{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp03_full320_20260801","cell_id":"invariant_mode_decomposition_design__pharmacology_toxicology","trajectory_id":"R","attempt_index":0,"archetype_slug":"invariant_mode_decomposition_design","domain_slug":"pharmacology_toxicology","decision":"CANDIDATE","problem_id":"coupled_biomarker_dynamics_mask_repeat_dose_toxicity","causal_lever_id":"harm_weighted_toxicity_mode_detection","proposal":{"problem":"In repeat-dose toxicology, organ-injury, exposure, and physiological measurements can change as coupled combinations while each reported variable remains below its individual alert threshold. Coordinate-by-coordinate review can therefore miss a persistent or growing toxicity trajectory until overt injury appears.","actors_substrate":["Study animals or other experimental biological systems under repeated exposure","Toxicologists and pharmacologists interpreting longitudinal measurements","Protocol-governance and animal-welfare authorities","Future trial participants or patients affected by downstream safety decisions","Longitudinal biomarker, exposure, pathology, and physiological data"],"observable_state":"Across successive observations under a defined exposure regimen, several measurements exhibit repeatable joint movement, yet univariate alerts are absent, inconsistent, or delayed; an estimated local state-transition model contains persistent or growing combinations that are not visually salient in the original variables.","consequence":"Potential toxicity may be detected late, dismissed as scattered noise, or attributed to the wrong variable, degrading safety-margin estimation, monitoring design, and dose-selection decisions.","affected_objective":"Earlier and more reliable identification of biologically consequential toxicity trajectories without increasing unsupported safety alarms.","structural_mapping":[{"archetype_element":"Transformation acting on many variables","domain_realization":"A prespecified one-step transition from the current multivariate biological state to the next observation under a fixed dose and sampling regime.","claim_kind":"INFERENCE"},{"archetype_element":"Invariant directions","domain_realization":"Approximately self-preserving combinations of exposure, injury, and physiological measurements within a bounded operating window.","claim_kind":"HYPOTHESIS"},{"archetype_element":"Scalar modal response","domain_realization":"Estimated persistence, decay, oscillation, or growth of each joint biological trajectory.","claim_kind":"HYPOTHESIS"},{"archetype_element":"Action-relevant mode ranking","domain_realization":"Modes ranked jointly by gain, association with prespecified harm endpoints, and sensitivity to feasible dose, schedule, or monitoring changes.","claim_kind":"HYPOTHESIS"},{"archetype_element":"Residual visibility","domain_realization":"Held-out prediction and reconstruction errors retain evidence of toxicity patterns excluded by the selected modes.","claim_kind":"INFERENCE"},{"archetype_element":"Drift and interpretation limits","domain_realization":"Modes are trusted only for the studied compound, population, exposure range, sampling interval, and regime in which stability and residual criteria hold.","claim_kind":"INFERENCE"}],"component_map":[{"component":"Transformation Scope","status":"adapted","domain_realization":"One-step longitudinal state transition under a specified compound, exposure regimen, population, and sampling interval."},{"component":"State-Vector Definition","status":"direct","domain_realization":"Prespecified exposure, organ-injury, physiological, and pathology-linked measurements with units, timing, and missingness rules fixed before fitting."},{"component":"Invariant Mode Basis","status":"adapted","domain_realization":"Eigenvectors of an estimated local transition operator, interpreted only as approximate joint trajectories."},{"component":"Modal Gain Spectrum","status":"direct","domain_realization":"Eigenvalues with uncertainty intervals describing each mode's local persistence or amplification."},{"component":"Dominant Mode Selection Rule","status":"adapted","domain_realization":"Retain a mode only when gain or persistence, harm-weighted sensitivity, and held-out fidelity exceed preregistered thresholds."},{"component":"Stable/Unstable Mode Partition","status":"adapted","domain_realization":"Classify modes relative to the discrete-time stability boundary, with an indeterminate class for uncertainty crossing that boundary."},{"component":"Modal Intervention Map","status":"adapted","domain_realization":"Map candidate dose, schedule, assay, or monitoring changes into predicted mode changes; treat these mappings as hypotheses until experimentally tested."},{"component":"Reconstruction Residual Check","status":"direct","domain_realization":"Measure held-out state-prediction error and inspect residuals for structured harm-related signals."},{"component":"Mode Drift Monitor","status":"direct","domain_realization":"Re-estimate modal direction, ordering, and gain across time blocks, dose groups, and replicate studies."