{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp06_four_proposal_generalization60_20260803","cell_id":"predictive_residual_processing__aviation_aeronautics","arm":"COMPLETE_PROPOSAL_PORTFOLIO","candidate_id":"prp-aviation-composite-inspection-residual-review-002","proposal_index":2,"version":0,"title":"Residual-First Review of Repeat Composite Ultrasound Inspections","problem":"During repeat phased-array ultrasonic inspection of an aircraft composite panel, an inspector may need to review a complete C-scan dominated by stable geometry, material texture, fastener effects, and previously documented features. Acquisition and registration differences can further clutter comparisons. A small new or changed indication may therefore compete for attention with predictable content, while simple image subtraction cannot reliably distinguish structural change from probe, coupling, calibration, or alignment variation.","actors":["Qualified nondestructive-testing inspector","Aircraft structural engineer","Maintenance quality-assurance reviewer","Inspection-data acquisition technician","Maintenance organization configuration manager","Aircraft operator or asset owner","Inspection workstation and model owner"],"observable_state":"For a panel with prior compatible inspections, the current full-resolution C-scan contains large regions whose registered signal structure resembles the expected response, yet the inspector still traverses the entire image. Fixed difference maps also highlight broad residuals after changes in probe path, coupling, calibration, repair configuration, or registration. Review records can identify which regions attracted attention, which required repeated rescanning, and which indications were ultimately accepted, monitored, or referred for engineering disposition.","consequence":"Predictable image content and acquisition artifacts can consume inspection attention, delay structural disposition, or obscure a localized change. Conversely, aggressive automation could suppress a material indication by treating it as expected texture or could elevate benign registration error as damage.","affected_objective":"Concentrate qualified inspector attention on trustworthy changes between expected and observed composite-panel response while preserving complete raw inspection access, explicit uncertainty, independent audit coverage, and sole human authority over structural disposition.","intervention":"Add a residual-review mode to an offline inspection workstation while retaining the complete current C-scan as the authoritative record. For an explicitly eligible panel and acquisition setup, a versioned model predicts the current registered ultrasonic response from panel geometry, prior compatible scans, documented repairs, acquisition parameters, and local uncertainty. The workstation compares the prediction with the actual amplitude, time-of-flight, and signal-quality fields to create signed, spatially structured residual tiles. A consequence- and precision-weighted gate orders tiles for first-pass review using residual structure, registration confidence, coupling quality, proximity to critical geometry, and uncertainty rather than residual magnitude alone. The inspector can reconstruct every tile as expected response plus residual and can reveal the raw waveform or entire scan at any time. Known repairs, mandated inspection zones, poor coupling, calibration failure, incompatible acquisition settings, weak registration, model staleness, and unexplained residual structure bypass suppression and require raw or full-scan review. Random raw tiles and risk-stratified regions are inserted into the review sequence without telling the production model which were selected. Inspector-confirmed outcomes enter a replay set, but model or threshold revisions occur only after separate engineering and quality review.","structural_mapping":[{"archetype_element":"Prediction target and observation boundary","domain_realization":"The target is the spatially registered amplitude, time-of-flight, and acquisition-quality response for declared regions of one identified composite panel under a compatible inspection procedure; raw waveforms remain outside suppression and available on demand."},{"archetype_element":"Generative model state","domain_realization":"A versioned expected-scan model combines prior compatible scans, panel geometry, documented repairs, probe configuration, calibration state, and local uncertainty."},{"archetype_element":"Model scope, horizon, and validity","domain_realization":"Prediction is authorized only for the identified panel, inspection method, acquisition configuration, registration tolerance, and recency window represented in the model manifest."},{"archetype_element":"Expected and actual behavior","domain_realization":"The workstation stores the predicted response before comparison and separately captures the current scan with aircraft, panel, probe, calibration, timestamp, operator, and quality provenance."},{"archetype_element":"Prediction comparator","domain_realization":"A registered comparator produces signed spatial residual fields and preserves whether discrepancies arise in amplitude, time of flight, waveform shape, or data quality."