{"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_production_rule_expiration_harness_04","proposal_index":4,"version":0,"title":"Production-Rule Expiration Harness for Cognitive Architectures","problem":"A computational cognitive model is extended across successive tasks by adding production rules, memory chunks, parameter overrides, and task-specific exceptions. Elements introduced for an earlier task or theoretical assumption can remain executable after the model, task environment, or explanation has changed. Because those deposits lack explicit lifecycle states, superseded elements may still fire, compete during rule selection, or influence predictions without being recognized as part of the current account. Removing them casually could break a dependency or make a published simulation impossible to reconstruct.","actors":["Cognitive modeler adding executable model elements","Theory lead responsible for the model's explanatory commitments","Cognitive-architecture maintainer","Researcher running simulations or fitting the model to a new task","Replication researcher reconstructing a reported simulation","Repository curator preserving executable research packages"],"observable_state":"For one cognitive-model family, an executable inventory lists production rules, chunks, overrides, and exception modules together with their creation revision, task bindings, last validation, firing traces, dependencies, and lifecycle states. Observable warning states include elements with no current hypothesis or task owner, rules that fire only under retired task configurations, overlapping rules that compete for the same conditions, overrides superseded by later assumptions, and published simulations whose exact executable package cannot be restored. Sandbox disabling and deterministic replay can show whether a candidate element changes current traces, predictions, benchmark results, or historical reconstruction.","consequence":"An accumulated active rule stack can make model behavior harder to attribute, allow obsolete assumptions to affect new simulations, and increase the effort required to determine which mechanisms generate a prediction. Ungoverned deletion can instead sever model dependencies or erase the executable history needed to reproduce an earlier scientific claim.","affected_objective":"Maintain a bounded, interpretable set of executable commitments in the current cognitive model while preserving complete, restorable packages for prior simulations and theoretical states.","intervention":"Install a lifecycle harness around executable cognitive-model elements. Every new production rule, chunk class, parameter override, or exception module receives an identity, author, creation revision, hypothesis and task bindings, dependency declarations, validation suite, review lease, and intended successor or expiry action. Instrumented traces identify where each element fires or participates in conflict resolution. Lease expiry, task retirement, supersession, validation drift, or prolonged lack of attributable use moves an element into review, not deletion. An age-weighted score ranks candidates using validation age, provenance, unique benchmark coverage, firing context, reconstruction value, and maintenance burden. Reviewers can renew, revise, shadow-disable, archive, or quarantine an element. Dependency checks and sandbox replays gate retirement. Historical executable packages move to archival tiers, remain subject to preservation exceptions, and undergo restore tests. Tombstones retain retired identities and successor mappings. The first deployment is read-only instrumentation and shadow disabling; it does not modify the authoritative model.","structural_mapping":[{"archetype_element":"Sequential deposits","domain_realization":"Production rules, declarative-memory structures, parameter overrides, and exception modules added during successive task adaptations and theory revisions."},{"archetype_element":"Changed active usefulness or validity","domain_realization":"An executable element can lose standing when its task is retired, its theoretical assumption is superseded, its parameterization changes, or newer mechanisms cover its former role."},{"archetype_element":"Old layers remaining active by default","domain_realization":"Earlier rules remain eligible to match and fire because executable inclusion is not separated from historical preservation."},{"archetype_element":"Stale layers mistaken for current authority","domain_realization":"A simulation trace includes an old rule whose presence is interpreted as part of the current theoretical account despite lacking a current binding."},{"archetype_element":"Age and decay rule","domain_realization":"Time since validation, task use, or explicit theoretical review reduces an element's review standing without independently determining deletion."},{"archetype_element":"Dependency and reconstruction check","domain_realization":"References among rules, buffers, chunks, modules, experiment configurations, reported runs, and random-seed packages are traced before retirement."},{"archetype_element":"Differentiated disposition paths","domain_realization":"Elements can remain active, be renewed, be revised, be shadow-disabled, be archived in an executable package, enter quarantine, or be destroyed after review."},{"archetype_element":"Preservation exceptions","domain_realization":"Published-result dependencies, canonical benchmark models, unresolved theoretical comparisons, and designated historical packages suspend ordinary disposition."},{"archetype_element":"Deletion audit and rollback","domain_realization":"Quarantine preserves a reversible copy, while a tombstone records retirement rationale, affected versions, and any replacement element."},{"archetype_element":"Revalidation loop","domain_realization":"Periodic trace audits, benchmark replays, exception reviews, and archive restorations reassess active and preserved model layers."