Crate sim_lib_numbers_stats

Expand description

Probability, descriptive and robust benchmark statistics, deterministic clustering, streaming quantiles, finite Markov and hidden-state sequence inference, and fairness helpers for number-domain data.

Descriptive statistics and disparate-impact helpers also expose Claim surfaces. The Claim values carry their subject, predicate, and evidence table as inspectable runtime data, so callers can browse both the computed metric and the inputs used to justify it. fit_markov keeps the finite vocabulary, exact counts, additive smoothing, held-out likelihood, deterministic serialization, and corpus provenance inspectable instead of hiding learned weights. QuantileSketch makes rank error and retained memory explicit; forward_backward, viterbi, and fit_hmm keep normalization, convergence, bounded work, numerical repair, and termination evidence. fit_kmeans and fit_gmm add seeded initialization, bounded convergence, regularized covariance, singular-component policy, and model-selection evidence without taking ownership of sequence alignment. exact_binary_interval, paired and clustered bootstrap, sealed RegisteredLookSequence contracts, and fit_isotonic provide the bounded mathematical owner for sequential study decisions. They reuse BootstrapControl and keep confidence, work, cluster independence, censoring, and replay evidence explicit.

Structs§

BinaryInterval: A two-sided Clopper–Pearson interval for a finite binary count.
BinaryOutcomeCounts: Counts for a binary selection or outcome table.
BootstrapControl: Controls a deterministic bootstrap of the candidate-minus-baseline mean.
BootstrapEffectInterval: A percentile bootstrap interval for the candidate-minus-baseline mean.
ClusterSample: One independent cluster, retaining its stable identity and paired rows.
CorpusProvenance: Stable provenance attached to every fitted transition model.
CoverageEvidence: Coverage and replay evidence returned with every design.
DisparateImpact: Disparate-impact summary for two binary-outcome groups.
FairnessClaimValue: A first-class runtime object wrapping the fairness Claim and its evidence.
FairnessEvidence: Evidence carried by the disparate-impact fairness Claim.
FiniteTransitionMatrix: A finite state vocabulary and row-stochastic transition matrix.
ForwardBackward: Normalized forward, backward, and smoothed posterior state probabilities.
GmmControl: Deterministic convergence and work policy for Gaussian-mixture EM.
GmmEvidence: Convergence, regularization, work, and selection evidence from EM.
GmmModel: Inspectable fitted Gaussian-mixture parameters.
GmmReport: Fitted mixture and complete EM evidence.
GmmSpec: Component count, covariance family, regularization, and singular policy.
HiddenMarkovModel: A finite hidden Markov model with inspectable transition and emission rows.
HmmFitControl: Deterministic initialization, convergence, and work policy for Baum-Welch.
HmmFitEvidence: Convergence, likelihood, repair, seed, and termination evidence.
HmmFitReport: A fitted hidden-state model together with complete termination evidence.
InferenceEvidence: Numerical and likelihood evidence from normalized sequence inference.
IsotonicFit: Inspectable weighted pool-adjacent-violators fit.
IsotonicPoint: One weighted raw point supplied to isotonic regression.
KMeansControl: Deterministic initialization, convergence, restart, and work policy for k-means.
KMeansModel: Inspectable k-means centroids and stable cluster assignments.
KMeansReport: Selected model and complete multi-restart evidence.
KMeansRestartEvidence: Convergence evidence for one bounded restart.
KsResult: One- or two-sample Kolmogorov-Smirnov report.
LatinHypercubePlan: Latin hypercube request.
MarkovModel: A finite first-order Markov model retaining exact counts and fitting policy.
MarkovPolicy: Explicit fitting and evaluation policy for a finite first-order model.
ModelReport: A fitted value together with training and held-out evidence.
ModelSelectionEvidence: AIC and BIC evidence for comparing fitted component counts.
PosteriorPath: Per-position maximum-posterior state path.
QuantileEstimate: One quantile estimate together with its retained rank evidence.
QuantilePolicy: Error and memory policy for a QuantileSketch.
QuantileSketch: A deterministic Greenwald-Khanna streaming quantile summary. The sketch stores observations exactly through QuantilePolicy::exact_threshold. Larger streams use rank intervals and deterministic compression. Compatible sketches can be merged without replaying their source streams. The hard entry limit makes memory admission explicit: an operation that cannot retain the requested rank error is rejected transactionally.
RegisteredLook: A pre-registered sequential look and its allocated false-elimination mass.
RegisteredLookSequence: An alpha-spent contract for bounded observations in [0, 1].
SampleDesign: A generated, reconstructable design and its coverage evidence.
SamplerReceipt: Replay receipt captured at an observable sampling boundary.
SamplerState: Canonical, architecture-independent sampler state.
SeededSampler: A public deterministic sampler with explicit version, work, state and forks.
SequentialInterval: A Hoeffding interval valid at its pre-registered look under the sealed budget.
SobolPlan: Sobol base-2 digital-net request. The reviewed direction table is bounded to four dimensions.
StandardizedMoments: Standardized third and fourth moment report.
StatsClaimEvidence: Evidence carried by a descriptive statistics Claim.
StatsClaimValue: A first-class runtime object wrapping a descriptive statistics Claim.
StatsNumbersLib: Library that installs the runtime statistics functions.
SweepPlan: Boundary-injection wrapper for any already generated unit-cube design.
TransitionScore: Aggregate likelihood evidence for a collection of state sequences.
UntestedRegion: Caller-described region intentionally absent from a sweep.
ViterbiPath: Maximum-probability hidden-state path and its joint log probability.

