{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp03_full320_20260801","cell_id":"invariant_mode_decomposition_design__tech_ethics_ai_governance","trajectory_id":"R","attempt_index":0,"archetype_slug":"invariant_mode_decomposition_design","domain_slug":"tech_ethics_ai_governance","decision":"CANDIDATE","problem_id":"coupled_ai_governance_risk_drift","causal_lever_id":"unstable_governance_mode_targeting","proposal":{"problem":"AI oversight programs often monitor fairness, reliability, privacy, human-override use, incident reporting, deployment scale, and remediation backlog as separate indicators. A recognizable failure occurs when a coupled combination deteriorates across review periods while every individual metric remains below its escalation threshold, delaying intervention until harm or organizational lock-in is substantial.","actors_substrate":["people subject to AI-assisted decisions","AI system operators and frontline reviewers","model owners and vendors","risk, compliance, audit, and ethics teams","executives or public officials authorizing deployment","periodic multivariate governance records linked to one system or comparable deployment units"],"observable_state":"A time-indexed vector of governance indicators shows repeatable cross-indicator movement; joint trajectories predict escalation, incidents, or control failure better than the dashboard's separate thresholds.","consequence":"A weakly damped or growing joint-risk pattern can evade coordinate-wise controls, producing late escalation, misdirected remediation, and avoidable impacts on affected people.","affected_objective":"Detect and contain material AI-governance risk early without treating a locally fitted statistical mode as proof of causation or as authority for adverse decisions.","structural_mapping":[{"archetype_element":"coupled transformation","domain_realization":"INFERENCE: transitions between review-period governance-state vectors can be represented by an explicitly estimated local operator for a bounded deployment regime.","claim_kind":"INFERENCE"},{"archetype_element":"hidden invariant directions","domain_realization":"HYPOTHESIS: recurring combinations such as rising reliance, falling override use, and growing remediation backlog may persist or amplify even when no constituent metric breaches its threshold.","claim_kind":"HYPOTHESIS"},{"archetype_element":"scalar modal response","domain_realization":"The estimated gain of each joint pattern indicates decay, persistence, oscillation, or growth within the fitted window; it is predictive rather than automatically causal.","claim_kind":"INFERENCE"},{"archetype_element":"mode-targeted control","domain_realization":"Candidate governance actions are ranked by their estimated effect on the risky joint pattern and checked for cross-effects, rather than selected from the largest raw metric alone.","claim_kind":"HYPOTHESIS"},{"archetype_element":"residual and drift governance","domain_realization":"Reconstruction error, out-of-window performance, basis rotation, and spectral-gap erosion determine when the modal account must be revised or retired.","claim_kind":"INFERENCE"}],"component_map":[{"component":"Transformation Scope","status":"adapted","domain_realization":"One AI deployment or a homogeneous deployment class, with a review-period-to-review-period state transition estimated only within a declared regime."},{"component":"State-Vector Definition","status":"adapted","domain_realization":"Predefined, auditable indicators covering impacts, exposure, reliance, overrides, incidents, complaints, drift, remediation, and deployment commitments."},{"component":"Invariant Mode Basis","status":"adapted","domain_realization":"Approximately invariant joint indicator directions from the fitted transition operator."},{"component":"Modal Gain Spectrum","status":"direct","domain_realization":"Estimated gain and uncertainty for every retained transition mode."},{"component":"Dominant Mode Selection Rule","status":"adapted","domain_realization":"Retain modes exceeding preregistered persistence, harm-weighted sensitivity, and out-of-sample fidelity thresholds."},{"component":"Stable/Unstable Mode Partition","status":"adapted","domain_realization":"Classify modes as decaying, marginal, growing, or oscillatory inside the local validity window."},{"component":"Modal Intervention Map","status":"adapted","domain_realization":"Map reversible controls—enhanced review, deployment pause, staffing, override redesign, or remediation priority—to predicted modal changes."},{"component":"Reconstruction Residual Check","status":"direct","domain_realization":"Compare reconstructed and observed governance states, including harm-relevant structure in residuals."},{"component":"Mode Drift Monitor","status":"direct","domain_realization":"Track mode rotation, gain changes, residual growth, and ordering changes after updates or policy shifts."},{"component":"Interpretation Scope Contract","status":"direct","domain_realization":"Document descriptive-versus-causal status, included populations, missing indicators, validity period, and prohibited uses."},{"component":"Mode-Coupling Register","status":"adapted","domain_realization":"Record near-degenerate modes and intervention cross-effects that defeat independent-mode interpretation."},{"component":"Local Linearization Window","status":"adapted","domain_realization":"Limit inference to specified deployment scale, model version, policy, population, and operating conditions."},{"component":"Spectral Gap Threshold","status":"direct","domain_realization":"Require adequate separation and uncertainty margins between retained and discarded modes before simplified decisions are permitted."}],"mechanism_dispositions":[{"slug":"eigendecomposition_workflow","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Decompose the explicitly estimated review-period transition operator; use uncertainty and conditioning checks because the operator may be non-normal.","counterfactual_removal":"Without a basis and gain spectrum, joint patterns cannot be classified or targeted modally."},{"slug":"modal_sensitivity_sweep","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Perturb candidate controls in a retrospective or simulated model and rank harm-relevant outcome response while registering cross-effects.","counterfactual_removal":"The analysis would rank prominent modes but could not identify which reversible controls plausibly change outcomes."},{"slug":"modal_stability_analysis","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Classify estimated modes within the declared model-version and policy regime.","counterfactual_removal":"A persistent or growing joint-risk direction would not be distinguished from transient variation."},{"slug":"mode_shape_testing","disposition":"incompatible","contribution_type":"NONE","adaptation_or_rejection":"Physical excitation and sensor-based mode identification do not map safely to people or live governance harms; observational validation is supplied by residual testing.","counterfactual_removal":"No change; ethically unacceptable live excitation is not part of the chain."