{"schema_version":1,"assessment_id":"eoa_inverse_innovation_exp03_opportunity320_20260801","source_experiment_id":"eoa_inverse_innovation_exp03_full320_20260801","cell_id":"layer_decay_and_expiration_management__cognitive_science","archetype_slug":"layer_decay_and_expiration_management","domain_slug":"cognitive_science","title":"Governed lifecycle states for stale episodic traces in long-running cognitive agents","opportunity_summary":"Test whether reversible demotion, archival, quarantine, refresh, and supersession states reduce obsolete-context interference beyond ordinary retention and context-sensitive retrieval gating, while preserving rare-event recall, provenance, explanation completeness, and restoration fidelity. The packet establishes a coherent and testable hypothesis but provides no external evidence of problem prevalence, stakeholder demand, incremental performance, or distinctiveness.","adopter_authorizer":"A system owner and designated research data steward would authorize research and retention classes; irreversible deletion additionally requires their approval, while represented-person rights and applicable consent, investigation, safety, and governance holds override automation.","scores":{"meaningful_impact":{"score":4,"rationale":"If the specified interference occurs, reducing outdated associations and source/context errors while protecting rare-event knowledge and provenance would materially improve memory-guided decisions. The frequency and magnitude of this harm are unsupported, preventing a top score."},"stakeholder_pull":{"score":2,"rationale":"Users, operators, researchers, represented people, and auditors are identified, but the packet contains no expressed demand, adoption inquiry, incident history, or evidence that any stakeholder prioritizes lifecycle governance."},"incremental_advantage":{"score":3,"rationale":"Explicit retention, archival, dependency, restoration, and deletion states offer a testable addition to query-time retrieval gating. Whether those states improve accuracy or retrieval cost enough to justify their complexity is the central unresolved intervention falsifier."},"distinctiveness_plausibility":{"score":2,"rationale":"The composition is specifically differentiated from context-sensitive retrieval gating, but prior art is explicitly unsearched and no evidence distinguishes it from episodic-memory consolidation, archival memory, retrieval tiering, interference control, or machine-unlearning approaches."},"technical_implementability":{"score":3,"rationale":"Shadow labels and reversible retrieval-priority changes appear implementable in an inspectable artificial episodic store. Reliable staleness detection, distributed dependency assessment, provenance-preserving restoration, and durable archives remain substantial technical uncertainties."},"adoption_authority_feasibility":{"score":4,"rationale":"The packet identifies a research data steward and system owner, reserves irreversible deletion for approval, and gives governance holds priority. Feasibility is reduced by represented-person rights, erasure conflicts, and the need to coordinate research, operational, and governance authority."},"evidence_readiness":{"score":3,"rationale":"The proposal supplies baselines, a nearest rival, separate problem and intervention falsifiers, reversible assignment, outcomes, and halt conditions. Eligibility rules, validated detectors, evaluation windows, thresholds, subgroup checks, and restoration criteria are not operationally prespecified."},"safety_net_benefit":{"score":4,"rationale":"Reduced-priority retrieval, retention holds, audit records, restoration testing, and immediate rollback directly limit harm from false staleness classifications during sandbox research. They cannot fully protect against missed distributed influence or lifecycle effects not captured by benchmark outcomes."},"scalability":{"score":2,"rationale":"Per-trace metadata, repeated revalidation, dependency scans, protected exceptions, archive maintenance, and restore testing may grow with the episodic store. The packet provides no evidence that these functions remain accurate or economical at long-running-agent scale."}},"score_confidence":"LOW","costs":{"first_evidence":{"band_2026_usd":"50K_TO_250K","scope":"Design and execute one sandbox study using matched long-running agents, shadow staleness labeling, randomized ordinary-versus-reduced-priority retrieval, comparison with context-sensitive gating, preregistered outcomes, subgroup checks, restore tests, and analysis.","confidence":"LOW","assumptions":["An inspectable episodic-agent benchmark and basic retrieval instrumentation already exist.","The study uses reversible demotion only and performs no hard deletion.","The evaluation can be completed with one research and engineering team without acquiring a large proprietary dataset.","Resource needs may rise if long-horizon memory accumulation must be generated from scratch."]},"initial_deployment_startup":{"band_2026_usd":"250K_TO_1M","scope":"Engineer lifecycle metadata and state transitions for one agent platform, retention holds, archival and restoration paths, audit logging, operator review controls, dependency probes, and preproduction validation.","confidence":"LOW","assumptions":["Deployment is limited to one inspectable artificial episodic-memory architecture.","Existing identity, authorization, storage, and observability infrastructure can be extended.","No automated irreversible deletion is included.","Legal and governance review is material but does not require multi-jurisdiction redesign."]},"operational_launch":{"band_2026_usd":"1M_TO_5M","scope":"Launch lifecycle governance for a production agent population with migration of existing traces, archive infrastructure, steward workflows, access controls, incident response, restore exercises, fairness evaluation, documentation, and staged rollout monitoring.","confidence":"LOW","assumptions":["The launch covers multiple agent instances but not an enterprise-wide heterogeneous fleet.","Consequential production decisions remain gated until sandbox and preproduction evidence passes.","Human review remains necessary for protected traces and high-impact dispositions.","Archive compatibility and trace provenance can be preserved without replacing the underlying memory system."]