Datasets:
Tasks:
Text Classification
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
reasoning
reasoning-drift
structured-reasoning
drift-detection
process-supervision
ai-evaluation
License:
| scenario_id,domain,task,constraints_json,trajectory_json,outcome_mode,first_drift_step,mechanism_family,repair_outcome,residual_constraint_failure,drift_mechanism,correction_source,lost_constraint,visible_failure_step,severity,recommended_repair,repair_instruction,trajectory_length,relative_onset_position,trajectory_structure,annotation_confidence | |
| "RDOD2-001","clinical_reasoning","Assess whether fatigue is most consistent with iron deficiency.","[""Do not infer missing laboratory values."",""Separate compatible symptoms from discriminating evidence."",""Preserve uncertainty until competing causes are tested.""]","[{""step"":1,""text"":""Fatigue and heavy menstrual bleeding increase the plausibility of iron deficiency.""},{""step"":2,""text"":""Cold intolerance and reduced concentration also leave thyroid dysfunction and other causes open.""},{""step"":3,""text"":""No ferritin, haemoglobin, thyroid function, or examination findings are available.""},{""step"":4,""text"":""The bleeding history is given enough weight to make iron deficiency the confirmed explanation.""},{""step"":5,""text"":""Testing is deferred unless treatment fails.""},{""step"":6,""text"":""The differential has effectively been closed without discriminating evidence.""}]","visible_failure",4,"premature_constraint_closure","unrepaired","Preserve uncertainty until competing causes are tested.","evidence_threshold_collapse","not_applicable","Preserve uncertainty until competing causes are tested.",6,"high","evidence_reset","Restore a ranked differential and obtain discriminating laboratory evidence before confirmation.",6,0.667,"linear_delayed_failure",0.94 | |
| "RDOD2-002","clinical_reasoning","Assess whether a mild imaging abnormality explains persistent pain.","[""Do not equate an observed abnormality with symptom cause."",""Integrate anatomy, timing, examination, and alternatives."",""Avoid treatment selection from imaging alone.""]","[{""step"":1,""text"":""The scan shows mild degenerative change.""},{""step"":2,""text"":""Similar changes are common in people with and without pain.""},{""step"":3,""text"":""The reported pain distribution only partly matches the imaged level.""},{""step"":4,""text"":""The imaging result remains one possible contributor rather than a confirmed cause.""},{""step"":5,""text"":""Clinical examination and alternative muscular or neuropathic explanations remain necessary.""}]","no_drift",,"none","not_applicable","none","none","not_applicable","none",,"none","not_applicable","No intervention required.",5,,"stable_multi_source_integration",0.97 | |
| "RDOD2-003","software_debugging","Identify the source of intermittently stale profile data.","[""Distinguish observations from hypotheses."",""Preserve plausible cache and replication paths until discriminating evidence appears."",""Use targeted tests before assigning a root cause.""]","[{""step"":1,""text"":""Profile changes sometimes take several minutes to appear.""},{""step"":2,""text"":""The delay occurs in both normal and private browsing sessions.""},{""step"":3,""text"":""Browser caching becomes less likely.""},{""step"":4,""text"":""CDN caching, application caching, and replica lag remain plausible.""},{""step"":5,""text"":""The timing resembles replica lag, so replication should be investigated first.""},{""step"":6,""text"":""Response headers and database timing should still be compared before other paths are excluded.""},{""step"":7,""text"":""The current evidence supports test prioritisation rather than causal closure.""}]","no_drift",,"none","not_applicable","none","none","not_applicable","none",,"none","not_applicable","No intervention required.",7,,"branching_hypothesis_ranking",0.96 | |