},{"slug":"network_spectral_centrality_analysis","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"The target is temporal coupling among indicators, not node-importance ranking; centrality would answer a different question.","counterfactual_removal":"No change to detection of joint temporal risk drift."},{"slug":"power_iteration_probe","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Dominant-only recovery would hide secondary harm-relevant modes and becomes unreliable near a small spectral gap.","counterfactual_removal":"No change because the bounded pilot uses the full estimated spectrum."},{"slug":"principal_component_analysis","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Variance directions do not establish review-period dynamics and may suppress low-variance harm signals.","counterfactual_removal":"No change; transition modes, not variance components, carry the proposed inference."},{"slug":"reduced_order_model","disposition":"selected_supporting","contribution_type":"TEST_DESIGN","adaptation_or_rejection":"Use a shadow, non-production surrogate for reversible counterfactual comparisons only inside the fitted regime.","counterfactual_removal":"The risky mode could still be detected, but inexpensive bounded comparisons of candidate controls would be lost."},{"slug":"residual_reconstruction_test","disposition":"selected_load_bearing","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Test out-of-sample reconstruction and inspect residuals for subgroup or harm-relevant structure.","counterfactual_removal":"Compression error could conceal omitted harms while the modal model appeared adequate."},{"slug":"singular_value_decomposition","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"SVD is reserved as a conditioning diagnostic; paired input-output singular directions are not invariant temporal modes.","counterfactual_removal":"No material change to the stated repeated-transition causal chain."},{"slug":"spectral_decomposition_report","disposition":"selected_load_bearing","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Bind each mode to uncertainty, population, regime, causal limits, prohibited uses, and known coupling.","counterfactual_removal":"Decision-makers could mistake descriptive latent patterns for causal facts or individual risk scores."},{"slug":"spectral_gap_monitor","disposition":"selected_load_bearing","contribution_type":"OPERATIONAL","adaptation_or_rejection":"Re-estimate separation, mode identity, and drift after data refreshes, model updates, or policy changes.","counterfactual_removal":"A once-valid simplification could remain operational after its modal separation disappeared."}],"causal_chain":["Separate metric thresholds ignore coordinated subthreshold movement.","A bounded transition operator estimates how the joint governance state changes between review periods.","Decomposition exposes approximately persistent or amplifying indicator combinations and their gains.","Stability and harm-weighted sensitivity identify a risky mode and candidate reversible controls with leverage over it.","Residual, coupling, gap, and drift checks gate whether that result may inform governance action.","Authorized reviewers use the mode as an early-warning prompt for investigation or reversible system-level controls, reducing delayed escalation if the hypothesis holds."],"baseline":"Ordinary governance uses separate dashboard thresholds, periodic aggregate audits, incident-triggered review, and expert judgment about each indicator.","nearest_rival":"A preregistered multivariate risk score or supervised incident-prediction model using the same indicators, without invariant-mode interpretation.","authority_safety":{"affected_parties":["people evaluated or served by the AI system","workers operating or reviewing the system","complainants and incident reporters","model owners and vendors","governance personnel"],"decision_authority":"The accountable deployment owner and independent risk or ethics review body, with data-owner approval and required worker or affected-party representation, authorize interpretation and any later control.","authorized_first_step":"Run a retrospective, shadow-mode pilot on historical review periods for one deployment: preregister state variables, compare held-out early warning against the baseline and nearest rival, test residuals by affected group, and make no production or individual-level decisions.","excluded_actions":["individual profiling or adverse decisions from modal coordinates","live excitation of risky behavior","automatic deployment suspension without authorized review","causal claims from observational modes alone","use beyond the declared population, model version, or operating window"],"halt_rollback":"Halt and retire the modal alert if held-out performance does not exceed the baseline, harm-relevant residuals exceed tolerance, modes are ill-conditioned or unstable across resamples, the spectral gap falls below its preregistered margin, or subgroup error is unacceptable; retain the existing governance process and delete pilot-derived operational scores."}},"negative_tests":{"strongest_counterevidence":"Separate thresholds or the nearest rival predict escalation equally well or better on held-out periods, while fitted modes rotate substantially across resamples or add no decision-relevant lead time.","analogy_break":"Governance indicators may be socially constructed, strategically reported, delayed, and changed by policy; unlike a stable physical state, their transition operator can change when definitions, incentives, populations, or model versions change.","failure_condition":"There are too few comparable time points, no reasonably stationary local regime, severe unmeasured confounding, or no repeatable coupled transition beyond one directly observable indicator.","problem_falsifier":"The alleged missed joint deterioration is absent: incidents and control failures are consistently preceded by an ordinary single-metric breach, and cross-indicator dynamics add no held-out predictive information.","intervention_falsifier":"A stable risky mode exists, but authorized reversible controls predicted to damp it do not change the mode or downstream harm-relevant outcomes in shadow evaluation or a separately approved bounded pilot.","risks":["Latent modes may launder biased or incomplete indicators into technical authority.","Aggregate fidelity may conceal subgroup-specific residual harm.","Goodhart effects may arise if teams learn which indicator combinations trigger review.","Near-degenerate or non-normal modes may yield unstable interpretations.","A descriptive mode may be misrepresented as a causal mechanism or individual trait.","Premature alerts may impose unnecessary pauses, workload, or compliance theater."]},"null_rationale":null,"classification":{"candidate_kind":"MECHANISM_ADAPTATION","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":0.74,"generator_notes":"Closed-book structural transfer. The candidate depends on a locally estimable, sufficiently stable transition operator and is rejected if ordinary thresholds explain the observed failures."}