},"annual_recurring":{"band_2026_usd":"250K_TO_1M","scope":"Operate storage tiers, periodic revalidation, steward review, governance-hold administration, audit retention, monitoring, detector recalibration, subgroup analysis, restore drills, and incident handling for one deployed platform.","confidence":"LOW","assumptions":["Trace volume and review queues remain bounded enough for a small dedicated operational function.","Hard deletion remains exceptional rather than a high-volume automated process.","Archive migrations and major regulatory changes are excluded.","Recurring costs could exceed the band if dependency analysis or manual review scales approximately with total trace volume."]}},"research_burden":"HIGH","earliest_credible_horizon":"3_TO_12_MONTHS","pipeline_gates":{"recognizable_externally_supportable_problem":{"status":"UNCERTAIN","reason":"The packet specifies observable interference, consequences, and a problem falsifier, but supplies no external observations showing that obsolete-context exposure actually predicts latency, context errors, or decision degradation."},"identifiable_adopter_or_authorizer":{"status":"YES","reason":"A system owner and designated research data steward are explicitly assigned approval authority, with represented-person rights and governance holds overriding automation."},"distinct_testable_incremental_claim":{"status":"YES","reason":"Governed reversible demotion is explicitly claimed to improve current-context accuracy or retrieval cost beyond both retain-all retrieval and context-sensitive retrieval gating, subject to preservation outcomes."},"bounded_next_evidence_step":{"status":"YES","reason":"The authorized first step is a fixed-window sandbox experiment that shadow-labels traces, randomizes eligible candidates to ordinary or reversibly reduced-priority retrieval, prohibits hard deletion, and includes predefined halt-and-restore conditions."},"no_unresolved_safety_or_authority_stop":{"status":"YES","reason":"For the bounded sandbox step, irreversible deletion and consequential production use are excluded, protected holds override assignment, and suppressed traces must be restored upon specified breaches. These controls do not authorize later deletion or production deployment."},"implementation_cost_scope_and_range":{"status":"UNCERTAIN","reason":"The required components are identifiable, but the packet does not specify trace volume, platform architecture, benchmark availability, archive requirements, review load, jurisdictions, or integration environment sufficiently to validate deployment cost ranges."}},"blocking_evidence":["Whether obsolete-context trace exposure predicts retrieval latency, source/context errors, or decision degradation after controlling for total memory volume and retrieval policy.","Whether governed reversible demotion outperforms both ordinary retain-all retrieval and context-sensitive retrieval gating on prespecified accuracy or retrieval-cost outcomes.","Whether rare-event recall, explanation completeness, subgroup outcomes, provenance, and restoration fidelity remain within acceptable bounds.","Whether stale-candidate detection and dependency probes are reliable enough to support reversible isolation without suppressing delayed-value, minority, or distributed experience.","Whether the lifecycle composition is meaningfully distinct from relevant existing approaches; prior art is unsearched."],"next_evidence_step":"Pre-register and run a fixed-window sandbox experiment on matched long-running agents: first test whether obsolete-context exposure predicts degradation after controlling for memory volume and retrieval policy, then randomly assign eligible stale candidates to ordinary retrieval or reversible reduced-priority retrieval and compare both with context-sensitive retrieval gating. Falsify progression if demotion yields no prespecified accuracy or retrieval-cost improvement, or breaches rare-event recall, explanation completeness, subgroup-impact, provenance, or restore-fidelity bounds; perform no hard deletion.","research_questions":["Does obsolete-context trace exposure independently predict retrieval latency, source/context errors, or decision degradation?","How accurately can stale candidates be identified without disproportionately labeling rare, minority, or delayed-value experiences?","Does governed demotion add benefit beyond context-sensitive retrieval gating, and under which memory loads or context shifts?","Can explicit and distributed dependencies be probed well enough to make reversible isolation behaviorally safe?","Do archived traces retain provenance, readability, retrieval timeliness, and behavioral usefulness after restoration?","What thresholds should trigger refresh, supersession, reduced priority, archive, quarantine, or continued retention?","How do lifecycle states affect autobiographical continuity and explanations beyond benchmark accuracy?","Which existing research or system approaches already implement materially similar lifecycle governance?","What trace volumes, review rates, and archive designs determine operational cost and scalability?","How should immediate-erasure obligations interact with quarantine, tombstones, audit metadata, and governance holds?"],"recommendation":"PARTNERED_RESEARCH","uncertainty_constraints":["Closed-book assessment provides no evidence of prevalence, realized impact, stakeholder demand, market size, prior art, or adoption history.","The proposal is a hypothesis for inspectable artificial episodic-memory systems and does not support intervention in biological human memory.","Cost bands are resource-equivalent planning ranges based on assumed scope, not quotations or observed implementation costs.","Dynamic relevance means an apparently obsolete trace may become useful in a novel context.","Absence of visible references does not establish behavioral independence in distributed representations.","The packet names outcomes and safeguards but does not provide calibrated eligibility rules, thresholds, detector performance, or evaluation duration.","Sandbox reversibility does not establish the safety or authority of irreversible deletion or consequential production deployment."],"closed_book_prior_art_boundary":"Prior-art status is explicitly UNSEARCHED. This assessment makes no claim that governed episodic-trace lifecycle states are novel, rare, prevalent, or absent from episodic-memory consolidation, interference control, archival memory, retrieval tiering, machine unlearning, or adjacent approaches."}