{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp06_four_proposal_generalization60_20260803","cell_id":"predictive_residual_processing__religious_studies_theology","arm":"COMPLETE_PROPOSAL_PORTFOLIO","candidate_id":"prp-recurrent-ritual-record-review-001","proposal_index":1,"version":0,"title":"Residual-Led Review of Recurrent Ritual Records","problem":"In a bounded corpus of repeated ritual records—such as successive service transcripts, ceremony descriptions, or editions of a ritual manual—researchers may repeatedly inspect largely recurring descriptive structures while locally consequential changes in sequence, participant role, wording category, omission, or material practice compete for limited review attention. A full-record workflow preserves context but makes every expected unit consume attention; a simple exception workflow risks treating one canonical form as normative and hiding changes that its baseline cannot represent.","actors":["Religious-studies researchers conducting comparative analysis","Corpus curators and descriptive coders","Research assistants reviewing records","Community data stewards or advisors governing sensitive materials","Research-ethics and data-governance reviewers"],"observable_state":"Within an already authorized corpus, each record can be segmented into provenance-linked descriptive units such as actor-role, action or utterance type, object, and sequence position. The relevant state is observable through reviewer time per record, the share of units reconstructed from an expectation versus supplied as residuals, disagreements between residual packets and full-record coding, residual distributions by collection and period, model-version mismatches, raw-fallback frequency, and misses found in independent full-record audits. Predictability must be evaluated separately within each collection and context; no universal model of normal religion is assumed.","consequence":"When recurring descriptive material occupies most first-pass review capacity, researchers may reach potentially important local variants late or inconsistently. If attention is compressed against an ungoverned template, the opposite harm can occur: dissenting, underrepresented, historically changing, or community-specific practice may be suppressed as expected noise or mischaracterized as abnormal. Either condition can weaken comparative interpretation and the traceability of claims to complete sources.","affected_objective":"Allocate bounded scholarly review attention toward model-relative differences while preserving reconstructable descriptive coding, immediate access to complete sources, collection-specific context, coverage of underrepresented practices, and human authority over interpretation.","intervention":"Build a read-only residual-review layer over an authorized, fully retained corpus. For each narrowly scoped collection, a versioned sequence model predicts the next descriptive event-code unit before a held-out record is revealed, conditioned only on declared metadata such as document genre, locality, date band, and ritual position. The prediction target is the descriptive code sequence, not doctrinal truth, authenticity, orthodoxy, or religious value. The comparator produces structured insertions, deletions, substitutions, reorderings, and uncertainty-bearing coding disagreements between predicted and observed units. A consequence-and-precision rule prioritizes residuals using source quality, coder agreement, underrepresentation, community-designated sensitivity, and interpretive materiality; it may collapse only high-confidence expected units in the first-pass interface. Each packet carries model version, scope, source identifiers, timestamps, uncertainty, and links that reconstruct the complete coded sequence and open the untouched source. Validated residuals may update descriptive transition beliefs only after human review, with fast operational thresholds separated from slower model revision. Random and risk-stratified records are always reviewed in full through an independent audit path. Low-confidence records, untranslated passages, consent or access ambiguities, community-designated sensitive content, allegations of coercion or harm, and material from contexts insufficiently represented in the model bypass suppression. Sustained structured residuals, stale scope, checksum mismatch, audit disagreement, or error-budget breach disables residual mode for the affected collection and restores full-record review.","structural_mapping":[{"archetype_element":"Prediction target definition","domain_realization":"The next provenance-linked descriptive event-code unit and its sequence position within one declared ritual-record collection and metadata context."},{"archetype_element":"Generative model state and scope","domain_realization":"A versioned, collection-specific probabilistic sequence template with uncertainty; it is prohibited from representing doctrinal correctness or serving outside its documented genre, community, period, and coding scheme."},{"archetype_element":"Expected input","domain_realization":"The descriptive unit or short sequence predicted before the corresponding held-out record segment is opened."},{"archetype_element":"Actual observation","domain_realization":"The coder-observed event unit linked to the untouched source span, coder identity, timestamp, and source-permission state."},{"archetype_element":"Prediction comparator","domain_realization":"A declared structured comparison that preserves insertion, deletion, substitution, reordering, and coding disagreement rather than reducing difference to a single anomaly score."},{"archetype_element":"Prediction-error signal","domain_realization":"A residual dossier containing the structured difference, expected context, actual source pointer, uncertainty, and model version."},{"archetype_element":"Precision weighting and residual budget","domain_realization":"A versioned table weights residuals by coder agreement, source quality, interpretive consequence, underrepresentation, sensitivity, and review cost; suppressed descriptive error must remain within a predeclared audit tolerance."},{"archetype_element":"Residual propagation channel","domain_realization":"A scholar-facing queue that foregrounds admitted residuals, collapses eligible expected units, verifies heartbeat and completeness, and permits immediate expansion to the full record."},{"archetype_element":"Reconstruction and synchronization","domain_realization":"The interface reconstructs the complete coded sequence from the matching model version plus residuals, verifies the version checksum, and periodically reanchors against a complete coding snapshot."