Advanced Analytics Slices

Tests of the Advanced Analytics slice pipeline, which prepares, validates and transforms datasets ahead of statistical and machine-learning work.

General

AdaBoostClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.AdaBoostClassifierExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInDecisionFunctionSpaceWithTheWeightedVoteMethod	Fact	To adaptation contribution is inexact and in decision function space with the weighted vote method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesProbabilitySpace	Fact	To adaptation describes calibration only when calibrated and uses probability space.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

AdaBoostClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.AdaBoostClassifierRecallTests — 12 tests.

Test	Kind	Verifies
SeparableDataProducesHighProbabilityForPositivesAndLowForNegatives	Fact	Separable data produces high probability for positives and low for negatives.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
PredictedProbabilitiesAreAlwaysBetweenZeroAndOne	Fact	Predicted probabilities are always between zero and one.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
RequiringBothClassesThrowsAClearExceptionWhenOnlyOneIsPresent	Fact	Requiring both classes throws a clear exception when only one is present.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleEnsembleSummariesAndFiniteNonNegativeAlphas	Fact	The evidence reports plausible ensemble summaries and finite non negative alphas.
ASingleRoundFitsExactlyOneTreeWithAFiniteNonNegativeAlpha	Fact	A single round fits exactly one tree with a finite non negative alpha.
ExplainRawScoreMatchesThePipelineRawScore	Fact	Explain raw score matches the pipeline raw score.

AdvancedAnalyticsRecallExtensionsTests

Jube.Test.AdvancedAnalytics.Slice.AdvancedAnalyticsRecallExtensionsTests — 5 tests. Category: Unit.

Test	Kind	Verifies
BuildScoreInputs_ReadsEachVariableByNameInOrder	Fact	Build score inputs reads each variable by name in order.
BuildScoreInputs_UsesNanForAMissingVariable	Fact	Build score inputs uses nan for a missing variable.
BuildScoreInputs_UsesNanForAnUnparseableOrNullValue	Fact	Build score inputs uses nan for an unparseable or null value.
BuildScoreInputs_ParsesBooleansAndIntegerTokensLikeExhaustiveDoes	Fact	Build score inputs parses booleans and integer tokens like exhaustive does.
RecallBodyHasValidTypesForAdvancedAnalytics_Boundaries	Theory	Recall body has valid types for advanced analytics boundaries.

BaggingClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.BaggingClassifierExplainToAdaptationTests — 7 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInVoteFractionSpaceWithTheTreePathContributionMethod	Fact	To adaptation contribution is inexact and in vote fraction space with the tree path contribution method.
ToAdaptation_TheDiscountedContributionsDoNotSumBackToRawScoreMinusBaseValue	Fact	To adaptation the discounted contributions do not sum back to raw score minus base value.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesProbabilitySpace	Fact	To adaptation describes calibration only when calibrated and uses probability space.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

BaggingClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.BaggingClassifierRecallTests — 12 tests.

Test	Kind	Verifies
SeparableDataProducesHighProbabilityForPositivesAndLowForNegatives	Fact	Separable data produces high probability for positives and low for negatives.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
PredictedProbabilitiesAreAlwaysBetweenZeroAndOne	Fact	Predicted probabilities are always between zero and one.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
RequiringBothClassesThrowsAClearExceptionWhenOnlyOneIsPresent	Fact	Requiring both classes throws a clear exception when only one is present.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleEnsembleSummaries	Fact	The evidence reports plausible ensemble summaries.
CalibrateFalseLeavesTheModelUncalibratedAndCalibrateTrueCalibratesIt	Fact	Calibrate false leaves the model uncalibrated and calibrate true calibrates it.
ExplainReportsAContributionPerVariableAndAPlausibleBaseValue	Fact	Explain reports a contribution per variable and a plausible base value.

BalancedKMeansClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.BalancedKMeansClustererRecallTests — 12 tests.

Test	Kind	Verifies
NearestCentroidDistanceIsGenuinelyTheMinimumAndIsMovedToTheFront	Fact	Nearest centroid distance is genuinely the minimum and is moved to the front.
PredictReturnsOneDistancePerCentroidInClusterIndexOrderExceptTheNearestIsFirst	Fact	Predict returns one distance per centroid in cluster index order except the nearest is first.
SameSeedProducesTheSameCentroidsAndTheSameRecallBitForBit	Fact	Same seed produces the same centroids and the same recall bit for bit.
RequestingMoreClustersThanDistinctRowsThrowsAClearException	Fact	Requesting more clusters than distinct rows throws a clear exception.
RequestingMoreClustersThanRowsThrowsAClearException	Fact	Requesting more clusters than rows throws a clear exception.
ExceedingTheMaximumTrainingRowsThrowsAClearException	Fact	Exceeding the maximum training rows throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleClusterSummaries	Fact	The evidence reports plausible cluster summaries.
ClusterSizesAreAsEqualAsPossibleEvenWhenTheNaturalGroupsAreNot	Fact	Cluster sizes are as equal as possible even when the natural groups are not.

BinarySplitClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.BinarySplitClustererRecallTests — 12 tests.

Test	Kind	Verifies
NearestCentroidDistanceIsGenuinelyTheMinimumAndIsMovedToTheFront	Fact	Nearest centroid distance is genuinely the minimum and is moved to the front.
PredictReturnsOneDistancePerCentroidInClusterIndexOrderExceptTheNearestIsFirst	Fact	Predict returns one distance per centroid in cluster index order except the nearest is first.
SameSeedProducesTheSameClustersAndNearlyTheSameRecall	Fact	Same seed produces the same clusters and nearly the same recall.
ADifferentSeedCanProduceDifferentCentroids	Fact	A different seed can produce different centroids.
RequestingASingleClusterIsRejectedBecauseItIsNeverBisected	Fact	Requesting a single cluster is rejected because it is never bisected.
RequestingMoreClustersThanDistinctRowsThrowsAClearException	Fact	Requesting more clusters than distinct rows throws a clear exception.
RequestingMoreClustersThanRowsThrowsAClearException	Fact	Requesting more clusters than rows throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleClusterSummaries	Fact	The evidence reports plausible cluster summaries.

DecisionTreeClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.DecisionTreeClassifierExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInProbabilitySpaceWithTheDecisionPathMethod	Fact	To adaptation contribution is inexact and in probability space with the decision path method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndAlwaysUsesProbabilitySpace	Fact	To adaptation describes calibration only when calibrated and always uses probability space.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

DecisionTreeClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.DecisionTreeClassifierRecallTests — 13 tests.

