{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp05_complete_proposal_portfolio20_20260803","cell_id":"invariant_mode_decomposition_design__literature_literary_theory","arm":"COMPLETE_PROPOSAL_PORTFOLIO","candidate_id":"cand_genre_prototype_stability_gate_004","proposal_index":4,"version":0,"title":"Genre-Prototype Stability Gate for Digital Literary Corpora","problem":"A digital literary corpus may recalibrate genre memberships by deriving category prototypes from currently assigned works and then rescoring those works against the updated prototypes. For hybrid works, this feedback can create a coupled reassignment direction: small changes to several memberships alter multiple prototypes, which then reinforce or reverse those changes in the next round. Aggregate genre counts can remain nearly constant while particular groups of works oscillate between categories or drift toward a self-confirming definition of a genre.","actors":["Corpus curator, who governs metadata releases and calibration rules","Genre scholars, who define permissible categories and interpret boundary cases","Cataloguers, who record textual evidence and candidate memberships","Corpus researchers, who depend on reproducible genre cohorts","Technical analyst, who estimates the local update operator without determining literary classifications"],"observable_state":"For a bounded corpus slice, the state vector contains one membership weight for every work–genre pair, together with explicit uncertainty status. A documented calibration function recomputes each genre prototype from eligible memberships and then updates work-level membership weights from catalogued textual evidence. Small sandbox perturbations around the current state produce a local Jacobian mapping one calibration round's membership deviations to the next. Historical or replayed rounds provide raw trajectories. Observable outcomes include work-level label switching, cohort overlap between releases, disagreement with independently adjudicated multi-label reference cases, and residual reconstruction error.","consequence":"Researchers can obtain materially different genre cohorts from adjacent metadata releases even when no new works or categories were added. Hybrid works may be forced into unstable single-label assignments, and the recalibration process may make its current genre boundaries appear empirically supported because those same boundaries determine which works update the prototypes.","affected_objective":"Maintain reproducible, traceable genre cohorts while representing hybrid or uncertain works without allowing them to create uncontrolled feedback in prototype recalibration.","intervention":"Before publishing a recalibrated corpus release, run the prototype update in a sandbox, estimate its local Jacobian, and decompose the coupled work–genre membership dynamics. Identify persistent, oscillatory, or growing modes that are stable under plausible evidence and weighting perturbations. A sensitivity sweep tests candidate controls, including soft-label retention, prototype-contribution caps, fixed reference anchors, and temporary exclusion of high-loading disputed memberships from prototype updates. The curator may activate a one-cycle stability gate: disputed memberships remain visible and searchable under all supported genres but do not update category prototypes until genre scholars adjudicate the underlying textual evidence. Stable reference cases anchor the recalibration, residual cases remain in a manual queue, and the operator is re-estimated before release.","structural_mapping":[{"archetype_element":"Transformation Scope","domain_realization":"The transformation is one execution of a fixed genre-prototype recalibration rule on a bounded corpus slice and fixed taxonomy."},{"archetype_element":"State-Vector Definition","domain_realization":"Coordinates are work–genre membership weights and uncertainty flags, traceable to catalogued textual evidence."},{"archetype_element":"Invariant Mode Basis","domain_realization":"Eigenvectors of the local update Jacobian represent coupled patterns of work and genre memberships that approximately retain their direction across recalibration rounds."},{"archetype_element":"Modal Gain Spectrum","domain_realization":"Eigenvalues estimate whether each coupled reassignment pattern is damped, preserved, reversed, oscillatory, or amplified by another round."},{"archetype_element":"Dominant Mode Selection Rule","domain_realization":"A mode becomes governance-relevant only if it is stable under resampling and plausible rule perturbations, predicts held-out membership movement, and affects declared cohort-reproducibility or reference-case criteria."},{"archetype_element":"Stable/Unstable Mode Partition","domain_realization":"Strongly damped modes proceed through ordinary recalibration; persistent, oscillatory, or growing modes are subject to the stability gate."},{"archetype_element":"Modal Intervention Map","domain_realization":"Sensitivity tests map prototype weights, anchors, contribution caps, and disputed memberships to changes in action-relevant modal gains and corpus outcomes."},{"archetype_element":"Reconstruction Residual Check","domain_realization":"Held-out recalibration trajectories are reconstructed from retained modes, while structured residual cases remain visible for manual review."},{"archetype_element":"Mode-Coupling Register","domain_realization":"Near-degenerate modes and shared boundary works are registered so coupled genre shifts are not presented as independent category changes."},{"archetype_element":"Local Linearization Window","domain_realization":"The Jacobian applies only near the current memberships, taxonomy, evidence representation, and calibration rule; major changes require a new analysis."