Crate quietset 

quietset — filter datasets by label stability.

Quick start

use quietset::{Observation, ScoreConfig, score_all};

let obs = vec![
    Observation { sample_id: "a".into(), label: Some("win".into()), score: Some(0.9), ..Default::default() },
    Observation { sample_id: "a".into(), label: Some("win".into()), score: Some(0.88), ..Default::default() },
];
let reports = score_all(obs, &ScoreConfig::default());
assert_eq!(reports[0].decision, quietset::Decision::Keep);

Re-exports

pub use active_review::ActionTargets;

pub use active_review::ActiveReviewCosts;

pub use active_review::ActiveReviewEntry;

pub use active_review::ActiveReviewWeights;

pub use active_review::RankBy;

pub use active_review::action_target_fields;

pub use active_review::rank_active_review;

pub use active_review::select_within_budget;

pub use agreement::compute_fleiss_kappa;

pub use agreement::compute_krippendorff_alpha;

pub use agreement::compute_spearman_rank_correlation;

pub use block::BlockClass;

pub use block::BlockStabilityReport;

pub use block::BlockThresholds;

pub use block::LossRecipeIssue;

pub use block::compute_block_report;

pub use block::loss_recipe_comparable;

pub use block::loss_recipe_issues;

pub use block::score_all_blocks;

pub use calibration::CalibrationOptions;

pub use calibration::CalibrationResult;

pub use calibration::GroupKey;

pub use calibration::GroupLeakage;

pub use calibration::compute_calibration;

pub use calibration::compute_calibration_with_options;

pub use calibration::group_leakage;

pub use calibration::leaked_group_keys;

pub use config::DecisionScore;

pub use config::LatentTruthSafety;

pub use config::MinRequirements;

pub use config::ScoreConfig;

pub use config::ScoreDispersion;

pub use config::ScoreWeights;

pub use decision::Thresholds;

pub use error::Error;

pub use error::Result;

pub use latent_truth::LatentTruthEstimate;

pub use latent_truth::compute_latent_truth;

pub use observation::Observation;

pub use observation::parse_csv;

pub use observation::parse_jsonl;

pub use preflight::BlockRunIdIssue;

pub use preflight::CheckpointCorrespondence;

pub use preflight::FieldCoverage;

pub use preflight::PreflightReport;

pub use preflight::PreflightTarget;

pub use preflight::SeedCoverage;

pub use preflight::SidePreflight;

pub use preflight::compute_preflight;

pub use preflight::required_fields;

pub use schema::Decision;

pub use schema::StabilityComponents;

pub use schema::StabilityReport;

pub use scoring::compute_report;

pub use scoring::score_all;

pub use stable_wrong::BudgetStableWrongRate;

pub use stable_wrong::EvaluatorStableWrongRate;

pub use stable_wrong::ModelStableWrongRate;

pub use stable_wrong::StableWrongBreakdown;

pub use stable_wrong::stable_wrong_breakdown;

pub use stream::StreamingScorer;

pub use trajectory_audit::ComparisonIssue;

pub use trajectory_audit::TrajectoryAuditEntry;

pub use trajectory_audit::compute_trajectory_audit;

pub use trajectory_audit::group_size_mismatches;

pub use weighting::compute_evaluator_label_weights;

pub use weighting::compute_evaluator_reliability;

pub use weighting::compute_evaluator_weights;

pub use weighting::compute_weighted_majority;

pub use weighting::compute_weighted_majority_by_label;

Modules

active_review

agreement

block

Block-level trajectory stability: groups observations by block_id (e.g. a 32-sample training block) instead of sample_id, and classifies each block’s training effect as stable growth/shrink, sensitive to one specific condition axis, or pathological.

calibration

config

decision

error

group

latent_truth

observation

preflight

Instrumentation preflight: dataset-level coverage and completeness checks, run before a real trajectory-audit, so a missing block_id/shuffle_seed is caught while the experiment can still be re-instrumented rather than discovered after the fact.

schema

scoring

stable_wrong

stream

trajectory_audit

Destructive-sample detection: diffs block-level trajectory stability between a “before” and “after” observation set (e.g. two training checkpoints), surfacing blocks whose dead/ saturated-unit rates spiked or whose net trajectory effect flipped.

weighting

Functions

score_all_gold_weighted

Score all samples with the gold-label-reliability-weighted vote driving stability_score/ decision (2-pass). See DecisionScore::GoldWeighted.

score_all_latent_truth

Score all samples with the Dawid-Skene EM latent-truth vote driving stability_score/ decision (2-pass). See DecisionScore::LatentTruth.

score_all_weighted

Score all samples with reliability-weighted majority voting (2-pass).