},{"archetype_element":"Precision weighting and residual budget","domain_realization":"Residual priority reflects registration confidence, coupling and calibration quality, repeatability, structural consequence zone, persistence across adjacent probe paths, and the tolerated burden of unreviewed residual mass."},{"archetype_element":"Residual propagation channel","domain_realization":"A review queue presents prioritized residual tiles with expected context, uncertainty, and reconstruction controls instead of initially drawing inspector attention uniformly across the complete scan."},{"archetype_element":"Synchronization and provenance","domain_realization":"Panel identity, coordinate transform, acquisition procedure, prior-scan identifiers, and model checksum must agree before a residual tile can be interpreted."},{"archetype_element":"Raw audit and reconstruction","domain_realization":"The workstation retains the complete current scan, inserts independently sampled raw regions into review, and verifies that expected response plus stored residual reproduces each selected region within declared tolerances."},{"archetype_element":"Drift and decompression","domain_realization":"Structured residuals, calibration or coupling failures, incompatible configuration, registration uncertainty, staleness, or cumulative audit disagreement suspend residual-first review for the affected region or entire scan."},{"archetype_element":"Bounded update loop","domain_realization":"Inspector dispositions and engineering findings are replayed after the inspection to classify model, data, registration, threshold, or scope errors; approved changes are versioned for future inspections rather than learned during the current disposition."}],"mechanism_mapping":[{"mechanism_slug":"anomaly_detection_model","role":"Models the context-conditioned expected ultrasonic response and identifies spatial observations that the model does not explain.","counterfactual_removal":"Without an expected-response model, the workstation can flag absolute limits or visual salience but cannot separate predictable panel texture from model-relative change."},{"mechanism_slug":"delta_or_differential_encoding","role":"Represents the current registered scan as expected response plus signed spatial corrections, making the residual the initial unit of inspector attention while retaining reconstructability.","counterfactual_removal":"The proposal becomes an ordinary overlay or classifier because the current observation is no longer explicitly represented and recoverable as a correction to a maintained expectation."},{"mechanism_slug":"precision_weighted_error_gate","role":"Orders residual tiles using measurement reliability, registration confidence, structural consequence, persistence, and review cost while preserving suppressed residuals for audit.","counterfactual_removal":"Large coupling or alignment artifacts could dominate the queue while smaller, reliable changes in consequential regions receive less attention."},{"mechanism_slug":"model_version_checksum_handshake","role":"Verifies that the panel baseline, coordinate transform, acquisition configuration, and predictor version used to produce a residual match those used to display and reconstruct it.","counterfactual_removal":"A residual could be displayed against the wrong prior scan, panel revision, or coordinate transform and still appear visually coherent."},{"mechanism_slug":"residual_comparison_test","role":"Tests residual fields for spatial correlation, bias, acquisition-linked structure, and disagreement with a simpler registered-difference model or independent raw sample.","counterfactual_removal":"Systematic misspecification or registration error could be mislabeled as random noise or as a localized structural indication."},{"mechanism_slug":"shadow_raw_channel_sampling","role":"Introduces random and risk-stratified full-resolution regions into an independent audit stream and compares raw observations with model-based reconstructions and review decisions.","counterfactual_removal":"The system would evaluate only the regions its own gate selected, allowing shared blind spots to remain invisible."},{"mechanism_slug":"raw_signal_fallback_switch","role":"Requires full-region or full-scan review when identity, calibration, coupling, registration, scope, staleness, audit, or consequence conditions invalidate residual suppression.","counterfactual_removal":"The predictor could continue filtering inspector attention when the scan is incompatible with its assumptions."},{"mechanism_slug":"bayesian_model_update","role":"After independent disposition, carries uncertainty forward when approved evidence revises the expected response for a panel region or acquisition condition.","counterfactual_removal":"The baseline would either remain fixed despite validated changes or be replaced without retaining uncertainty and attribution."