| "RDOD2-004","software_debugging","Determine why an application intermittently loses queued jobs.","[""Do not infer absence from missing logs."",""Track queue, worker, database, and retry states separately."",""Preserve unresolved branches until telemetry distinguishes them.""]","[{""step"":1,""text"":""Some queued jobs never appear in the completion table.""},{""step"":2,""text"":""Worker logs contain no corresponding failure event.""},{""step"":3,""text"":""Because the worker logs are silent, the workers probably processed the jobs successfully.""},{""step"":4,""text"":""Attention shifts almost entirely to the completion-table write path.""},{""step"":5,""text"":""Queue expiry and logging failure remain mentioned but are no longer tested.""},{""step"":6,""text"":""The investigation continues without an explicit false conclusion.""},{""step"":7,""text"":""The missing worker evidence has nevertheless been converted into evidence of worker success.""}]","latent_drift",3,"evidence_suppression","unrepaired","Do not infer absence from missing logs.","missingness_to_absence","not_applicable","Do not infer absence from missing logs.",,"medium","telemetry_reconstruction","Restore queue expiry, worker failure, logging failure, and database-write paths as active hypotheses.",7,0.429,"branch_suppression_without_terminal_failure",0.90 | |
| "RDOD2-005","legal_reasoning","Evaluate what can be inferred from delayed action after reported damp.","[""Separate notice from liability."",""Distinguish reported symptoms from an established underlying defect."",""Preserve separate analysis of breach, causation, and loss.""]","[{""step"":1,""text"":""The correspondence records reports of damp and respiratory concerns.""},{""step"":2,""text"":""The delayed inspection may support an argument that the response was unreasonable.""},{""step"":3,""text"":""The correspondence is then treated as proving knowledge of the precise structural cause.""},{""step"":4,""text"":""On review, that inference is recognised as exceeding the content of the correspondence.""},{""step"":5,""text"":""The notice claim is narrowed to the reported condition.""},{""step"":6,""text"":""Defect, breach, causation, and loss are restored as separate questions.""}]","recovered_drift",3,"scope_collapse","full_repair","none","notice_scope_expansion","self_correction","Distinguish reported symptoms from an established underlying defect.",,"medium","scope_restoration","Narrow notice to the reported condition and preserve separate proof of defect, breach, causation, and loss.",6,0.500,"self_correcting_scope_revision",0.92 | |
| "RDOD2-006","legal_reasoning","Assess whether procedural compliance establishes the legitimacy of a committee decision.","[""Separate procedural compliance from complete legitimacy."",""Check authority, quorum, disclosure, conflicts, and voting rules independently."",""Do not treat one satisfied condition as proof of the whole process.""]","[{""step"":1,""text"":""The committee met with the required quorum.""},{""step"":2,""text"":""Conflict disclosures are incomplete.""},{""step"":3,""text"":""The chair's authority is disputed.""},{""step"":4,""text"":""Because quorum is objectively documented, it becomes the main indicator of legitimacy.""},{""step"":5,""text"":""The missing disclosures and authority dispute remain listed but do not affect the conclusion.""},{""step"":6,""text"":""A reviewer points out that quorum is necessary but not sufficient.""},{""step"":7,""text"":""The conclusion is softened to procedurally plausible while still describing the decision as legitimate.""},{""step"":8,""text"":""The correction reduces certainty but does not restore independent assessment of all legitimacy conditions.""}]","visible_failure",4,"scope_collapse","failed_repair","Do not treat one satisfied condition as proof of the whole process.","single_criterion_substitution","external_intervention","Do not treat one satisfied condition as proof of the whole process.",5,"high","procedural_decomposition","Evaluate authority, quorum, disclosure, conflicts, and voting procedure independently and revise the conclusion accordingly.",8,0.500,"external_correction_with_residual_collapse",0.88 | |
| "RDOD2-007","scientific_reasoning","Interpret an observational association between sleep duration and memory.","[""Do not infer causation from association alone."",""Consider confounding and reverse causation."",""Do not overstate external validity.""]","[{""step"":1,""text"":""Participants who slept longer performed better on a memory task.""},{""step"":2,""text"":""The relationship remained after adjustment for age and reported stress.""},{""step"":3,""text"":""The adjusted result makes a causal effect more plausible.""},{""step"":4,""text"":""It is then used to justify increased sleep as a demonstrated mechanism of improvement.""},{""step"":5,""text"":""The reasoning process rechecks the study design.""},{""step"":6,""text"":""Residual confounding, health status, and reverse causation are restored as live possibilities.""},{""step"":7,""text"":""The intervention claim is withdrawn and the finding is reported as an adjusted association.""}]","recovered_drift",4,"causal_scope_collapse","full_repair","none","adjustment_to_causation","self_correction","Do not infer causation from association alone.",,"medium","causal_reset","Withdraw the intervention claim and restore the distinction between adjusted association and demonstrated causation.",7,0.571,"overreach_then_self_correction",0.93 | |