},{"archetype_element":"Spectral Gap Threshold","domain_realization":"A reduced modal representation is used only when retained and omitted modes remain sufficiently separated under resampling and parameter perturbation."},{"archetype_element":"Mode Drift Monitor","domain_realization":"Mode directions, gains, work loadings, residuals, and spectral separation are recomputed after adjudication or any calibration-rule change."},{"archetype_element":"Interpretation Scope Contract","domain_realization":"A modal loading measures participation in recalibration feedback, not the work's essential genre, literary value, or canonical status."}],"mechanism_mapping":[{"mechanism_slug":"eigendecomposition_workflow","role":"Factor the explicit local Jacobian into a complete basis of coupled membership modes and their scalar gains.","counterfactual_removal":"Without the full decomposition, cataloguers can observe individual label changes but cannot distinguish the distributed directions preserved or amplified by the recalibration rule."},{"mechanism_slug":"modal_stability_analysis","role":"Classify modes as decaying, persistent, reversing, oscillatory, or growing within the current calibration window.","counterfactual_removal":"Without stability classification, ordinary boundary uncertainty cannot be separated from membership patterns that compound or alternate across rounds."},{"mechanism_slug":"power_iteration_probe","role":"Check the dominant mode and spectral separation in large work–genre state spaces without requiring repeated full factorization.","counterfactual_removal":"Without a scalable dominant-mode probe, the stability gate may be impractical to refresh on the bounded corpus slice as memberships change."},{"mechanism_slug":"modal_sensitivity_sweep","role":"Perturb anchors, prototype contributions, soft-label rules, and candidate gated memberships to rank their effects on modal gains and declared corpus outcomes.","counterfactual_removal":"Without sensitivity testing, detecting instability would not reveal which bounded governance control can alter it without suppressing the disputed metadata."},{"mechanism_slug":"reduced_order_model","role":"Provide a sandbox surrogate for replaying a small set of action-relevant membership modes under candidate calibration controls.","counterfactual_removal":"Without the surrogate, curators would need to rerun and inspect the complete high-dimensional recalibration for every candidate gate configuration."},{"mechanism_slug":"residual_reconstruction_test","role":"Test whether the retained modes reproduce held-out membership trajectories and identify boundary cases omitted by the reduction.","counterfactual_removal":"Without residual testing, a low-amplitude but theoretically consequential hybrid cluster could disappear from the governance view."},{"mechanism_slug":"spectral_gap_monitor","role":"Detect narrowing separation or rotation of retained modes before a fixed stability gate is reused.","counterfactual_removal":"Without gap and drift monitoring, the corpus could continue applying controls derived from obsolete genre boundaries."},{"mechanism_slug":"spectral_decomposition_report","role":"Document mode loadings, gains, control sensitivities, residual cases, evidence links, and prohibited essentialist interpretations.","counterfactual_removal":"Without a bounded report, numerical instability findings could be mistaken for authoritative genre assignments."}],"causal_chain":["Current work memberships determine which works contribute to each genre prototype.","Updated prototypes rescore work memberships, so classification outputs become inputs to the next calibration round.","Hybrid works can couple several prototypes, causing a small membership change to propagate across multiple genres.","Repeated recalibration can preserve, reverse, or amplify a distributed reassignment mode even when aggregate category counts appear stable.","Local decomposition exposes the coupled mode and estimates its scalar behavior across further rounds.","Held-out reconstruction and reference-case outcomes test whether the mode represents actionable feedback rather than harmless uncertainty.","Sensitivity analysis identifies which prototype contributions or rule parameters sustain the mode.","A temporary stability gate breaks the feedback path while retaining disputed multi-label metadata for discovery and scholarly adjudication.","Re-estimation after adjudication determines whether ordinary recalibration can resume or the local modal model must be retired."],"baseline":"The baseline recalibrates prototypes from all currently assigned works, applies per-work confidence thresholds, and sends only low-confidence individual assignments to manual review. Release comparison focuses on total genre counts and isolated label changes rather than coupled membership trajectories.","nearest_rivals":["Fixed controlled-vocabulary cataloguing, which avoids prototype feedback but cannot use evidence-based recalibration without manual revision of every affected work.","Per-work confidence thresholds, which identify uncertain assignments but can miss a coupled mode distributed across individually moderate-confidence memberships.","Manual genre-panel adjudication, which supplies contextual scholarly judgment but does not reveal how accepted assignments feed back into later prototype updates.","Versioned metadata snapshots, which preserve reproducibility after release but document drift rather than diagnosing or controlling its transformation.","Cluster analysis or principal-component visualization, which can display groups of works but does not classify their behavior under repeated