},{"mechanism_slug":"prediction_error_review","role":"Requires material misses and false escalations to be classified as model, acquisition, registration, scope, threshold, or structural-change issues before any revision is authorized.","counterfactual_removal":"Errors could produce ad hoc threshold changes or untraceable model updates without determining what actually caused the mismatch."}],"causal_chain":["A compatible prior scan, panel configuration, and acquisition manifest establish a versioned prediction of the current ultrasonic response before the new scan is interpreted.","The workstation registers the actual scan, preserves its provenance, and computes signed spatial differences between expected and observed amplitude, timing, waveform, and quality fields.","Precision and consequence weighting separates trustworthy localized mismatch from uncertain coupling, calibration, and registration artifacts and assigns residual tiles to a bounded review queue.","The inspector reviews prioritized residuals together with the expected context and can reconstruct or reveal the corresponding full-resolution raw data immediately.","Independent random raw tiles and protected structural regions test what the predictor and gate would otherwise suppress.","Persistent residual structure, configuration mismatch, poor data quality, audit disagreement, or model staleness triggers decompression to regional or complete raw review.","The qualified inspector and structural engineer retain disposition authority and record whether a residual represented structural change, acquisition error, benign variation, or model failure.","Post-inspection review uses validated discrepancies to revise the model, uncertainty, scope, or threshold under configuration control, improving the next prediction without altering the completed inspection record."],"baseline":"The comparison condition is complete manual traversal of the current C-scan, optionally assisted by side-by-side viewing of a prior scan and fixed display thresholds. The qualified inspector accesses raw waveforms and documents indications without a model-governed residual queue.","nearest_rivals":["Complete manual review of every current-scan region with side-by-side prior imagery","Registered pixel- or voxel-wise subtraction against the immediately preceding scan using fixed thresholds","An anomaly detector or defect classifier applied only to the current scan without reconstructive expected-response modeling","Rule-based prioritization of known critical zones and previously documented indications","Automated change detection that produces a heat map but lacks independent raw sampling, model synchronization, an explicit residual budget, and full-scan fallback"],"remaining_contrastive_claim":"The proposal is not merely a change map or defect classifier: it makes a scoped, versioned expected scan the suppressed representation, preserves the signed correction needed to reconstruct the current observation, allocates human attention through uncertainty- and consequence-weighted residuals, and conditions continued suppression on independent raw audits and decompression triggers. If a simpler registered-difference or full-review workflow satisfies the same preregistered safety, fidelity, and attention criteria with less model-governance burden, this contrastive rationale fails.","authority_safety":{"decision_authority":"Only the qualified nondestructive-testing inspector may accept or reject an indication under the applicable inspection procedure, with structural-engineering and quality-assurance authority unchanged. The residual system may order evidence and request fuller review but may not determine airworthiness, return to service, repair scope, or inspection completion.","authorized_first_step":"Run a retrospective, offline, blinded evaluation on previously completed paired inspections with preserved full-resolution scans and independently established dispositions. The system may simulate review ordering and collect study decisions but may not modify maintenance records, aircraft status, inspection procedures, or actual dispositions.","excluded_actions":["No autonomous airworthiness, return-to-service, repair, or no-fault decision","No deletion, lossy replacement, or withholding of the authoritative current scan or raw waveform data","No use on an unregistered panel, undocumented repair state, incompatible inspection method, or out-of-scope acquisition configuration","No online learning from an inspector click or provisional label during the active inspection","No reduction of procedure-mandated coverage, calibration checks, protected-zone review, or independent quality review","No threshold adjustment to meet a queue-size, review-time, or alert-count target","No representation of a quiet residual queue as proof that the panel is free of reportable indications"],"halt_rollback":"Stop the offline study for unrecoverable provenance gaps, leakage of adjudicated outcomes into predictions, failed reconstruction, or inability to reproduce the source scan. In any later separately authorized shadow trial, checksum mismatch, weak registration, calibration or coupling failure, undocumented configuration change, out-of-scope geometry, stale model state, structured residual drift, or audit-budget breach disables residual-first ordering and returns the affected region or entire inspection to the existing full-review workflow. Re-entry requires corrected acquisition or registration, a compatible model manifest, and approval by the responsible inspector or quality reviewer."