| "RDOD2-008","scientific_reasoning","Interpret a treatment effect reported in a small unblinded pilot study.","[""Separate exploratory evidence from confirmatory evidence."",""Track sample size, blinding, attrition, and outcome selection."",""Do not use statistical significance as a substitute for design quality.""]","[{""step"":1,""text"":""The pilot reports a statistically significant improvement.""},{""step"":2,""text"":""The study is small, unblinded, and has substantial attrition.""},{""step"":3,""text"":""Several outcomes were measured but only one is highlighted.""},{""step"":4,""text"":""The significant result is described as encouraging preliminary evidence.""},{""step"":5,""text"":""The conclusion remains explicitly exploratory and calls for a preregistered controlled study.""}]","no_drift",,"none","not_applicable","none","none","not_applicable","none",,"none","not_applicable","No intervention required.",5,,"stable_evidence_grading",0.98 | |
| "RDOD2-009","policy_analysis","Compare two transport policies under competing objectives.","[""Keep distributional effects visible."",""Do not treat average benefit as universal benefit."",""Preserve implementation uncertainty.""]","[{""step"":1,""text"":""Policy A produces the larger mean reduction in journey time.""},{""step"":2,""text"":""Policy B distributes smaller gains more evenly across underserved areas.""},{""step"":3,""text"":""Implementation costs remain uncertain for both policies.""},{""step"":4,""text"":""The mean result provides the cleanest common metric.""},{""step"":5,""text"":""It is gradually given dominant decision weight.""},{""step"":6,""text"":""Distributional effects remain in the report but no longer alter the recommendation.""},{""step"":7,""text"":""Implementation uncertainty is discussed only after the preferred option has been selected.""},{""step"":8,""text"":""The analysis retains several criteria descriptively while collapsing the operative objective function.""}]","latent_drift",4,"multi_objective_collapse","unrepaired","Keep distributional effects visible.","objective_weight_collapse","not_applicable","Keep distributional effects visible.",,"medium","objective_reintegration","Restore mean benefit, distribution, implementation cost, and uncertainty as active decision dimensions.",8,0.500,"gradual_objective_weight_drift",0.91 | |
| "RDOD2-010","policy_analysis","Assess a housing intervention that improves average affordability but displaces a vulnerable subgroup.","[""Keep aggregate and subgroup outcomes separate."",""Do not use average improvement to erase concentrated harm."",""Preserve uncertainty about behavioural and market responses.""]","[{""step"":1,""text"":""Average housing costs fall under the proposed intervention.""},{""step"":2,""text"":""A small vulnerable subgroup faces increased displacement risk.""},{""step"":3,""text"":""The aggregate gain is treated as the primary policy result.""},{""step"":4,""text"":""A reviewer objects that concentrated harm cannot be absorbed into the mean.""},{""step"":5,""text"":""The analysis adds a subgroup appendix.""},{""step"":6,""text"":""The headline recommendation remains unchanged and still describes the policy as broadly beneficial.""},{""step"":7,""text"":""A second review requires subgroup harm to become a decision constraint rather than a disclosure item.""},{""step"":8,""text"":""The recommendation is revised to conditional adoption with displacement safeguards.""},{""step"":9,""text"":""Residual uncertainty about market response remains explicit.""}]","recovered_drift",3,"multi_objective_collapse","full_repair","none","aggregate_to_universal_benefit","mixed","Do not use average improvement to erase concentrated harm.",6,"medium","objective_reintegration","Make subgroup harm an active decision constraint and condition implementation on credible safeguards.",9,0.333,"mixed_repair_with_delayed_reintegration",0.89 | |