},{"archetype_element":"Update rule","domain_realization":"Only human-validated residuals can revise probabilities over descriptive sequences; changes are versioned, attributable, reversible, and reviewed separately from interpretations."},{"archetype_element":"Drift and validity control","domain_realization":"Residual structure, audit disagreement, context shifts, model age, and new coding-schema versions determine whether the model remains valid for a collection."},{"archetype_element":"Raw-signal audit","domain_realization":"Independent reviewers receive a random baseline sample plus risk-stratified records in complete, uncollapsed form and compare their coding with the residual reconstruction."},{"archetype_element":"Safety-critical bypass and fallback","domain_realization":"Sensitive, rights-relevant, low-confidence, insufficiently represented, permission-ambiguous, or harm-related material travels in full; any validity failure returns the affected scope to full-record review."},{"archetype_element":"Attention and bandwidth budget","domain_realization":"The scarce resource is first-pass researcher attention, measured together with model maintenance, audit, fallback, and reconstruction costs rather than by queue reduction alone."}],"mechanism_mapping":[{"mechanism_slug":"forecast_backtesting","role":"Walk-forward testing on previously coded records defines the collections, contexts, and horizons in which the descriptive sequence predictor may be used.","counterfactual_removal":"Without held-out testing, there is no bounded evidence that expected units are predictable enough to suppress, so residual mode lacks a license to operate."},{"mechanism_slug":"predictive_codec","role":"The review layer represents an eligible record as a matched versioned expectation plus structured corrections, from which the full descriptive code sequence is reconstructed.","counterfactual_removal":"Without prediction-plus-correction reconstruction, the intervention becomes ordinary anomaly highlighting rather than predictive residual processing."},{"mechanism_slug":"precision_weighted_error_gate","role":"It allocates limited review slots using uncertainty, source reliability, underrepresentation, sensitivity, and interpretive consequence while logging what was collapsed.","counterfactual_removal":"Without the gate, either every difference floods the queue or a magnitude-only cutoff systematically hides small but consequential or well-supported changes."},{"mechanism_slug":"model_version_checksum_handshake","role":"It prevents a residual generated against one collection model or coding scheme from being interpreted against another.","counterfactual_removal":"Without version compatibility checks, a well-formed packet could silently reconstruct the wrong descriptive sequence or baseline context."},{"mechanism_slug":"bayesian_model_update","role":"After human validation, descriptive transition beliefs and their uncertainty are revised while preserving prior and posterior versions for challenge and rollback.","counterfactual_removal":"Without an uncertainty-bearing update rule, the expectation remains static and recurring residuals do not improve or invalidate the model in a controlled way."},{"mechanism_slug":"model_drift_monitoring","role":"It detects sustained residual structure, context change, calibration loss, or staleness that makes a collection-specific expectation unsafe to reuse.","counterfactual_removal":"Without drift monitoring, historical change could be repeatedly treated as isolated exceptions while an obsolete baseline continues suppressing context."},{"mechanism_slug":"shadow_raw_channel_sampling","role":"A random and risk-stratified set of complete records is independently reviewed to reveal meaning-bearing material that the predictor or coding scheme failed to surface.","counterfactual_removal":"Without independent full-record sampling, the residual stream would grade itself and could not expose systematic suppression blind spots."},{"mechanism_slug":"raw_signal_fallback_switch","role":"It immediately restores complete-record presentation for invalid, sensitive, mismatched, or out-of-scope cases and uses hysteresis before residual mode resumes.","counterfactual_removal":"Without full-record fallback, model failure would leave scholars dependent on an incomplete representation precisely when the expectation is least trustworthy."},{"mechanism_slug":"prediction_error_review","role":"Researchers and data stewards classify material misses as model, coding, source, scope, or interpretation-boundary problems before authorizing changes.","counterfactual_removal":"Without structured review, residuals might attract attention but would not produce accountable learning, scope correction, or retirement decisions."}],"causal_chain":["A collection-specific model predicts a descriptive ritual-event sequence before each held-out record is exposed.","The observed provenance-linked coding is compared with that prediction to form structured, signed residuals.","Uncertainty, source quality, underrepresentation, sensitivity, consequence, and attention cost determine which residuals are foregrounded and which expected units may be collapsed.","The matching model version plus residual packet reconstructs the complete coded sequence, while the untouched source remains directly accessible.","Researchers spend first-pass attention on admitted residuals and expand complete context before making interpretive judgments.","Independent full-record audits measure what the residual interface suppressed and whether reconstruction stayed within the declared tolerance.","Validated residuals revise descriptive expectations or narrow their authorized scope; structured errors, drift, or audit failures trigger complete-record fallback.","The loop continues only while attention saved exceeds model, audit, synchronization, and fallback costs without breaching fidelity or governance constraints."],"baseline":"A controlled comparator condition in which reviewers read and code every authorized record sequentially in full, using the same source access, codebook, personnel, and interpretive questions but without predicted collapse or residual prioritization. Time, coding disagreements, identified material variants, source expansions, and governance incidents are recorded for both conditions.","nearest_rivals":["A plain textual diff against one selected canonical record, which exposes literal edits but lacks a context-conditioned generative expectation, uncertainty, model updating, and independent raw-audit