}],"mechanism_mapping":[{"mechanism_slug":"stale_layer_detection_dashboard","role":"Provides an identity-resolved inventory and flags executable elements with obsolete task bindings, missing owners, superseding rules, failed validations, or unexplained participation in current traces.","counterfactual_removal":"Without the inventory and semantic flags, modelers must infer the active stack from source files and may overlook elements that remain executable under outdated assumptions."},{"mechanism_slug":"time_to_live_ttl_policy","role":"Assigns each new element a review lease based on its class and intended task scope; expiry triggers revalidation, shadow disabling, or archival review rather than destruction.","counterfactual_removal":"Without a precommitted review point, provisional patches can become permanent executable commitments through inattention."},{"mechanism_slug":"age_weighted_value_score","role":"Ranks review candidates using age since validation together with provenance, unique task coverage, firing context, reconstruction value, and maintenance burden.","counterfactual_removal":"Without ranking, the harness supplies an inventory but no bounded way to prioritize accumulated review work; the score itself has no retirement authority."},{"mechanism_slug":"dependency_safe_delete_check","role":"Traces inbound references from other rules, memory structures, task configurations, benchmark suites, reported simulations, and reconstruction packages before allowing retirement or destruction.","counterfactual_removal":"Without this gate, an apparently obsolete element could be removed while remaining load-bearing for an indirect execution path or historical result."},{"mechanism_slug":"lifecycle_storage_tiering_policy","role":"Keeps current executable elements in the active model while moving superseded but reconstructively useful elements into version-bound archival packages.","counterfactual_removal":"Without tiering, inactive theory history must either remain executable in the current model or be discarded."},{"mechanism_slug":"retention_schedule","role":"Defines preservation classes for exploratory patches, current theory components, benchmark models, published-result dependencies, and historically designated packages.","counterfactual_removal":"Without class-level policy and exceptions, individual modelers decide preservation inconsistently and may treat lack of current use as authority to destroy scientific history."},{"mechanism_slug":"soft_delete_quarantine_window","role":"Holds an approved retirement candidate in a recoverable, nonauthoritative state while regression and reconstruction checks continue.","counterfactual_removal":"Without quarantine, an overlooked dynamic dependency or rare execution path can turn a cleanup decision into an irreversible model break."},{"mechanism_slug":"tombstone_or_deletion_marker","role":"Preserves the retired element's identity, rationale, affected versions, and successor mapping so traces and citations do not resolve to an unexplained absence.","counterfactual_removal":"Without a tombstone, later modelers cannot distinguish deliberate theoretical retirement from accidental omission or an element that never existed."},{"mechanism_slug":"archive_restore_test","role":"Restores sampled historical model packages and reruns their preserved configurations, inputs, and seeds to test readability and executable reconnection.","counterfactual_removal":"Without an end-to-end restore test, an archived model may retain source files while losing the environment or dependency versions required to execute them."}],"causal_chain":["Successive task fits and theoretical revisions deposit new executable rules, chunks, overrides, and exceptions.","Earlier elements remain eligible for execution because historical preservation and current theoretical authority share the same active model stack.","Superseded elements can participate in matching, conflict resolution, memory retrieval, or prediction generation without an explicit current justification.","Unattributed participation makes the causal source of model behavior and the boundary of the current theory harder to determine.","The lifecycle harness binds each executable element to an owner, task, hypothesis, validation history, review lease, and dependency graph.","Trace-based staleness signals and age-weighted ranking identify candidates for bounded human review.","Sandbox disabling and dependency-gated replay test whether a candidate can leave the current executable stack safely.","Approved elements move through reversible quarantine into version-bound archives, while tombstones preserve successor and retirement information.","Restore drills preserve executable scientific history while the current model contains only explicitly renewed commitments."],"baseline":"Modelers use source-control history, comments, task-specific filenames, regression tests, and occasional manual deletion of apparently unused code. Version control preserves snapshots but does not consistently identify the lifecycle state of each executable element, distinguish historical presence from current theoretical authority, trace scientific reconstruction dependencies, or test whether archived environments still run.","nearest_rivals":["Source-control tags and immutable releases for each reported model version","Static dead-code analysis that identifies unreachable functions or rules","Regression testing that detects behavioral changes after model edits","Separate model branches or modules for each cognitive task","Model selection, regularization, or complexity penalties applied to the model as a whole"],"remaining_contrastive_claim":"The proposal's testable differentiator is lifecycle governance at the level of executable cognitive commitments. Source history preserves versions, static analysis finds unreachable code, and model selection compares whole models, but none alone determines whether an old yet reachable rule should retain current theoretical authority, requires dependency-gated shadow retirement, or preserves the retired rule in a tested executable reconstruction package.","authority_safety":{"decision_authority":"The theory lead and designated model owner jointly approve changes to the authoritative executable stack. Task owners validate task-specific dependencies, and the repository curator controls preservation of published