},{"component":"Interpretation Scope Contract","status":"direct","domain_realization":"Forbid extrapolation beyond the compound, population, dose range, sampling cadence, and local state region validated."},{"component":"Mode-Coupling Register","status":"direct","domain_realization":"Record near-degeneracy, non-orthogonality, correlated uncertainty, and cross-mode responses to perturbations."},{"component":"Local Linearization Window","status":"direct","domain_realization":"Define the exposure, time, and biological-state window in which held-out errors support a linear approximation."},{"component":"Spectral Gap Threshold","status":"adapted","domain_realization":"Require a preregistered, uncertainty-adjusted separation between retained and discarded modes before using a reduced modal interpretation."}],"mechanism_dispositions":[{"slug":"eigendecomposition_workflow","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Apply complete eigendecomposition to the explicitly fitted local transition matrix and report conditioning and eigenvalue uncertainty.","counterfactual_removal":"Without it, invariant trajectories and their gains are not obtained, so the proposed lever disappears."},{"slug":"modal_sensitivity_sweep","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Perturb modal coordinates and feasible control inputs locally, ranking changes by prespecified toxicity outcomes and logging cross-effects.","counterfactual_removal":"Removing it leaves a descriptive spectrum without evidence about actionable or harm-relevant leverage."},{"slug":"modal_stability_analysis","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Classify estimated modes within the declared linearization window and preserve an uncertainty-aware indeterminate category.","counterfactual_removal":"Removing it prevents distinction between decaying variation and trajectories that persist or grow."},{"slug":"mode_shape_testing","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"The packet's physical excitation-and-sensor protocol does not directly match a bounded retrospective longitudinal toxicology test; empirical validation is supplied by held-out trajectories instead.","counterfactual_removal":"No change because it is not used."},{"slug":"network_spectral_centrality_analysis","disposition":"incompatible","contribution_type":"NONE","adaptation_or_rejection":"Biomarker coordinates are not justified as graph nodes whose dominant-eigenvector entries represent biological importance.","counterfactual_removal":"No change because network centrality is not part of the causal chain."},{"slug":"power_iteration_probe","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"A dominant-only estimate would hide secondary harmful modes and is unstable when the leading gap is small; the bounded test requires the complete fitted spectrum.","counterfactual_removal":"No change because complete eigendecomposition is used."},{"slug":"principal_component_analysis","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Variance directions do not encode longitudinal transition or harm relevance and could suppress a low-variance toxicity signal.","counterfactual_removal":"No change because covariance variance is not the proposed lever."},{"slug":"reduced_order_model","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"A runnable surrogate is unnecessary for the first shadow test and would add extrapolation risk before modal validity is established.","counterfactual_removal":"No change to detection in the bounded test."},{"slug":"residual_reconstruction_test","disposition":"selected_load_bearing","contribution_type":"TEST_DESIGN","adaptation_or_rejection":"Use temporally and biologically held-out observations, assess both residual magnitude and structured association with harm endpoints, and sweep retained-mode count.","counterfactual_removal":"Without it, omitted toxicity structure and overfitting cannot be detected, hard-gating viability."},{"slug":"singular_value_decomposition","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Singular vectors may diagnose amplification or conditioning but are not invariant directions of the fitted transition; use only in a later non-normality audit.","counterfactual_removal":"No change to the stated eigenmode test."},{"slug":"spectral_decomposition_report","disposition":"selected_supporting","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Document biological loadings, uncertainty, couplings, residuals, descriptive-versus-causal limits, and the validated operating window.","counterfactual_removal":"The analysis could still run, but decision-makers would lack enforceable interpretation boundaries."},{"slug":"spectral_gap_monitor","disposition":"selected_load_bearing","contribution_type":"OPERATIONAL","adaptation_or_rejection":"Track uncertainty-adjusted retained/discarded separation and modal rotation across time blocks and groups.","counterfactual_removal":"A fixed basis could remain in use after loss of dominance or mode identity, making continued use unsafe."