},"negative_tests":{"strongest_counterevidence":"A simple registered-difference display or existing full-scan workflow meets the same preregistered indication-capture, reconstruction, uncertainty, and review criteria with less false prioritization and lower configuration, audit, and maintenance burden. This would show that a governed predictive residual loop is not causally necessary for the task.","problem_falsifier":"Task observation and inspection records show that complete-scan traversal does not create a relevant attention or disposition bottleneck for eligible repeat inspections, or the applicable inspection procedure requires an independent full-fidelity evaluation of every region in a way that residual-first ordering cannot assist without duplicating the entire workload.","intervention_falsifier":"In blinded retrospective replay, residual-first ordering fails any preregistered protected-indication criterion, cannot reconstruct selected regions within tolerance, prioritizes acquisition artifacts ahead of adjudicated material changes, shows systematic disagreement in random raw samples, or requires full-scan fallback so frequently that it provides no usable attention distinction from the baseline.","risks":["A strong prior may explain away a subtle new delamination, disbond, impact feature, fluid ingress pattern, or repair-edge change.","Registration error may create plausible residual contours or spatially displace a real indication.","Changes in probe, calibration, coupling, temperature, surface condition, or acquisition procedure may be mistaken for structural change.","A model trained mainly on previously flagged indications may reproduce existing blind spots and underrepresent unanticipated signal forms.","Risk-stratified audits may concentrate on known concerns while missing genuinely novel regions; random audits may still miss rare events.","Inspectors may anchor on residual ordering and give insufficient attention to low-ranked or reconstructed regions.","Repeated accepted indications may be absorbed into the expected model even when their future change remains consequential.","Displaying expected response plus residual may conceal reconstruction or quantization error unless the raw waveform remains immediately accessible.","Model and panel-configuration identifiers may agree while both encode an incorrect or incomplete repair history.","Pressure to shorten review may encourage thresholds that optimize queue size rather than tolerated inspection error."]},"next_evidence_step":"Pre-register one bounded retrospective reader study using no more than forty paired, full-resolution inspections from one documented aircraft-panel and inspection-method family, with adjudicated outcomes concealed from the prediction and readers. Freeze eligibility, registration procedure, model version, protected zones, rival methods, residual budget, reconstruction tolerances, indication-level halt criteria, and analysis plan before scoring. Have qualified readers review cases under the existing full-scan workflow, a fixed registered-difference display, and the residual-first interface in counterbalanced sessions. Record review sequence, raw-data reveals, regions examined, indication decisions, false prioritizations, reconstruction disagreement, random-audit findings, fallback triggers, and total model, synchronization, and audit burden. Deliberately introduce bounded registration, coupling, calibration, and model-version faults to verify decompression behavior. The result may support or reject a later shadow-mode evaluation but cannot authorize reduced inspection coverage or operational disposition.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"Proposal 1 addressed constrained air-to-ground bandwidth during active flight testing by placing matched predictors onboard and on the ground, transmitting residual telemetry, and falling back to full frames. This proposal addresses a different problem: qualified human attention during ground-based repeat ultrasonic inspection of composite structure. It retains the complete scan locally, uses a panel-specific expected image to order inspection evidence rather than conserve a flight link, and closes its causal loop through inspector disposition and configuration-controlled future baselines rather than real-time telemetry reconstruction. It is independently adoptable in a maintenance inspection workstation and neither depends on nor extends the flight-test telemetry system.","revision_record":{"parent_version":null,"progress_targets_addressed":["Created a second complete proposal with a materially different maintenance-inspection problem, attention-allocation intervention, and post-inspection learning path.","Specified explicit contrast with sealed proposal 1 while preserving reconstructive residual processing, independent raw audit, synchronization, and fallback."],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}