Enums§

ClusteringError: Errors returned by clustering and mixture-model fitting.
CovarianceType: Covariance representation fitted for every Gaussian component.
DesignError: Fail-closed design and sampler errors.
EmissionModel: Discrete or scalar Gaussian emissions for a finite hidden-state model.
GaussianCovariance: Inspectable covariance parameters for one Gaussian component.
GmmTermination: Why bounded EM stopped.
HmmError: Failure while constructing, fitting, or running hidden-state inference.
HmmSpec: Hidden-state and emission family requested from fit_hmm.
HmmTermination: Why bounded Baum-Welch stopped.
KMeansSearchTermination: Why the bounded multi-restart search stopped.
KMeansTermination: Why one k-means restart stopped.
KsMethod: Kolmogorov-Smirnov evaluation policy and applicability identity.
MarkovError: Failure while validating, fitting, scoring, or serializing a Markov model.
MomentConvention: Explicit finite-sample estimator convention.
QuantileError: Failure while configuring, updating, merging, or querying a quantile sketch.
SamplerAlgorithm: Stable identity of a deterministic sampler algorithm.
Scramble: Optional reviewed digital scrambling policy.
Sequence: One homogeneous observation sequence accepted by HMM fitting.
SingularComponentPolicy: Policy for an EM component with negligible responsibility or singular covariance.
StatsError: Errors returned by probability, statistics, and fairness helpers.
ThresholdReadout: Threshold crossing evidence, including censoring beyond the tested range.
TransitionError: Failure while constructing a finite row-stochastic transition matrix.

Statics§

RECIPES: Cookbook recipes for this lib, embedded at build time.

Traits§

HmmObservation: Observation accepted by generic HMM inference.

Functions§

bayesian_update: Computes prior * likelihood / evidence and validates the posterior.
bayesian_update_binary: Computes a binary-test posterior from prior, true-positive, and false-positive rates.
bootstrap_mean_difference_interval: Bootstraps the difference between candidate and baseline arithmetic means.
clustered_bootstrap_interval: Deterministically resamples whole independent clusters with replacement.
disparate_impact: Computes disparate impact and the four-fifths pass/fail flag.
entropy: Computes Shannon entropy in bits for a probability vector that sums to one.
exact_binary_interval: Computes an exact equal-tailed finite-count binary interval.
exact_quantile: Computes the exact linearly interpolated quantile of finite small data.
fairness_claim: Builds the public disparate-impact fairness Claim and interns its evidence object.
fairness_claim_value: Wraps a fairness Claim as a runtime value.
fit_gmm: Fits a regularized diagonal or full-covariance Gaussian mixture.
fit_hmm: Fits a discrete- or continuous-emission HMM with bounded Baum-Welch.
fit_isotonic: Fits a weighted nondecreasing curve with pool-adjacent-violators.
fit_kmeans: Fits deterministic seeded k-means with k-means++ initialization.
fit_markov: Fits an inspectable finite first-order model and reports held-out evidence.
fnv1a64: Computes a stable FNV-1a digest for small transparent fixture corpora.
forward_backward: Runs normalized forward/backward inference in the log domain.
four_fifths_ratio: Computes the comparison/reference rate ratio used by the four-fifths rule.
kolmogorov_smirnov_one_sample: Computes a one-sample KS statistic against a caller-supplied continuous CDF.
kolmogorov_smirnov_two_sample: Computes the two-sample KS statistic. Inputs must represent independent samples.
mean: Computes the arithmetic mean of finite values.
median_absolute_deviation: Computes the raw median absolute deviation from the sample median.
normal_cdf: Standard normal cumulative probability.
normal_density: Standard normal probability density.
normal_quantile: Standard normal quantile, found by bounded monotone inversion.
normal_survival: Standard normal survival probability, evaluated directly in the positive tail.
paired_bootstrap_interval: Deterministically bootstraps paired candidate-minus-baseline effects.
population_variance: Computes population variance with divisor n.
posterior_decode: Decodes the independently most probable hidden state at each position.
sample_variance: Computes sample variance with divisor n - 1.
standardized_moments: Computes standardized third and fourth moments under an explicit convention.
stats_claims_symbol: Returns the symbol bound to the stats/claims batch operation.
stats_clustered_bootstrap_symbol: Returns the symbol bound to cluster-preserving bootstrap intervals.
stats_disparate_impact_claim_symbol: Returns the symbol bound to the stats/disparate-impact-claim operation.
stats_entropy_claim_symbol: Returns the symbol bound to the stats/entropy-claim operation.
stats_exact_binary_interval_symbol: Returns the symbol bound to exact finite binary intervals.
stats_gmm_symbol: Returns the symbol bound to regularized Gaussian-mixture EM.
stats_isotonic_symbol: Returns the symbol bound to weighted isotonic fitting.
stats_kmeans_symbol: Returns the symbol bound to deterministic bounded k-means.
stats_mean_claim_symbol: Returns the symbol bound to the stats/mean-claim operation.
stats_paired_bootstrap_symbol: Returns the symbol bound to paired bootstrap intervals.
stats_registered_look_symbol: Returns the symbol bound to registered-look intervals.
stats_result_claim: Builds the public descriptive statistics Claim and interns its evidence object.
stats_result_claim_value: Wraps a descriptive statistics Claim as a runtime value.
stats_variance_claim_symbol: Returns the symbol bound to the stats/variance-claim operation.
student_t_cdf: Student-t cumulative probability.
student_t_density: Student-t probability density for positive degrees of freedom.
student_t_quantile: Student-t quantile, found by bounded monotone inversion.
student_t_survival: Student-t survival probability, using the direct beta tail for positive values.
variance: Computes population variance.
viterbi: Finds the maximum joint-probability hidden-state path in the log domain.

Type Aliases§

StateId: Stable numeric hidden-state identifier produced by fit_hmm.
StatsResult: Result alias for the statistics helpers, fixing the error to StatsError.