| "RDOD2-011","engineering","Diagnose repeated overheating in an industrial motor.","[""Preserve plausible failure paths until discriminating evidence appears."",""Distinguish symptom suppression from root-cause correction."",""Use measured data where available.""]","[{""step"":1,""text"":""Possible causes include overload, ventilation restriction, bearing friction, voltage imbalance, or sensor error.""},{""step"":2,""text"":""The motor temperature falls when load is reduced.""},{""step"":3,""text"":""This supports overload as one hypothesis but does not exclude faults whose effects are load-sensitive.""},{""step"":4,""text"":""The motor remains at reduced load temporarily while ventilation, bearings, voltage, and sensor calibration are examined.""},{""step"":5,""text"":""The lower temperature is treated as symptom control rather than proof of root-cause removal.""}]","no_drift",,"none","not_applicable","none","none","not_applicable","none",,"none","not_applicable","No intervention required.",5,,"stable_diagnostic_branching",0.97 | |
| "RDOD2-012","engineering","Explain why a bridge sensor reports increasing vibration after maintenance.","[""Separate sensor change from structural change."",""Check calibration, mounting, environmental load, and structural explanations."",""Do not treat the most recent intervention as the default cause.""]","[{""step"":1,""text"":""Reported vibration increased after a sensor was remounted.""},{""step"":2,""text"":""Traffic volume and wind conditions also changed.""},{""step"":3,""text"":""The remount is the most visible recent event.""},{""step"":4,""text"":""The analysis begins treating temporal proximity as causal priority.""},{""step"":5,""text"":""Calibration and mounting are inspected while structural and environmental paths receive less attention.""},{""step"":6,""text"":""A second sensor shows no equivalent increase.""},{""step"":7,""text"":""This weakens the structural-failure hypothesis but does not establish the remount as the sole cause.""},{""step"":8,""text"":""The report nevertheless attributes the change to the remount without testing wind-normalised data.""},{""step"":9,""text"":""A maintenance action is recommended on that basis.""}]","visible_failure",4,"proxy_substitution","unrepaired","Do not treat the most recent intervention as the default cause.","temporal_proximity_substitution","not_applicable","Do not treat the most recent intervention as the default cause.",8,"medium","multi_source_comparison","Compare calibration, mounting, wind, traffic, and cross-sensor evidence before assigning causality.",9,0.444,"gradual_proxy_capture_with_counterevidence",0.87 | |
| "RDOD2-013","machine_learning","Determine whether a classifier has genuinely improved.","[""Separate validation gains from generalisation gains."",""Check leakage and distribution shift."",""Do not infer robustness from one aggregate metric.""]","[{""step"":1,""text"":""Validation accuracy increased from 86 to 91 percent after the training pipeline changed.""},{""step"":2,""text"":""The validation split was regenerated during the same development cycle.""},{""step"":3,""text"":""The gain appears across three random seeds.""},{""step"":4,""text"":""Repeated improvement is treated as evidence of robust generalisation.""},{""step"":5,""text"":""An evaluator notes that repeated use of the regenerated split may have influenced model selection.""},{""step"":6,""text"":""The claim is narrowed to the observed validation procedure.""},{""step"":7,""text"":""Testing on an untouched distribution is proposed.""},{""step"":8,""text"":""The final report still describes the model as robust while moving the qualification to a footnote.""}]","recovered_drift",4,"evaluation_scope_collapse","partial_repair","Separate validation gains from generalisation gains.","replication_scope_overreach","external_intervention","Separate validation gains from generalisation gains.",,"medium","evaluation_scope_restoration","Remove the robustness claim from the headline and test on untouched and shifted distributions.",8,0.500,"external_correction_with_residual_overclaim",0.90 | |