Test	Kind	Verifies
PredictReturnsAHighProbabilityForAClearlyPositiveRowAndLowForAClearlyNegativeRow	Fact	Predict returns a high probability for a clearly positive row and low for a clearly negative row.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedianAndStillProducesAFiniteProbability	Fact	A missing input is imputed with the training median and still produces a finite probability.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleMetricsAndImportancesThatSumToOne	Fact	The evidence reports plausible metrics and importances that sum to one.
CalibrateFalseLeavesTheModelUncalibratedAndCalibrateTrueCalibratesIt	Fact	Calibrate false leaves the model uncalibrated and calibrate true calibrates it.
MaxDepthOneProducesAtMostOneSplit	Fact	Max depth one produces at most one split.
AConstantColumnIsExcludedFromTheFitAndWarnedAbout	Fact	A constant column is excluded from the fit and warned about.
BothClassesMustBePresentInTheTrainingSplit	Fact	Both classes must be present in the training split.
ExplainsRootProbabilityMatchesTheEvidenceAndContributionsMoveTowardTheLeaf	Fact	Explains root probability matches the evidence and contributions move toward the leaf.

DecisionTreeRegressorExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.DecisionTreeRegressorExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInScoreSpaceWithTheDecisionPathMethod	Fact	To adaptation contribution is inexact and in score space with the decision path method.
ToAdaptation_IsNeverCalibrated	Fact	To adaptation is never calibrated.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

DecisionTreeRegressorRecallTests

Jube.Test.AdvancedAnalytics.Slice.DecisionTreeRegressorRecallTests — 11 tests.

Test	Kind	Verifies
PredictReturnsAHighValueForAClearlyPositiveRowAndLowForAClearlyNegativeRow	Fact	Predict returns a high value for a clearly positive row and low for a clearly negative row.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedianAndStillProducesAFiniteValue	Fact	A missing input is imputed with the training median and still produces a finite value.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleMetricsAndImportancesThatSumToOne	Fact	The evidence reports plausible metrics and importances that sum to one.
MaxDepthOneProducesAtMostOneSplit	Fact	Max depth one produces at most one split.
AConstantColumnIsExcludedFromTheFitAndWarnedAbout	Fact	A constant column is excluded from the fit and warned about.
ExplainsRootValueMatchesTheEvidenceAndContributionsMoveTowardTheLeaf	Fact	Explains root value matches the evidence and contributions move toward the leaf.

GaussianMixtureClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.GaussianMixtureClustererRecallTests — 11 tests.

Test	Kind	Verifies
HighestPosteriorIsGenuinelyTheMaximumAndIsMovedToTheFront	Fact	Highest posterior is genuinely the maximum and is moved to the front.
PredictReturnsOnePosteriorPerComponentInComponentIndexOrderExceptTheHighestIsFirst	Fact	Predict returns one posterior per component in component index order except the highest is first.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RequestingMoreComponentsThanDistinctRowsThrowsAClearException	Fact	Requesting more components than distinct rows throws a clear exception.
RequestingMoreComponentsThanRowsThrowsAClearException	Fact	Requesting more components than rows throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleClusterSummaries	Fact	The evidence reports plausible cluster summaries.

GaussianMixtureDensityExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.GaussianMixtureDensityExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInTheLogLikelihoodSpaceWithTheMahalanobisTermMethod	Fact	To adaptation contribution is inexact and in the log likelihood space with the mahalanobis term method.
ToAdaptation_IsNeverCalibratedAndUsesLogLikelihoodSpace	Fact	To adaptation is never calibrated and uses log likelihood space.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

GaussianMixtureDensityRecallTests

Jube.Test.AdvancedAnalytics.Slice.GaussianMixtureDensityRecallTests — 9 tests.

Test	Kind	Verifies
ARowNearAFittedComponentScoresAHigherLogDensityThanARowFarFromEveryComponent	Fact	A row near a fitted component scores a higher log density than a row far from every component.
ExplainDecomposesPerDimensionAndDoesNotSumExactlyToTheLogDensity	Fact	Explain decomposes per dimension and does not sum exactly to the log density.
AConstantColumnIsExcludedFromTheMixtureAndCarriesZeroWeight	Fact	A constant column is excluded from the mixture and carries zero weight.
EveryVariableConstantThrowsALoudException	Fact	Every variable constant throws a loud exception.
TooFewRowsForTheRequestedComponentCountThrowsAClearException	Fact	Too few rows for the requested component count throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
RecallIsThreadSafeAndDeterministicUnderContention	Fact	Recall is thread safe and deterministic under contention.

GradientBoostingClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.GradientBoostingClassifierExplainToAdaptationTests — 7 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsExactAndInLogOddsSpaceWithTheTreePathContributionMethod	Fact	To adaptation contribution is exact and in log odds space with the tree path contribution method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesProbabilitySpaceEitherWay	Fact	To adaptation describes calibration only when calibrated and uses probability space either way.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.
ToAdaptation_ProvesExactnessAcrossManyRealFitsThroughTheRecallAndAdaptationRoundTrip	Fact	To adaptation proves exactness across many real fits through the recall and adaptation round trip.

GradientBoostingClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.GradientBoostingClassifierRecallTests — 14 tests.

Test	Kind	Verifies
SeparableDataProducesHighProbabilityForPositivesAndLowForNegatives	Fact	Separable data produces high probability for positives and low for negatives.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
PredictedProbabilitiesAreAlwaysBetweenZeroAndOne	Fact	Predicted probabilities are always between zero and one.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
RequiringBothClassesThrowsAClearExceptionWhenOnlyOneIsPresent	Fact	Requiring both classes throws a clear exception when only one is present.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleEnsembleSummaries	Fact	The evidence reports plausible ensemble summaries.
ExplainRawScoreMatchesThePipelineRawScore	Fact	Explain raw score matches the pipeline raw score.
ContributionsReconstructTheRawLogOddsMarginExactlyAcrossManyProbes	Fact	Contributions reconstruct the raw log odds margin exactly across many probes.
ALargerLearningRateReachesAGivenTrainingLossInFewerRounds	Fact	A larger learning rate reaches a given training loss in fewer rounds.
ASingleRoundStillProducesAValidCalibratedModel	Fact	A single round still produces a valid calibrated model.

GradientBoostingRegressorExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.GradientBoostingRegressorExplainToAdaptationTests — 7 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsExactAndInScoreSpaceWithTheTreePathContributionMethod	Fact	To adaptation contribution is exact and in score space with the tree path contribution method.
ToAdaptation_IsNeverCalibrated	Fact	To adaptation is never calibrated.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.
ToAdaptation_ProvesExactnessAcrossManyRealFitsThroughTheRecallAndAdaptationRoundTrip	Fact	To adaptation proves exactness across many real fits through the recall and adaptation round trip.

GradientBoostingRegressorRecallTests

Jube.Test.AdvancedAnalytics.Slice.GradientBoostingRegressorRecallTests — 15 tests.

Test	Kind	Verifies
PredictionsTrackTheOutcomeReasonablyWellOnASimpleRelationship	Fact	Predictions track the outcome reasonably well on a simple relationship.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
PredictionsAreAlwaysFinite	Fact	Predictions are always finite.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleEnsembleSummaries	Fact	The evidence reports plausible ensemble summaries.
ExplainPredictedValueMatchesThePipelinePrediction	Fact	Explain predicted value matches the pipeline prediction.
ContributionsReconstructThePredictedValueExactlyAcrossManyProbesAndAllTrainingRows	Fact	Contributions reconstruct the predicted value exactly across many probes and all training rows.
ALargerLearningRateReachesAGivenTrainingLossInFewerRounds	Fact	A larger learning rate reaches a given training loss in fewer rounds.
ASingleRoundStillProducesAValidModel	Fact	A single round still produces a valid model.
FewerRowsThanVariablesStillTrainsSinceTreesHaveNoDesignMatrixRankRequirement	Fact	Fewer rows than variables still trains since trees have no design matrix rank requirement.
TestSplitRowsNeverTouchTheFittedTrainingStatistics	Fact	Test split rows never touch the fitted training statistics.

IsolationForestAnomalyExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.IsolationForestAnomalyExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInTheScoreSpaceWithTheTreePathContributionMethod	Fact	To adaptation contribution is inexact and in the score space with the tree path contribution method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesScoreSpaceOtherwise	Fact	To adaptation describes calibration only when calibrated and uses score space otherwise.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

IsolationForestAnomalyRecallTests

Jube.Test.AdvancedAnalytics.Slice.IsolationForestAnomalyRecallTests — 10 tests.

Test	Kind	Verifies
OutlierRowsScoreHigherThanNormalRows	Fact	Outlier rows score higher than normal rows.
RecallReproducesTrainingTimeScoresBitForBit	Fact	Recall reproduces training time scores bit for bit.
TrainingIsDeterministicForTheSameSeed	Fact	Training is deterministic for the same seed.
MissingValuesAreImputedWithTheTrainingMedian	Fact	Missing values are imputed with the training median.
UnlabelledTrainingHasNoCalibratorAndReportsARawScoreInUnitRange	Fact	Unlabelled training has no calibrator and reports a raw score in unit range.
LabelledTrainingWithBothClassesInValidationFitsACalibratorAndProducesARealRocAuc	Fact	Labelled training with both classes in validation fits a calibrator and produces a real roc auc.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
RecallIsThreadSafeAndDeterministicUnderContention	Fact	Recall is thread safe and deterministic under contention.

KMeansClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.KMeansClustererRecallTests — 11 tests.

Test	Kind	Verifies
NearestCentroidDistanceIsGenuinelyTheMinimumAndIsMovedToTheFront	Fact	Nearest centroid distance is genuinely the minimum and is moved to the front.
PredictReturnsOneDistancePerCentroidInClusterIndexOrderExceptTheNearestIsFirst	Fact	Predict returns one distance per centroid in cluster index order except the nearest is first.
SameSeedProducesTheSameCentroidsAndTheSameRecallBitForBit	Fact	Same seed produces the same centroids and the same recall bit for bit.
ADifferentSeedCanProduceDifferentCentroids	Fact	A different seed can produce different centroids.
RequestingMoreClustersThanDistinctRowsThrowsAClearException	Fact	Requesting more clusters than distinct rows throws a clear exception.
RequestingMoreClustersThanRowsThrowsAClearException	Fact	Requesting more clusters than rows throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleClusterSummaries	Fact	The evidence reports plausible cluster summaries.

KMedoidsClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.KMedoidsClustererRecallTests — 12 tests.

Test	Kind	Verifies
NearestMedoidDistanceIsGenuinelyTheMinimumAndIsMovedToTheFront	Fact	Nearest medoid distance is genuinely the minimum and is moved to the front.
PredictReturnsOneDistancePerMedoidInClusterIndexOrderExceptTheNearestIsFirst	Fact	Predict returns one distance per medoid in cluster index order except the nearest is first.
SameSeedProducesTheSameMedoidsAndTheSameRecallBitForBit	Fact	Same seed produces the same medoids and the same recall bit for bit.
ADifferentSeedCanProduceDifferentMedoidsUnderUniformSeeding	Fact	A different seed can produce different medoids under uniform seeding.
RequestingMoreClustersThanDistinctRowsThrowsAClearException	Fact	Requesting more clusters than distinct rows throws a clear exception.
RequestingMoreClustersThanRowsThrowsAClearException	Fact	Requesting more clusters than rows throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
EveryMedoidIsALiteralTrainingRowNotAnAverage	Fact	Every medoid is a literal training row not an average.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleClusterSummaries	Fact	The evidence reports plausible cluster summaries.

KModesClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.KModesClustererRecallTests — 13 tests.

Test	Kind	Verifies
NearestModeDistanceIsGenuinelyTheMinimumAndIsMovedToTheFront	Fact	Nearest mode distance is genuinely the minimum and is moved to the front.
PredictReturnsOneMismatchCountPerModeInClusterIndexOrderExceptTheNearestIsFirst	Fact	Predict returns one mismatch count per mode in cluster index order except the nearest is first.
SameSeedProducesTheSameModesAndTheSameRecallBitForBit	Fact	Same seed produces the same modes and the same recall bit for bit.
ADifferentSeedCanProduceDifferentModes	Fact	A different seed can produce different modes.
RequestingMoreClustersThanDistinctRowsThrowsAClearException	Fact	Requesting more clusters than distinct rows throws a clear exception.
RequestingMoreClustersThanRowsThrowsAClearException	Fact	Requesting more clusters than rows throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMode	Fact	A missing input is imputed with the training mode.
ANonIntegerInputIsRoundedToTheNearestCategory	Fact	A non integer input is rounded to the nearest category.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleClusterSummaries	Fact	The evidence reports plausible cluster summaries.
ANonIntegerTrainingColumnRaisesAWarning	Fact	A non integer training column raises a warning.

KNearestNeighboursClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.KNearestNeighboursClassifierRecallTests — 11 tests.

Test	Kind	Verifies
APointNearAKnownClassClusterPredictsThatClassWithHighProbability	Fact	A point near a known class cluster predicts that class with high probability.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RequestingMoreNeighboursThanAvailableTrainingRowsThrowsAClearException	Fact	Requesting more neighbours than available training rows throws a clear exception.
RequestingMoreNeighboursThanStoredRowsAtRecallTimeThrowsAClearException	Fact	Requesting more neighbours than stored rows at recall time throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsCalibratedTestMetricsOnWellSeparatedClasses	Fact	The evidence reports calibrated test metrics on well separated classes.
TheExplainSharesAreNonNegativeAndSumToOne	Fact	The explain shares are non negative and sum to one.

KNearestNeighboursRegressorExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.KNearestNeighboursRegressorExplainToAdaptationTests — 8 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndUsesTheTrainingMeanAsBaseValue	Fact	To adaptation contribution is inexact and uses the training mean as base value.
ToAdaptation_IsNeverCalibrated	Fact	To adaptation is never calibrated.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.
ToAdaptation_NeverCarriesPerContributionSignificance	Fact	To adaptation never carries per contribution significance.
ToAdaptation_ContributionWeightsSumToOne	Fact	To adaptation contribution weights sum to one.

KNearestNeighboursRegressorRecallTests

Jube.Test.AdvancedAnalytics.Slice.KNearestNeighboursRegressorRecallTests — 12 tests.

Test	Kind	Verifies
APointNearAKnownTargetClusterPredictsCloseToThatCluster	Fact	A point near a known target cluster predicts close to that cluster.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RequestingMoreNeighboursThanAvailableTrainingRowsThrowsAClearException	Fact	Requesting more neighbours than available training rows throws a clear exception.
RequestingMoreNeighboursThanStoredRowsAtRecallTimeThrowsAClearException	Fact	Requesting more neighbours than stored rows at recall time throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleMetricsOnWellSeparatedTargets	Fact	The evidence reports plausible metrics on well separated targets.
TheExplainSharesAreNonNegativeAndSumToOne	Fact	The explain shares are non negative and sum to one.
ThePredictedValueIsThePlainMeanOfTheNeighbourTargets	Fact	The predicted value is the plain mean of the neighbour targets.

KernelRidgeRegressorRecallTests

Jube.Test.AdvancedAnalytics.Slice.KernelRidgeRegressorRecallTests — 11 tests.

Test	Kind	Verifies
AGaussianKernelFitsASinusoidThatALinearKernelCannot	Fact	A gaussian kernel fits a sinusoid that a linear kernel cannot.
APolynomialKernelFitsAQuadraticSurface	Fact	A polynomial kernel fits a quadratic surface.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleMetricsAndExpansionSize	Fact	The evidence reports plausible metrics and expansion size.
GammaResolvesAutomaticallyFromTheNumberOfPredictorsWhenLeftAtZero	Fact	Gamma resolves automatically from the number of predictors when left at zero.
ANearSingularSystemIsHandledEitherWithAWarningOrARefusalRatherThanSilentGarbage	Fact	A near singular system is handled either with a warning or a refusal rather than silent garbage.

KernelSupportVectorClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.KernelSupportVectorClassifierRecallTests — 12 tests.

Test	Kind	Verifies
AGaussianKernelSeparatesAnXorPatternThatIsNotLinearlySeparable	Fact	A gaussian kernel separates an xor pattern that is not linearly separable.
APolynomialKernelSeparatesAnXorPattern	Fact	A polynomial kernel separates an xor pattern.
ALaplacianKernelSeparatesConcentricCircles	Fact	A laplacian kernel separates concentric circles.
ASigmoidKernelFitsAndRecallsWithoutError	Fact	A sigmoid kernel fits and recalls without error.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsCalibratedTestMetricsOnWellSeparatedClasses	Fact	The evidence reports calibrated test metrics on well separated classes.
GammaResolvesAutomaticallyFromTheNumberOfPredictorsWhenLeftAtZero	Fact	Gamma resolves automatically from the number of predictors when left at zero.

LinearSupportVectorClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.LinearSupportVectorClassifierExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsExactAndInDecisionFunctionSpaceWithTheCoefficientMethod	Fact	To adaptation contribution is exact and in decision function space with the coefficient method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesProbabilitySpaceEitherWay	Fact	To adaptation describes calibration only when calibrated and uses probability space either way.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

LinearSupportVectorClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.LinearSupportVectorClassifierRecallTests — 11 tests.

Test	Kind	Verifies
ClearlyPositiveAndNegativeProbesRecallOnTheCorrectSideOfOneHalf	Fact	Clearly positive and negative probes recall on the correct side of one half.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ExplainContributionsPlusTheBiasSumExactlyToTheRawDecisionFunction	Fact	Explain contributions plus the bias sum exactly to the raw decision function.
ExplainContributionsAreGenuinelyExactNotJustFlagged	Fact	Explain contributions are genuinely exact not just flagged.
ConstantFeatureAcrossTheWholeTrainingSplitIsWarnedAboutButDoesNotThrow	Fact	Constant feature across the whole training split is warned about but does not throw.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleMetricsAndCalibration	Fact	The evidence reports plausible metrics and calibration.
VaryingComplexityValuesAllProduceAValidFiniteModel	Theory	Varying complexity values all produce a valid finite model.

LogisticRegressionCalibrationTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionCalibrationTests — 7 tests.

Test	Kind	Verifies
CalibrationRepairsADeliberatelyDistortedScoreOnTheTestSplit	Fact	Calibration repairs a deliberately distorted score on the test split.
CalibrationIsFittedOnValidationRowsOnly	Fact	Calibration is fitted on validation rows only.
PriorShiftBringsBalancedFitProbabilitiesBackToTheTwoPercentBaseRate	Theory	Prior shift brings balanced fit probabilities back to the two percent base rate.
PriorShiftOffsetMatchesTheClosedFormAndHonoursAnOverride	Fact	Prior shift offset matches the closed form and honours an override.
HeldOutMetricsAreComputedOnTheTestPredictionsAndMatchTheSharedDefinitions	Fact	Held out metrics are computed on the test predictions and match the shared definitions.
TheBayesMinimumRiskThresholdIsTheCostMinimiserOnCalibratedProbabilities	Fact	The bayes minimum risk threshold is the cost minimiser on calibrated probabilities.
AnExplicitThresholdWinsOverTheCostThreshold	Fact	An explicit threshold wins over the cost threshold.

LogisticRegressionExplainTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionExplainTests — 5 tests.

Test	Kind	Verifies
Explain_ContributionsAreWeightTimesInputAndSumWithInterceptToTheRawLogit	Fact	Explain contributions are weight times input and sum with intercept to the raw logit.
Explain_RawLogitMatchesTheExistingRawLogitMethodExactly	Fact	Explain raw logit matches the existing raw logit method exactly.
Explain_ContributionsSumWithInterceptExactlyEvenWithStandardisation	Fact	Explain contributions sum with intercept exactly even with standardisation.
Explain_AMissingInputContributesThroughItsIndicatorWeightNotTheRealWeight	Fact	Explain a missing input contributes through its indicator weight not the real weight.
Explain_APresentInputWithAnIndicatorContributesZeroFromTheIndicatorWeight	Fact	Explain a present input with an indicator contributes zero from the indicator weight.

LogisticRegressionExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionExplainToAdaptationTests — 7 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsExactAndReconstructsTheRawLogit	Fact	To adaptation contribution is exact and reconstructs the raw logit.
ToAdaptation_DescribesCalibrationOnlyWhenCalibrated	Fact	To adaptation describes calibration only when calibrated.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.
ToAdaptation_CarriesPerCoefficientSignificanceWhenPresentAndLeavesItNullOtherwise	Fact	To adaptation carries per coefficient significance when present and leaves it null otherwise.

LogisticRegressionLeakageAndDeterminismTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionLeakageAndDeterminismTests — 6 tests.

Test	Kind	Verifies
NoTestRowInfluencesPreprocessingFitOrCalibration	Fact	No test row influences preprocessing fit or calibration.
ValidationRowsChangeOnlyTheCalibrationAndTrainRowsChangeTheFit	Fact	Validation rows change only the calibration and train rows change the fit.
PreprocessingStatisticsComeFromTrainRowsAlone	Fact	Preprocessing statistics come from train rows alone.
SplitsAreDisjointCoverEveryRowAndKeepClassProportions	Fact	Splits are disjoint cover every row and keep class proportions.
TheSameSeedGivesByteIdenticalResultsAndADifferentSeedDoesNot	Fact	The same seed gives byte identical results and a different seed does not.
TrainWithSplitRejectsOverlappingOrOutOfRangeSplits	Fact	Train with split rejects overlapping or out of range splits.

LogisticRegressionRecallTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionRecallTests — 12 tests.

Test	Kind	Verifies
RecallReproducesTrainingTimePredictionsBitForBitIncludingMissingInputs	Fact	Recall reproduces training time predictions bit for bit including missing inputs.
RecallReproducesTheUncalibratedAndUnstandardisedPipelinesToo	Fact	Recall reproduces the uncalibrated and unstandardised pipelines too.
AMissingInputIsImputedWithTheTrainingMedianAndFlagged	Fact	A missing input is imputed with the training median and flagged.
RecallRejectsMissingValuesWhenTheModelWasTrainedToReject	Fact	Recall rejects missing values when the model was trained to reject.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
TheSnapshotCarriesSignificancesMatchingTheCoefficientTableAndExplainSurfacesThem	Fact	The snapshot carries significances matching the coefficient table and explain surfaces them.
RecallFallsBackToNullSignificancesWhenTheSnapshotPredatesTheField	Fact	Recall falls back to null significances when the snapshot predates the field.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
RecallIsThreadSafeAndDeterministicUnderContention	Fact	Recall is thread safe and deterministic under contention.
TheToolsScoreExactlyLikeTheTrainerAndRecall	Fact	The tools score exactly like the trainer and recall.
TheResultSerialisesForAnLlmWithoutNaNAndWithoutBulkPredictionArrays	Fact	The result serialises for an llm without na n and without bulk prediction arrays.

LogisticRegressionRecoveryTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionRecoveryTests — 5 tests.

Test	Kind	Verifies
EstimatesRecoverTheKnownCoefficientsAtLargeSample	Fact	Estimates recover the known coefficients at large sample.
WaldIntervalsCoverTheTruthAtAboutTheNominalRate	Fact	Wald intervals cover the truth at about the nominal rate.
TruePredictorsAreSignificantAndNoiseFollowsTheFalsePositiveRate	Fact	True predictors are significant and noise follows the false positive rate.
CoefficientTableMatchesAnIndependentPythonImplementation	Fact	Coefficient table matches an independent python implementation.
RidgePenaltyMatchesAnIndependentPythonImplementationAndLeavesTheInterceptUnpenalised	Fact	Ridge penalty matches an independent python implementation and leaves the intercept unpenalised.

LogisticRegressionValidationTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionValidationTests — 11 tests.

Test	Kind	Verifies
RaggedRowsAreRejected	Fact	Ragged rows are rejected.
LabelsOtherThanZeroAndOneAreRejected	Fact	Labels other than zero and one are rejected.
NaNInputsAreRejectedUnderTheRejectPolicyAndImputedOtherwise	Fact	Na n inputs are rejected under the reject policy and imputed otherwise.
ASingleClassIsRejected	Fact	A single class is rejected.
FewerRowsThanVariablesIsRejected	Fact	Fewer rows than variables is rejected.
MismatchedLengthsBlankAndDuplicateNamesAndEmptyInputAreRejected	Fact	Mismatched lengths blank and duplicate names and empty input are rejected.
NonsensicalOptionsAreRejected	Fact	Nonsensical options are rejected.
AVariableWithNoObservedTrainingValueIsRejected	Fact	A variable with no observed training value is rejected.
EveryVariableConstantIsRejected	Fact	Every variable constant is rejected.
ACancelledTokenStopsTrainingBeforeAnyWork	Fact	A cancelled token stops training before any work.
CancellationDuringALongFitIsHonoured	Fact	Cancellation during a long fit is honoured.

LogisticRegressionWarningTests

Jube.Test.AdvancedAnalytics.Slice.LogisticRegressionWarningTests — 14 tests.