recalibration or map governance controls into modal gains."],"remaining_contrastive_claim":"The proposal is warranted only if the locally estimated coupled modes predict held-out membership trajectories or cohort instability better than individual confidence, aggregate counts, and non-transformational clustering, and if a sensitivity-selected gate improves the declared stability criteria without concealing disputed memberships or degrading independently adjudicated reference cases.","authority_safety":{"decision_authority":"The corpus curator controls metadata releases and calibration settings; genre scholars adjudicate literary evidence and category meaning. The modal analysis may trigger review but cannot assign a work's genre or redefine the taxonomy.","authorized_first_step":"Replay the existing recalibration rule in a nonpublishing sandbox on an authorized corpus slice and produce a read-only stability report.","excluded_actions":["Automatically changing, deleting, or publishing any genre membership","Treating modal loadings as statements about a work's essential identity, quality, or canonical importance","Permanently excluding hybrid works from prototype formation without scholarly review","Hiding disputed memberships from users or forcing them into a single genre","Changing the taxonomy, evidence rubric, or public API without the corpus governance process","Applying the fitted operator to another corpus, taxonomy, language, or calibration regime without re-estimation"],"halt_rollback":"Halt if membership evidence cannot be traced, reference adjudication is unavailable, modes are unstable under resampling, spectral separation is inadequate, held-out residuals exceed tolerance, the gate disproportionately removes the evidence needed to represent hybrid genres, or curatorial authorization is withdrawn. Discard the sandbox release candidate, restore the unchanged published metadata snapshot, and return disputed works to ordinary manual review."},"negative_tests":{"strongest_counterevidence":"Observed membership changes follow an explicit scholarly redefinition of the taxonomy and remain stable after that change, while individual confidence thresholds and reference-case adjudication fully explain the transitions without a persistent recalibration mode.","problem_falsifier":"The prototype update does not materially depend on prior assignments, repeated sandbox rounds converge without stable oscillatory or growing directions, or apparent modes disappear under coder, evidence, and parameter perturbations.","intervention_falsifier":"On held-out replay rounds, the stability gate fails to reduce unsupported switching or cohort divergence relative to confidence-threshold and manual-review baselines, or it merely freezes existing prototype bias and worsens agreement with independently adjudicated multi-label cases.","risks":["Temporarily withholding boundary works from prototype updates may make genre prototypes artificially homogeneous.","The chosen evidence representation may encode one theoretical account of genre as neutral metadata.","Stable reference anchors may preserve inherited classification errors.","A locally stable calibration can still embody unsuitable genre categories.","Modal reduction may omit a small hybrid tradition with high scholarly significance.","Repeated gating may delay necessary taxonomy revision by treating conceptual conflict as numerical instability.","Corpus users may mistake multi-label uncertainty for poor cataloguing rather than a substantive genre boundary.","Large state spaces and small spectral gaps may make mode identity numerically fragile."]},"next_evidence_step":"Select one authorized corpus slice containing at most 150 works and three adjacent genre categories with an existing documented recalibration rule. Preregister the state representation, perturbation range, reference-case criteria, residual tolerance, modal-stability threshold, and rival comparisons. In a sandbox, replay five calibration rounds from the same starting snapshot, fit the local Jacobian on early rounds, and test prediction on later rounds and a held-out set of independently adjudicated multi-label works. Compare no gate, per-work confidence review, fixed anchors, and the sensitivity-selected modal gate. Preserve all candidate memberships in the outputs and publish nothing from the pilot.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"This proposal is independent of all three earlier proposals. Proposal 1 concerns creative revision of an unreliable narrator, models prose-feature changes between manuscript drafts, and offers alternative edit batches to preserve author-defined ambiguity; this proposal concerns corpus metadata feedback, models work–genre memberships under prototype recalibration, and gates prototype contributions without editing literary text. Proposal 2 concerns factual and interpretive dependencies among critical-edition notes, uses a claim network to order source verification, and seeks apparatus consistency; this proposal neither traces claim dependencies nor prioritizes fact checking, but controls feedback between genre assignments and category prototypes. Proposal 3 concerns live seminar dynamics, models turn-to-turn interpretive convergence, and applies a temporary facilitation reset; this proposal has no classroom interaction or discussion intervention and instead governs reproducibility of released corpus classifications. Its artifact, actors, state representation, transformation, objective, causal path, and authorized action are distinct from each earlier opportunity, and it can be adopted without any manuscript, apparatus, or seminar system.","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}