or historically designated packages. Automated traces and scores may nominate candidates but cannot retire them.","authorized_first_step":"Instrument one model family in a sandbox, build a read-only element inventory, assign provisional lifecycle states, and shadow-disable selected review candidates one at a time while preserving the authoritative source and executable packages.","excluded_actions":["Automatic deletion or disabling in the authoritative model","Declaring a theoretical mechanism false because it is old or infrequently activated","Changing a published reconstruction package in place","Removing an element without dependency tracing and benchmark replay","Using firing frequency as the sole measure of scientific value","Treating a passing regression suite as proof that every relevant cognitive phenomenon is preserved","Rewriting reported results or interpretations"],"halt_rollback":"Halt if instrumentation changes model behavior, an element identity cannot be resolved, a known task or publication dependency is absent, a shadow-disabled candidate affects an unreviewed benchmark, or an archived package cannot be restored. Rollback consists of removing the instrumentation and provisional lifecycle metadata from the sandbox and returning to the unchanged authoritative model."},"negative_tests":{"strongest_counterevidence":"Strict task-specific modules or immutable model branches may already isolate historical rules from current execution while preserving reported versions. If current behavior is fully attributable and problematic interactions arise from the architecture's intended dynamics rather than superseded deposits, element-level lifecycle governance adds review machinery without addressing the causal problem.","problem_falsifier":"Every executable element in the selected model has a verified current task or hypothesis binding, no superseded element participates in current traces, all interactions are attributable through existing tools, and every reported model version can already be restored and executed.","intervention_falsifier":"The harness cannot reliably distinguish stale elements from dormant elements needed for rare conditions, shadow retirement repeatedly changes reviewed benchmark behavior or defeats reconstruction, or lifecycle review requires more interpretive effort than inspecting isolated versioned modules without clarifying causal attribution.","risks":["Rare but necessary rules can appear unused under an incomplete trace suite.","Age-weighted ranking can privilege recently studied tasks over enduring theoretical commitments.","Lifecycle labels can create false confidence that active rules are scientifically warranted.","Benchmark-gated retirement can overfit model maintenance to the available test suite.","Dynamic or indirect dependencies may escape the reference graph.","Frequent review leases can impose disproportionate curation work.","Preservation exceptions can accumulate until archives become difficult to navigate.","Archived environments can become insecure or unusable even when their files remain intact.","Tombstones and metadata can themselves become an unmanaged layer stack.","Modelers may use retirement status rhetorically as evidence against a competing theory." ]},"next_evidence_step":"Choose one existing cognitive-model family with two to four task configurations and at least one preserved reported version. In a sandbox, inventory executable elements, instrument matching and firing traces across the existing task suite, and have model owners verify task and hypothesis bindings. Select a bounded set of candidates exhibiting supersession, missing ownership, or expired validation; shadow-disable each candidate separately and replay current benchmarks plus one historical configuration. Record changed traces, prediction differences, failed dependencies, reviewer disagreements, time required for classification, and whether the archived historical package restores and executes. Do not alter the authoritative model or delete any element.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"Proposal 1 manages repository artifacts used to assemble experimental protocols and analyses; proposal 4 instead manages executable rules and memory structures inside a cognitive model, with the causal concern being hidden participation in simulated cognition rather than selection of the wrong document or script version. Proposal 2 manages accumulated behavioral evidence used to estimate a person's changing state; proposal 4 leaves observations and their inferential weights untouched and instead governs the model mechanisms that transform inputs into predictions. Proposal 3 manages human-facing handoff cues that compete for an operator's attention; proposal 4 has no shared handoff workspace and addresses causal attribution within executable cognitive architectures. Its lifecycle harness, trace-gated retirement, and restoration of historical simulations form an independently adoptable model-development practice even when protocol repositories, user-evidence stores, and operational cue boards require no intervention.","revision_record":{"parent_version":null,"progress_targets_addressed":["Initial complete specification for proposal index 4","Explicit material differentiation from sealed proposals 1, 2, and 3","Causal mapping of layer accumulation to executable cognitive-model elements","Operational authority, safeguards, rivals, falsifiers, and bounded sandbox evidence"],"conceptual_changes":["Initial version; no parent proposal","Targets accumulation of executable theoretical commitments within cognitive architectures"],"operational_changes":["Initial version; defines element review leases, trace-based staleness detection, dependency-gated shadow disabling, archival packages, quarantine, and tombstones"],"evidence_changes":["Initial version; specifies sandbox trace instrumentation, one-at-a-time shadow retirement, benchmark replay, and historical restoration"],"claim_changes":["Initial version; limits the contrastive claim to lifecycle governance of reachable executable model elements","Makes no claim of novelty, prevalence, demand, or effect size"]}}