Test	Kind	Verifies
AWellPosedProblemRaisesNoModelQualityWarnings	Fact	A well posed problem raises no model quality warnings.
PerfectSeparationIsDetected	Fact	Perfect separation is detected.
QuasiSeparationFromAZeroCellIsFlaggedByHugeCoefficientsOrNonConvergence	Fact	Quasi separation from a zero cell is flagged by huge coefficients or non convergence.
ConstantColumnsAreExcludedAndReported	Fact	Constant columns are excluded and reported.
NearConstantColumnsAreReported	Fact	Near constant columns are reported.
LowEventsPerVariableIsReportedAndAdequateEventsAreNot	Fact	Low events per variable is reported and adequate events are not.
HighVarianceInflationIsMeasuredAndReported	Fact	High variance inflation is measured and reported.
ExactlyDuplicatedPredictorsAreReportedAsInfiniteVariance	Fact	Exactly duplicated predictors are reported as infinite variance.
TinyAndMissingTestSplitsAreFlagged	Fact	Tiny and missing test splits are flagged.
ImbalanceBeyondOneToOneHundredWithoutWeightsIsFlaggedAndWeightsSilenceIt	Fact	Imbalance beyond one to one hundred without weights is flagged and weights silence it.
ACalibrationSlopeFarFromOneIsFlagged	Fact	A calibration slope far from one is flagged.
NonConvergenceIsReportedWhenTheIterationBudgetIsTooSmall	Fact	Non convergence is reported when the iteration budget is too small.
TheNoSignificantPredictorsWarningAppearsExactlyWhenEveryPValueExceedsFivePercent	Fact	The no significant predictors warning appears exactly when every p value exceeds five percent.
BalancedWeightsCarryTheApproximateStandardErrorWarning	Fact	Balanced weights carry the approximate standard error warning.

MahalanobisAnomalyExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.MahalanobisAnomalyExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInTheDistanceSpaceWithTheMahalanobisTermMethod	Fact	To adaptation contribution is inexact and in the distance space with the mahalanobis term method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesDistanceSpaceOtherwise	Fact	To adaptation describes calibration only when calibrated and uses distance space otherwise.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

MahalanobisAnomalyRecallTests

Jube.Test.AdvancedAnalytics.Slice.MahalanobisAnomalyRecallTests — 12 tests.

Test	Kind	Verifies
RecallReproducesTrainingTimeDistancesBitForBit	Fact	Recall reproduces training time distances bit for bit.
TheFittedDistanceMatchesTheTheoreticalMahalanobisDistanceOnAKnownTwoDimensionalGaussian	Fact	The fitted distance matches the theoretical mahalanobis distance on a known two dimensional gaussian.
ExplainDecomposesPerDimensionAndDoesNotSumExactlyToTheDistance	Fact	Explain decomposes per dimension and does not sum exactly to the distance.
AConstantColumnIsExcludedFromTheCovarianceAndCarriesZeroWeight	Fact	A constant column is excluded from the covariance and carries zero weight.
ANearSingularCovarianceIsRegularisedRatherThanThrowing	Fact	A near singular covariance is regularised rather than throwing.
EveryVariableConstantThrowsALoudException	Fact	Every variable constant throws a loud exception.
LabelledTrainingCalibratesWhenBothClassesFallInTheValidationSplitAndUnlabelledDoesNot	Fact	Labelled training calibrates when both classes fall in the validation split and unlabelled does not.
ALargerDistanceCalibratesToAHigherOrEqualProbability	Fact	A larger distance calibrates to a higher or equal probability.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
RecallIsThreadSafeAndDeterministicUnderContention	Fact	Recall is thread safe and deterministic under contention.

MeanShiftClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.MeanShiftClustererRecallTests — 9 tests.

Test	Kind	Verifies
NearestModeDistanceIsGenuinelyTheMinimumAndIsMovedToTheFront	Fact	Nearest mode distance is genuinely the minimum and is moved to the front.
PredictReturnsOneDistancePerModeInModeIndexOrderExceptTheNearestIsFirst	Fact	Predict returns one distance per mode in mode index order except the nearest is first.
SameDataProducesTheSameModesAndTheSameRecallBitForBit	Fact	Same data produces the same modes and the same recall bit for bit.
ADifferentBandwidthCanProduceADifferentModeCount	Fact	A different bandwidth can produce a different mode count.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleModeSummaries	Fact	The evidence reports plausible mode summaries.

MiniBatchKMeansClustererRecallTests

Jube.Test.AdvancedAnalytics.Slice.MiniBatchKMeansClustererRecallTests — 13 tests.

Test	Kind	Verifies
NearestCentroidDistanceIsGenuinelyTheMinimumAndIsMovedToTheFront	Fact	Nearest centroid distance is genuinely the minimum and is moved to the front.
PredictReturnsOneDistancePerCentroidInClusterIndexOrderExceptTheNearestIsFirst	Fact	Predict returns one distance per centroid in cluster index order except the nearest is first.
SameSeedProducesTheSameCentroidsAndTheSameRecallBitForBit	Fact	Same seed produces the same centroids and the same recall bit for bit.
ADifferentSeedCanProduceDifferentCentroids	Fact	A different seed can produce different centroids.
RequestingMoreClustersThanDistinctRowsThrowsAClearException	Fact	Requesting more clusters than distinct rows throws a clear exception.
RequestingMoreClustersThanRowsThrowsAClearException	Fact	Requesting more clusters than rows throws a clear exception.
RequestingABatchSizeLargerThanTheTrainingSplitThrowsAClearException	Fact	Requesting a batch size larger than the training split throws a clear exception.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleClusterSummaries	Fact	The evidence reports plausible cluster summaries.
AnUnsetInitializationBatchSizeIsCappedToTheAvailableTrainingRowsWithoutHanging	Fact	An unset initialization batch size is capped to the available training rows without hanging.

MultipleLinearRegressionExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.MultipleLinearRegressionExplainToAdaptationTests — 7 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsExactAndReconstructsThePredictedValue	Fact	To adaptation contribution is exact and reconstructs the predicted value.
ToAdaptation_IsNeverCalibrated	Fact	To adaptation is never calibrated.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.
ToAdaptation_NeverCarriesPerCoefficientSignificance	Fact	To adaptation never carries per coefficient significance.

MultipleLinearRegressionRecallTests

Jube.Test.AdvancedAnalytics.Slice.MultipleLinearRegressionRecallTests — 13 tests.