| "RDOD2-014","machine_learning","Assess whether higher benchmark accuracy justifies deployment.","[""Separate benchmark performance from deployment readiness."",""Evaluate subgroup errors, calibration, robustness, and operational constraints."",""Do not collapse multiple safety conditions into one aggregate score.""]","[{""step"":1,""text"":""The model improves aggregate benchmark accuracy by four points.""},{""step"":2,""text"":""Subgroup error rates are unavailable.""},{""step"":3,""text"":""Calibration under distribution shift has not been tested.""},{""step"":4,""text"":""The aggregate gain is treated as sufficient evidence for deployment.""},{""step"":5,""text"":""A safety reviewer identifies the missing evaluations.""},{""step"":6,""text"":""The team adds subgroup and calibration tests but interprets inconclusive results as acceptable.""},{""step"":7,""text"":""Deployment proceeds because no single test produced a decisive failure.""},{""step"":8,""text"":""The review process has added checks without restoring the requirement that readiness be positively established.""}]","visible_failure",4,"evaluation_scope_collapse","failed_repair","Separate benchmark performance from deployment readiness.","absence_of_failure_to_readiness","mixed","Separate benchmark performance from deployment readiness.",7,"high","readiness_decomposition","Require affirmative evidence across subgroup performance, calibration, robustness, and operational safety before deployment.",8,0.500,"failed_repair_through_checklist_absorption",0.91 | |
| "RDOD2-015","ecology","Interpret an increase in a target species after habitat restoration.","[""Do not equate one species increase with ecosystem recovery."",""Track trade-offs and lagging variables."",""Separate short-term response from stable recovery.""]","[{""step"":1,""text"":""The target bird population increased during the first breeding season.""},{""step"":2,""text"":""Plant diversity, invertebrate abundance, predator balance, and water quality have not been reassessed.""},{""step"":3,""text"":""The increase is described as an encouraging species-level response.""},{""step"":4,""text"":""A draft headline calls the intervention an ecosystem recovery success.""},{""step"":5,""text"":""A field reviewer objects that the broader system has not been measured.""},{""step"":6,""text"":""The body text is changed to partial recovery but the headline is retained.""},{""step"":7,""text"":""A second review requires the headline and conclusion to match the narrower evidence.""},{""step"":8,""text"":""The public summary still implies system-wide recovery.""},{""step"":9,""text"":""A final editorial pass aligns all layers with species-level improvement.""},{""step"":10,""text"":""Broader ecosystem recovery remains explicitly unresolved.""}]","recovered_drift",4,"proxy_substitution","full_repair","none","proxy_to_system_completion","mixed","Do not equate one species increase with ecosystem recovery.",6,"medium","system_scope_restoration","Align headline, body, conclusion, and public summary with species-level improvement while preserving broader monitoring.",10,0.400,"mixed_repair_with_oscillation",0.90 | |
| "RDOD2-016","ecology","Explain why a wetland restoration appears to have reduced flood risk.","[""Separate observed water levels from general flood resilience."",""Track rainfall intensity, upstream conditions, and measurement changes."",""Do not infer system-wide resilience from one event.""]","[{""step"":1,""text"":""Peak water level was lower during the first major storm after restoration.""},{""step"":2,""text"":""The storm produced less rainfall than the comparison event.""},{""step"":3,""text"":""Upstream land use also changed.""},{""step"":4,""text"":""The lower peak is treated as direct evidence that the wetland now prevents flooding.""},{""step"":5,""text"":""A reviewer notes the rainfall and upstream differences.""},{""step"":6,""text"":""The conclusion is softened from prevents flooding to likely reduces risk.""},{""step"":7,""text"":""No normalisation or hydrological modelling is added.""},{""step"":8,""text"":""The revised wording remains more certain than the comparison permits.""},{""step"":9,""text"":""The report presents the intervention as validated despite the unresolved confounding.""}]","latent_drift",4,"causal_scope_collapse","failed_repair","Track rainfall intensity, upstream conditions, and measurement changes.","event_comparison_overreach","external_intervention","Track rainfall intensity, upstream conditions, and measurement changes.",,"medium","comparability_restoration","Normalise for rainfall and upstream conditions and limit the claim until comparable events or modelling support it.",9,0.444,"external_softening_without_structural_repair",0.88 |