governance.","Keyword, concordance, or topic-based search, which retrieves specified features but does not reconstruct each record as expectation plus observed residual.","A standalone anomaly detector that ranks unusual records but does not synchronize a reconstructive baseline, distinguish missingness from confirmation, or use residuals in a governed learning loop.","Random or stratified manual sampling of full records, which protects raw access but does not use maintained predictions to concentrate attention between sampled records.","Manual exception reporting by coders, which can foreground differences but leaves the baseline, thresholds, suppressed stream, and update history implicit."],"remaining_contrastive_claim":"Relative to the listed rivals, the testable distinction is the joint architecture: a pre-observation, collection-scoped expectation; structured residual representation; version-compatible reconstruction; uncertainty-and-consequence routing; validated model revision; independent full-record auditing; and automatic decompression when validity fails. The proposal should be rejected as an instance of this archetype if those elements do not operate as one loop.","authority_safety":{"decision_authority":"A corpus-governance committee controls eligible sources, model scope, bypass classes, and stopping rules; community data stewards retain authority over consent and culturally restricted materials; the principal investigator may authorize only the bounded pilot; individual scholars retain sole authority over interpretation and claims.","authorized_first_step":"Run a read-only retrospective pilot on already authorized and fully coded records, with a frozen model, predeclared audit sample, and no alteration of sources, permissions, published claims, or community practices.","excluded_actions":["No automated judgment of orthodoxy, authenticity, sincerity, doctrinal correctness, or religious value","No universal or cross-tradition model of normal religious practice","No deletion, replacement, or lossy alteration of complete source records","No expansion of access to restricted, sacred, personal, or community-governed material","No model update from an unreviewed residual or from consequences created by the interface itself","No publication, personnel evaluation, pastoral action, community intervention, or adverse inference based solely on a residual score","No lowering of bypass or audit protections merely to reduce reviewer workload"],"halt_rollback":"Immediately disable residual mode for the affected collection upon a permission conflict, sensitive-content exposure, bypass failure, model-version mismatch, missing heartbeat, systematic audit disagreement, subgroup or context-specific reconstruction failure, or excess total review cost. Revert all users to the untouched full-record interface, freeze model updates, preserve attributable logs, notify the governance authority, and require explicit reauthorization before a new version can resume."},"negative_tests":{"strongest_counterevidence":"Independent full-record readers repeatedly find interpretively material context inside units the residual interface classified as expected—especially in underrepresented, dissenting, historically changing, or community-designated sensitive records—while residual-led readers do not recover that context when answering the same research questions.","problem_falsifier":"Time-and-task observation in the scoped corpus shows that descriptive sequences are not sufficiently predictable, that full-record reading is not the binding attention cost, or that researchers already reach material variants without a residual queue; under any of these findings, the inferred structural problem is absent for this setting.","intervention_falsifier":"On held-out records, the expectation-plus-residual representation cannot reconstruct the full descriptive coding within the predeclared tolerance, misses any mandatory-bypass item, produces systematic errors by collection or context, or consumes at least as much total researcher effort after model maintenance, audits, and fallback as the full-record baseline.","risks":["A collection-specific expectation could acquire unwarranted normative authority and be mistaken for a canonical ritual form.","Small or underrepresented traditions could be assigned low precision and suppressed unless representation and consequence protections override frequency.","Semantic significance may reside in apparently routine wording, performance, silence, or material context that a descriptive event code omits.","Residual queues can overemphasize novelty and encourage scholars to treat difference as intrinsically more important than continuity.","Sensitive outliers may become easier to identify because residuals concentrate atypical information.","Coder disagreement or translation uncertainty may be mistaken for change in the religious practice itself.","Human reviewers may tune thresholds toward a comfortable queue size rather than the declared fidelity and harm budget.","Model updates may absorb gradual change as expected and thereby erase evidence of historical transformation.","The interface may contaminate later coding by teaching coders what the model expects."]},"next_evidence_step":"Select 60 already authorized, fully coded records from one narrowly bounded corpus and freeze their permissions and codebook. Use the earliest 30 records only to fit a transparent descriptive sequence model; reserve the later 30 as untouched walk-forward cases. Before testing, register prediction scope, reconstruction tolerance, mandatory full-display classes, audit rules, and stopping conditions. Have one reviewer use the residual interface and another use complete records, counterbalancing record order; require both to reconstruct the descriptive sequence and answer the same limited variant-identification questions. Independently inspect a random sample of residual-collapsed records plus every risk-triggered record in full. Record review time, reconstruction disagreement, missed prespecified variants, context expansions, version or missingness failures, fallback use, and governance incidents. This step licenses only a decision to reject, revise, or continue testing the interface; it does not authorize substantive religious conclusions or deployment.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"No prior proposal context was consulted, and distinctness from any prior candidate is not asserted; this candidate is defined solely by its religious-studies realization of the supplied predictive-residual structure.","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}