Test	Kind	Verifies
RecallReproducesTrainingTimePredictionsBitForBitIncludingMissingInputs	Fact	Recall reproduces training time predictions bit for bit including missing inputs.
RecallReproducesTheUnstandardisedPipelineToo	Fact	Recall reproduces the unstandardised pipeline too.
APredictionIsExactlyTheDotProductOfWeightsAndStandardisedInputsPlusTheIntercept	Fact	A prediction is exactly the dot product of weights and standardised inputs plus the intercept.
AMissingInputIsImputedWithTheTrainingMedianAndFlagged	Fact	A missing input is imputed with the training median and flagged.
RecallRejectsMissingValuesWhenTheModelWasTrainedToReject	Fact	Recall rejects missing values when the model was trained to reject.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
FewerRowsThanVariablesThrowsAClearException	Fact	Fewer rows than variables throws a clear exception.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
RecallIsThreadSafeAndDeterministicUnderContention	Fact	Recall is thread safe and deterministic under contention.
TestSplitRowsNeverTouchTheFittedTrainingStatistics	Fact	Test split rows never touch the fitted training statistics.
SameSeedProducesTheSameSnapshotAndDifferentSeedsDoNot	Fact	Same seed produces the same snapshot and different seeds do not.
TheResultSerialisesForAnLlmWithoutBulkPredictionArrays	Fact	The result serialises for an llm without bulk prediction arrays.

NaiveBayesClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.NaiveBayesClassifierExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsExactAndInLogOddsSpaceWithTheLogLikelihoodRatioMethod	Fact	To adaptation contribution is exact and in log odds space with the log likelihood ratio method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesProbabilitySpaceEitherWay	Fact	To adaptation describes calibration only when calibrated and uses probability space either way.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

NaiveBayesClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.NaiveBayesClassifierRecallTests — 11 tests.

Test	Kind	Verifies
ClearlyPositiveAndNegativeProbesRecallOnTheCorrectSideOfOneHalf	Fact	Clearly positive and negative probes recall on the correct side of one half.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ExplainContributionsPlusThePriorSumExactlyToTheRawLogOdds	Fact	Explain contributions plus the prior sum exactly to the raw log odds.
ExplainContributionsAreGenuinelyExactNotJustFlagged	Fact	Explain contributions are genuinely exact not just flagged.
ConstantFeatureAcrossTheWholeTrainingSplitIsWarnedAboutButDoesNotThrow	Fact	Constant feature across the whole training split is warned about but does not throw.
ASingleTrainingRowInAClassFloorsTheVarianceInsteadOfThrowing	Fact	A single training row in a class floors the variance instead of throwing.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleMetricsAndCalibration	Fact	The evidence reports plausible metrics and calibration.

NeuralNetworkClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.NeuralNetworkClassifierExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInTheLogOddsSpaceWithTheGradientMethod	Fact	To adaptation contribution is inexact and in the log odds space with the gradient method.
ToAdaptation_AlwaysDescribesCalibrationInTheProbabilitySpace	Fact	To adaptation always describes calibration in the probability space.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

NeuralNetworkClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.NeuralNetworkClassifierRecallTests — 10 tests.

Test	Kind	Verifies
RecallLearnsACleanLinearlySeparableBoundaryReliably	Fact	Recall learns a clean linearly separable boundary reliably.
PredictReturnsASingleProbabilityBetweenZeroAndOne	Fact	Predict returns a single probability between zero and one.
SameSeedProducesTheSameWeightsAndTheSameRecallBitForBit	Fact	Same seed produces the same weights and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
BothClassesMustBePresentInTheTrainingSplit	Fact	Both classes must be present in the training split.
ExplainGradientsAreFiniteAndOneEntryPerVariable	Fact	Explain gradients are finite and one entry per variable.

NeuralNetworkRegressorExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.NeuralNetworkRegressorExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInTheScoreSpaceWithTheGradientMethod	Fact	To adaptation contribution is inexact and in the score space with the gradient method.
ToAdaptation_NeverCalibratesAndUsesTheScoreSpace	Fact	To adaptation never calibrates and uses the score space.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

NeuralNetworkRegressorRecallTests

Jube.Test.AdvancedAnalytics.Slice.NeuralNetworkRegressorRecallTests — 10 tests.

Test	Kind	Verifies
RecallLearnsANonlinearRelationshipReliably	Fact	Recall learns a nonlinear relationship reliably.
PredictReturnsASingleFiniteValue	Fact	Predict returns a single finite value.
SameSeedProducesTheSameWeightsAndTheSameRecallBitForBit	Fact	Same seed produces the same weights and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
AConstantTrainingTargetProducesAFiniteConstantPredictionInsteadOfDiverging	Fact	A constant training target produces a finite constant prediction instead of diverging.
ExplainGradientsAreFiniteAndOneEntryPerVariable	Fact	Explain gradients are finite and one entry per variable.

OneClassSupportVectorAnomalyRecallTests

Jube.Test.AdvancedAnalytics.Slice.OneClassSupportVectorAnomalyRecallTests — 11 tests.

Test	Kind	Verifies
FarOutliersScoreHigherThanTheNormalClusterTheyWereFittedOn	Fact	Far outliers score higher than the normal cluster they were fitted on.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
UnlabelledTrainingFitsWithNoCalibratorAndReportsTheRawScore	Fact	Unlabelled training fits with no calibrator and reports the raw score.
LabelledTrainingFitsACalibratorAndProducesARealRocAuc	Fact	Labelled training fits a calibrator and produces a real roc auc.
NuRoughlyControlsHowManyTrainingRowsBecomeSupportVectors	Fact	Nu roughly controls how many training rows become support vectors.
TheEvidenceReportsPlausibleScorePercentilesAndSupportVectorCount	Fact	The evidence reports plausible score percentiles and support vector count.
RecallIsThreadSafeAndDeterministicUnderContention	Fact	Recall is thread safe and deterministic under contention.

PredictorSignificanceTests

Jube.Test.AdvancedAnalytics.Slice.PredictorSignificanceTests — 6 tests.

Test	Kind	Verifies
MatchesAnIndependentPythonUnivariateLogisticImplementation	Fact	Matches an independent python univariate logistic implementation.
BenjaminiHochbergMatchesTheOriginalPaperExample	Fact	Benjamini hochberg matches the original paper example.
BenjaminiHochbergKeepsInputOrderSkipsUndefinedAndNeverExceedsOne	Fact	Benjamini hochberg keeps input order skips undefined and never exceeds one.
TruePredictorsAreFoundAndAdjustedNoiseFalsePositivesAreRare	Fact	True predictors are found and adjusted noise false positives are rare.
SeparatedConstantAndSparseVariablesReturnANoteInsteadOfNumbers	Fact	Separated constant and sparse variables return a note instead of numbers.
InvalidArgumentsAreRejectedAndCancellationIsHonoured	Fact	Invalid arguments are rejected and cancellation is honoured.

PrincipalComponentRecallTests

Jube.Test.AdvancedAnalytics.Slice.PrincipalComponentRecallTests — 10 tests.

