evidence that the history remains usable."}],"causal_chain":["The adaptive system deposits new trial-derived evidence during each session while retaining earlier deposits.","User state, task parameters, scoring logic, and relevant context can change, but older deposits retain inferential standing by default.","The current estimator consequently mixes evidence generated under potentially incompatible states or measurement conditions.","Inference leases and semantic staleness triggers make continued active standing an explicit, inspectable decision.","Expired or incompatible evidence leaves the active estimator but remains archived with provenance and historical model bindings.","Dependency checks and preservation overrides prevent removal of layers needed to explain derived scores or prior adaptations.","Quarantine and tombstones make approved cleanup reversible and historically legible.","Archive-restore drills test whether prior estimates and recommendations remain reconstructable.","The live evidence set stays bounded and context-specific without collapsing historical retention into continued inferential authority."],"baseline":"The platform retains trial history and either fits a cumulative user score, applies one global recency weight, or resets state manually when an operator notices a major change. Raw history may remain available, but individual evidence layers do not have explicit inferential states, context-specific leases, dependency gates, preservation exceptions, or tested reconstruction paths.","nearest_rivals":["A fixed sliding window that estimates current state from only the most recent trials","Exponential forgetting applied uniformly inside the user-state model","A state-space or change-point model that represents ability as temporally varying","A complete user-model reset at the beginning of each session","Immutable event logging combined with periodic full-model retraining"],"remaining_contrastive_claim":"The proposal's testable differentiator is the separation of inferential activation from historical retention at the evidence-layer level. Recency weighting or temporal modeling may address nonstationarity, and immutable logs may preserve events, but neither alone supplies explicit expiry states, context-sensitive renewal, dependency-gated disposition, preservation exceptions, reversible cleanup, and tested reconstruction of prior adaptations.","authority_safety":{"decision_authority":"The cognitive scientist and model owner approve inference-lease rules; the data steward controls retention and destruction under participant permissions and institutional requirements. Operators may request review but cannot override preservation holds. Scores and automated detectors have no deletion authority.","authorized_first_step":"Run offline shadow replays on one completed, appropriately governed task dataset; assign provisional leases and lifecycle states without changing the live estimator, participant experience, or source records.","excluded_actions":["Automatic hard deletion of trial or participant records","Live changes to task difficulty or feedback during the first evidence step","Use of lifecycle status as a clinical, diagnostic, employment, educational, or safety-critical judgment","Expiry based solely on chronological age","Overriding participant permissions or institutional retention requirements","Removing evidence required to reconstruct a prior recommendation or reported analysis","Treating the value score as a measure of a person's cognitive worth or stable ability"],"halt_rollback":"Halt if shadow processing exposes data outside its authorized scope, assigns inconsistent identities, misses a known dependency, cannot reproduce the baseline estimator, or fails to restore a sampled historical state. Rollback consists of discarding the derived shadow registry and provisional states while leaving the governed source data, live model, and task operation unchanged."},"negative_tests":{"strongest_counterevidence":"A well-specified temporal user model may already distinguish current state from historical observations while immutable source logs preserve complete reconstruction. If older observations continue improving current-state prediction after task version, scoring version, and recent evidence are accounted for, lifecycle expiry may discard useful signal without solving a separate problem.","problem_falsifier":"In the selected dataset, old evidence never changes a current estimate or recommendation after conditioning on recent evidence and model state, task or scoring changes do not create incompatible layers, and every historical recommendation is already reconstructable from explicit version bindings.","intervention_falsifier":"In offline replay, the lifecycle-managed estimator is no better than the cumulative baseline or a simple recency rival at matching held-out current-session behavior, while producing unstable recommendations, false staleness classifications, unmanageable review burden, or failures to reconstruct archived historical estimates.","risks":["Short or poorly chosen leases can create excessive recency bias and discard stable individual information from active inference.","Context-sensitive expiry rules can encode sensitive attributes or apply unevenly across users.","Feedback between task selection and observed performance can make recent evidence appear more authoritative than it is.","A composite value score can conceal contestable judgments behind false numerical precision.","Incomplete dependencies can allow destruction of evidence needed to interpret a prior recommendation.","Preservation exceptions can accumulate and recreate an unbounded historical store.","Archival latency can obstruct a time-sensitive audit or investigation.","Quarantine retains data that may be subject to a binding destruction requirement.","Frequent state transitions can make the model harder for researchers and users to understand." ]},"next_evidence_step":"Select one completed dataset from a single working-memory task with multiple sessions and at least two documented task or scoring contexts, excluding any use not already authorized by its governance conditions. In a sandbox, reproduce the existing cumulative estimator, implement provisional inference leases, and compare three offline replays: cumulative baseline, a fixed recency rival, and lifecycle-managed activation. Evaluate held-out next-block performance alignment, recommendation stability, the number and reasons for layers changing state, false staleness judgments reviewed by the task owner, dependency completeness, and successful reconstruction of sampled historical recommendations. Make no live adaptation and destroy no source evidence during this bounded test.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"Earlier proposal 1 governs the lifecycle of laboratory protocol artifacts so researchers do not select superseded instructions, stimuli, scoring rules, or scripts while reconstructive packages remain available. This proposal instead governs behavioral evidence inside a longitudinal adaptive user model. Its affected actors, accumulated objects, decision point, and causal path are different: old observations distort an online estimate and subsequent task adaptation, rather than old research artifacts causing assembly of an incompatible study package. Its intervention is an inference-lease system that removes evidence from current model activation while preserving historical records, not a registry for retiring protocol versions. It is independently adoptable by an adaptive-system team even if its protocol repository requires no lifecycle intervention.","revision_record":{"parent_version":null,"progress_targets_addressed":["Initial complete specification for proposal index 2","Material differentiation from sealed proposal 1","Causal preservation of accumulation, decay, expiry, archival, dependency protection, exceptions, reversibility, and restoration","Bounded offline evidence with explicit authority limits and falsifiers"],"conceptual_changes":["Initial version; no parent proposal","Targets lifecycle management of longitudinal cognitive evidence rather than lifecycle management of research protocol artifacts"],"operational_changes":["Initial version; defines evidence leases, context-sensitive review triggers, archival from active inference, and offline shadow replay"],"evidence_changes":["Initial version; specifies comparison with cumulative and fixed-recency estimators plus reconstruction testing"],"claim_changes":["Initial version; limits the contrastive claim to explicit evidence-layer lifecycle governance","Makes no claim of novelty, prevalence, demand, or effect size"]}}