}],"causal_chain":["Repeated exposure induces coupled biological state changes that may be weak in individual coordinates.","A bounded longitudinal transition model estimates how combinations of measurements propagate between observations.","Eigendecomposition exposes approximate joint trajectories and their persistence or growth.","Stability analysis and harm-weighted sensitivity distinguish actionable risk modes from loud but irrelevant variation.","Residual, conditioning, gap, and drift checks determine whether the modal account is fit for use.","Only validated modes inform confirmatory assay priorities or hypotheses about dose, schedule, and monitoring changes; governed experiments must test any actual control effect."],"baseline":"Ordinary practice reviews each biomarker trend and threshold, exposure summary, clinical observation, and pathology endpoint largely in its named coordinate, with expert integration after alerts occur.","nearest_rival":"A supervised multivariate toxicity predictor or conventional PK/PD model that predicts outcomes directly but does not expose invariant transition modes, modal stability, or spectral-gap failure.","authority_safety":{"affected_parties":["Study animals","Laboratory and toxicology personnel","Sponsors and regulators relying on safety interpretation","Future trial participants or patients"],"decision_authority":"The toxicology study director may authorize shadow analysis; any prospective exposure or sampling change requires the applicable protocol owner, veterinarian, animal-welfare body, and organizational safety governance.","authorized_first_step":"Run a preregistered shadow analysis on one completed repeat-dose study, fixing outcomes and thresholds in advance, fitting on designated animals and early time points, and comparing blinded held-out detection timing, false alerts, residuals, conditioning, and drift against the baseline and nearest rival. No decisions are changed.","excluded_actions":["Changing an ongoing dose or schedule from model output alone","Adding animals, procedures, or samples without protocol approval","Using a mode as proof of a biological mechanism","Suppressing discordant biomarkers or pathology findings","Applying the result to another compound, population, or clinical setting without revalidation"],"halt_rollback":"Halt if conditioning is unacceptable, the spectral gap or modal identity is unstable, held-out residuals exceed the preregistered harm-weighted budget, or false alerts exceed baseline tolerance. Discard the modal decision layer, retain ordinary review, and archive the model and failure evidence."}},"negative_tests":{"strongest_counterevidence":"Toxic responses may be dominated by nonlinear thresholds, irreversible lesions, changing metabolism, sparse sampling, or one directly observable marker; an estimated linear operator may then yield unstable mathematical mixtures rather than reproducible toxicity trajectories.","analogy_break":"Unlike independent physical vibration modes, biological eigenmodes need not correspond to separable mechanisms or independently controllable processes. Non-normal transient amplification can also produce harm even when all eigenvalues appear locally stable.","failure_condition":"No usable candidate exists if no local window produces reproducible modes with acceptable conditioning, stable identity, adequate spectral separation, and harm-weighted held-out residuals within budget.","problem_falsifier":"Across blinded replicate data, univariate review detects consequential toxicity at least as early and accurately as multivariate methods, and no repeatable coupled transition signal remains after accounting for exposure, time, and measurement error.","intervention_falsifier":"Even when reproducible risk modes exist, mode-ranked confirmatory assays or governed dose/schedule hypotheses fail to improve detection, prediction, or control over the nearest rival in a prospective bounded test.","risks":["Spurious modes from small samples, missingness, scaling choices, or correlated measurement error","False reassurance from local stability despite nonlinear or transient toxicity","Overinterpretation of mathematical loadings as biological mechanisms","Mode swapping or rotation near a small spectral gap","Animal-welfare harm if exploratory output changes exposure without approval","Loss of rare but consequential signals during modal compression"]},"null_rationale":null,"classification":{"candidate_kind":"MECHANISM_COMPOSITION","prior_art_status":"UNSEARCHED","evidence_maturity":"HYPOTHESIS"},"revision_change_log":{"revision_kind":"ORIGINAL","prior_problem_id":null,"prior_causal_lever_id":null,"problem_changed":false,"causal_lever_changed":false,"conceptual_changes":[],"operational_changes":[],"repairs_addressed":[]},"confidence":0.78,"generator_notes":"Closed-book structural candidate. Domain-specific performance, reproducibility, thresholds, and biological interpretations are explicitly hypotheses or inferences rather than established claims."}