Test	Kind	Verifies
RecallReproducesTrainingTimeProjectionsBitForBit	Fact	Recall reproduces training time projections bit for bit.
Pc1IsTheDominantDirectionOfPositivelyCorrelatedVariables	Fact	Pc1 is the dominant direction of positively correlated variables.
RecallMatchesAccordsOwnTransformOnTheSameStandardisedRows	Fact	Recall matches accords own transform on the same standardised rows.
TrainingIsDeterministicForTheSameSeed	Fact	Training is deterministic for the same seed.
FewerRowsThanVariablesStillTrainsButCapsTheAvailableComponents	Fact	Fewer rows than variables still trains but caps the available components.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
ElementZeroIsAlwaysThePc1Score	Fact	Element zero is always the pc1 score.

RandomForestClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.RandomForestClassifierExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactAndInVoteFractionSpaceWithTheTreePathContributionMethod	Fact	To adaptation contribution is inexact and in vote fraction space with the tree path contribution method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesProbabilitySpace	Fact	To adaptation describes calibration only when calibrated and uses probability space.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

RandomForestClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.RandomForestClassifierRecallTests — 11 tests.

Test	Kind	Verifies
SeparableDataProducesHighProbabilityForPositivesAndLowForNegatives	Fact	Separable data produces high probability for positives and low for negatives.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
PredictedProbabilitiesAreAlwaysBetweenZeroAndOne	Fact	Predicted probabilities are always between zero and one.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
RequiringBothClassesThrowsAClearExceptionWhenOnlyOneIsPresent	Fact	Requiring both classes throws a clear exception when only one is present.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleForestSummaries	Fact	The evidence reports plausible forest summaries.
ExplainRawScoreMatchesThePipelineRawScore	Fact	Explain raw score matches the pipeline raw score.

RandomForestRegressorExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.RandomForestRegressorExplainToAdaptationTests — 7 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamily	Fact	To adaptation carries the value and model name and family.
ToAdaptation_ContributionIsInexactInScoreSpaceWithTheTreePathContributionMethod	Fact	To adaptation contribution is inexact in score space with the tree path contribution method.
ToAdaptation_IsNeverCalibratedRegardlessOfRequestedValue	Fact	To adaptation is never calibrated regardless of requested value.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.
ToAdaptation_NeverCarriesPerVariableSignificance	Fact	To adaptation never carries per variable significance.

RandomForestRegressorRecallTests

Jube.Test.AdvancedAnalytics.Slice.RandomForestRegressorRecallTests — 13 tests.

Test	Kind	Verifies
StrongSignalDataProducesAHighTestRSquared	Fact	Strong signal data produces a high test r squared.
RecallReproducesTrainingTimePredictionsBitForBitIncludingMissingInputs	Fact	Recall reproduces training time predictions bit for bit including missing inputs.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ADifferentSeedCanProduceADifferentSnapshot	Fact	A different seed can produce a different snapshot.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleForestSummaries	Fact	The evidence reports plausible forest summaries.
ExplainPredictedValueEqualsBaseValuePlusContributions	Fact	Explain predicted value equals base value plus contributions.
FeatureFractionResolvesToAPlausibleFeaturesPerSplit	Theory	Feature fraction resolves to a plausible features per split.
FewerRowsThanTreesStillTrainsSuccessfully	Fact	Fewer rows than trees still trains successfully.
RecallIsThreadSafeAndDeterministicUnderContention	Fact	Recall is thread safe and deterministic under contention.

TreeAugmentedNaiveBayesClassifierExplainToAdaptationTests

Jube.Test.AdvancedAnalytics.Slice.TreeAugmentedNaiveBayesClassifierExplainToAdaptationTests — 6 tests.

Test	Kind	Verifies
ToAdaptation_IsAssignableToTheSharedExplanationInterface	Fact	To adaptation is assignable to the shared explanation interface.
ToAdaptation_CarriesTheValueAndModelNameAndFamilyAndChowLiuStructureLearning	Fact	To adaptation carries the value and model name and family and chow liu structure learning.
ToAdaptation_ContributionIsExactAndInLogOddsSpaceWithTheLogLikelihoodRatioMethod	Fact	To adaptation contribution is exact and in log odds space with the log likelihood ratio method.
ToAdaptation_DescribesCalibrationOnlyWhenCalibratedAndUsesProbabilitySpaceEitherWay	Fact	To adaptation describes calibration only when calibrated and uses probability space either way.
ToAdaptation_MapsSourceFromProcessingTypeIdsWhenGiven	Fact	To adaptation maps source from processing type ids when given.
ToAdaptation_LeavesSourceNullWhenProcessingTypeIdsAreNotGiven	Fact	To adaptation leaves source null when processing type ids are not given.

TreeAugmentedNaiveBayesClassifierRecallTests

Jube.Test.AdvancedAnalytics.Slice.TreeAugmentedNaiveBayesClassifierRecallTests — 12 tests.

Test	Kind	Verifies
ClearlyPositiveAndNegativeProbesRecallOnTheCorrectSideOfOneHalf	Fact	Clearly positive and negative probes recall on the correct side of one half.
SameSeedProducesTheSameSnapshotAndTheSameRecallBitForBit	Fact	Same seed produces the same snapshot and the same recall bit for bit.
ExplainContributionsPlusThePriorSumExactlyToTheRawLogOdds	Fact	Explain contributions plus the prior sum exactly to the raw log odds.
ExplainContributionsAreGenuinelyExactNotJustFlagged	Fact	Explain contributions are genuinely exact not just flagged.
ConstantFeatureAcrossTheWholeTrainingSplitIsWarnedAboutButDoesNotThrow	Fact	Constant feature across the whole training split is warned about but does not throw.
RecallRejectsARowOfTheWrongLength	Fact	Recall rejects a row of the wrong length.
AMissingInputIsImputedWithTheTrainingMedian	Fact	A missing input is imputed with the training median.
TheSnapshotRoundTripsThroughTheEnvelopeUnchanged	Fact	The snapshot round trips through the envelope unchanged.
ATamperedOrForeignSnapshotIsRefused	Fact	A tampered or foreign snapshot is refused.
TheEvidenceReportsPlausibleMetricsAndCalibration	Fact	The evidence reports plausible metrics and calibration.
TheChowLiuTreeDiscoversAGenuineWithinClassDependencyAndFitsAMeaningfulSlope	Fact	The chow liu tree discovers a genuine within class dependency and fits a meaningful slope.
ASingleVariableModelHasNoParentAndStillPredicts	Fact	A single variable model has no parent and still predicts.