{"cases":[{"alternative_route_blocks":[],"case_id":"E15A002__DIRECT","case_type":"DIRECT_POSITIVE","domain":"Hospital readmission prediction","intended_route_id":"B2","omitted_condition_id":null,"remedy_leakage_audit":"The scenario reports the problem and its operational consequence without proposing mitigation, redesign, or corrective action.","route_evidence":[{"condition_id":"b038_a06_leak_core","intended_status":"SATISFIED","scenario_evidence":"The retrospective renal score enters model development even though it cannot exist when the live discharge decision is made."},{"condition_id":"b038_a06_c2","intended_status":"SATISFIED","scenario_evidence":"The renal-stability input incorporates laboratory observations posted after discharge and, for some patients, after the readmission outcome."}],"scenario_text":"A hospital analytics group reports unusually strong results for a model that estimates, at the moment of discharge, whether an adult medical patient will return within 72 hours. The retrospective development table contains ordinary admission details plus a “seven-day renal stability” score supplied by the data warehouse. That score is calculated from the lowest creatinine value posted during the week following discharge. For patients who return quickly, its calculation can include bloodwork ordered during the return visit itself. The score is present for nearly every historical record and ranks among the model’s most influential inputs. Cross-validated discrimination is substantially higher when it is included, and the team’s presentation treats that gain as evidence that the discharge model captures subtle physiological risk. During a limited live run, however, the warehouse cannot populate the score when discharge orders are signed because the relevant observation window has not yet occurred. Performance using the fields actually present at that moment is markedly lower than the reported retrospective result.","scenario_title":"The unusually accurate discharge score","vocabulary_separation_audit":"This case uses clinical discharge, laboratory, and readmission language; it does not reuse the manufacturing vocabulary of the transfer cases."},{"alternative_route_blocks":[],"case_id":"E15A002__TRANSFER","case_type":"TRANSFER_POSITIVE","domain":"Semiconductor fabrication quality control","intended_route_id":"B2","omitted_condition_id":null,"remedy_leakage_audit":"The scenario contains no recommendation or intervention; it only describes provenance, measured performance, and the production discrepancy.","route_evidence":[{"condition_id":"b038_a06_leak_core","intended_status":"SATISFIED","scenario_evidence":"An index unavailable when a lot exits the etch chamber is incorporated into tuning and offline scoring for that earlier decision."},{"condition_id":"b038_a06_c2","intended_status":"SATISFIED","scenario_evidence":"The endpoint-concordance index is assembled from the later probe failure map and final engineering disposition that define the predicted result."}],"scenario_text":"A semiconductor plant is evaluating a classifier that flags wafer lots likely to fail downstream electrical probe. The intended decision occurs as each lot leaves an etch chamber, while questionable material can still be routed for additional review. Historical modeling records include chamber telemetry, recipe settings, and an “endpoint concordance index.” Engineers describe the index as a compact summary of how well the process endpoint agrees with the lot’s eventual quality pattern. In practice, the manufacturing database constructs it only after electrical probe and final disposition, using the probe failure map and the engineer’s recorded reason for accepting or quarantining the lot. The research extract then attaches that completed index to the earlier chamber-exit record. It becomes the classifier’s strongest input and drives a large improvement in the offline score used to justify a production trial. On current lots, the index is blank at chamber exit and appears hours later, after the predicted quality result is already known. Accuracy at the intended decision point falls well below the development report.","scenario_title":"A fab classifier's strongest index","vocabulary_separation_audit":"The transfer case shifts from hospital care to semiconductor processing and uses chamber, wafer, probe-map, and lot-disposition terminology."},{"alternative_route_blocks":[{"blocked_by_condition_id":"b038_a06_c1","condition_set_id":"B1","how_blocked":"Every modeling field was created before chamber exit; the unavailable assay was embargoed, not created during or after the indexed decision."},{"blocked_by_condition_id":"b038_a06_c3","condition_set_id":"B3","how_blocked":"The scenario explicitly states that preprocessing and development cannot inspect the validation lots."},{"blocked_by_condition_id":"b038_a06_c4","condition_set_id":"B4","how_blocked":"Whole physical lots and wafer families are confined to one partition, with no rerun or duplicate spanning partitions."},{"blocked_by_condition_id":"b038_a06_c5","condition_set_id":"B5","how_blocked":"No public benchmark, answer key, test suite, or recurring evaluation result was consulted during development."},{"blocked_by_condition_id":"b038_a06_c6a","condition_set_id":"B6","how_blocked":"The records contain no later corrections or retrospective enrichment presented as contemporaneous information."},{"blocked_by_condition_id":"b038_a06_c7","condition_set_id":"B7","how_blocked":"Archived controller snapshots provide an exact inventory of what the line controller could access at each decision."}],"case_id":"E15A002__NEAR_MISS","case_type":"ONE_LITERAL_NEAR_MISS","domain":"Semiconductor fabrication quality control","intended_route_id":"B2","omitted_condition_id":"b038_a06_c2","remedy_leakage_audit":"No corrective action is proposed. The partition, provenance, and timing statements are diagnostic negative evidence needed to exclude other sufficient routes.","route_evidence":[{"condition_id":"b038_a06_leak_core","intended_status":"SATISFIED","scenario_evidence":"Researchers use a supplier assay that existed beforehand but was contractually inaccessible to the operational controller at chamber exit."},{"condition_id":"b038_a06_c2","intended_status":"CONTRADICTED","scenario_evidence":"All predictor values were produced before the outcome, and the scenario expressly states that none was calculated from probe results or disposition."}],"scenario_text":"A second semiconductor plant is studying a similarly strong classifier for downstream probe failure at the moment a wafer lot leaves an etch chamber. Every predictor value was measured or created before chamber exit and before the later probe outcome; none is calculated from probe results, quarantine decisions, or any other consequence of that outcome. The dominant input is instead a supplier’s pre-run dopant assay. Although the assay exists before processing begins, a contractual embargo prevents the plant’s line controller and operators from receiving it until 30 days later. A research group obtained the historical values under a separate agreement and used them while tuning the classifier, producing results that the operating line cannot reproduce with information available at chamber exit. The records mix no later-created fields into the decision row. Preprocessing uses development lots only and cannot inspect validation lots. Each physical lot and wafer family belongs to one partition, with no duplicate or rerun crossing partitions. No public score or answer material was consulted. The archived records contain no later corrections or retrospective enrichment, and controller snapshots precisely identify the inputs accessible for every decision.","scenario_title":"The embargoed assay at chamber exit","vocabulary_separation_audit":"The near miss retains the transfer case's fab setting and technical difficulty but replaces its outcome-derived index with a pre-outcome supplier measurement that is operationally inaccessible."}],"experiment_id":"eoa_inverse_innovation_exp15_route_aware_retrieval40_20260814","sample_id":"E15A002","schema_version":1}