foxhunt/crates/ml-ensemble/src/confidence.rs

//! Confidence-weighted voting and uncertainty quantification for ensemble predictions
//!
//! This module provides confidence aggregation mechanisms that combine predictions
//! from multiple models while accounting for prediction uncertainty, model reliability,
//! and ensemble confidence intervals.
//!
//! ## Key types
//! - [`ConfidenceAggregator`] — main entry point for uncertainty-aware ensemble voting
//! - [`UncertaintyQuantifier`] — decomposes uncertainty into epistemic / aleatoric / disagreement
//! - [`ReliabilityScorer`] — exponential-decay model reliability tracking
//! - [`IntervalCombiner`] — prediction intervals at 68% / 95% / 99% confidence levels

use anyhow::Result;
use serde::{Deserialize, Serialize};
use std::collections::HashMap;
use tracing::{debug, warn};

use crate::ModelPrediction;

// ── Backwards-compatible stub kept for aggregator.rs ──────────────────────────

/// Simple confidence calculator (used by [`super::aggregator::SignalAggregator`]).
#[derive(Debug)]
pub struct ConfidenceCalculator;

impl ConfidenceCalculator {
    pub fn new() -> Self {
        Self
    }
}

// ── Core aggregator ───────────────────────────────────────────────────────────

/// Confidence aggregator for uncertainty-aware ensemble voting.
///
/// Combines predictions from multiple models with proper uncertainty
/// decomposition, reliability tracking, and prediction interval estimation.
#[derive(Debug)]
pub struct ConfidenceAggregator {
    /// Uncertainty quantification engine
    uncertainty_quantifier: UncertaintyQuantifier,
    /// Model reliability scoring system
    reliability_scorer: ReliabilityScorer,
    /// Prediction interval combination engine
    interval_combiner: IntervalCombiner,
    /// Aggregation configuration
    config: AggregationConfig,
}

// ── Sub-components ────────────────────────────────────────────────────────────

/// Uncertainty quantification engine.
///
/// Decomposes total prediction uncertainty into:
/// - **Epistemic** — model disagreement (reducible with more data / models)
/// - **Aleatoric** — inherent data noise (irreducible)
/// - **Disagreement** — recent historical model spread
#[derive(Debug)]
pub struct UncertaintyQuantifier {
    epistemic_config: EpistemicConfig,
    aleatoric_config: AleatoricConfig,
    disagreement_tracker: DisagreementTracker,
}

/// Model reliability scoring system with exponential decay.
#[derive(Debug)]
pub struct ReliabilityScorer {
    reliability_history: HashMap<String, Vec<ReliabilityRecord>>,
    decay_factor: f64,
    min_observations: usize,
}

/// Prediction interval combination engine.
#[derive(Debug)]
pub struct IntervalCombiner {
    confidence_levels: Vec<f64>,
}

/// Model disagreement tracker with weighted recency.
#[derive(Debug)]
pub struct DisagreementTracker {
    disagreement_history: Vec<DisagreementRecord>,
    max_history_length: usize,
    warning_threshold: f64,
}

// ── Output types ──────────────────────────────────────────────────────────────

/// Enhanced ensemble prediction with full uncertainty quantification.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct EnsemblePredictionWithUncertainty {
    /// Central prediction value
    pub prediction: f64,
    /// Overall ensemble confidence (0.0–1.0)
    pub confidence: f64,
    /// Uncertainty bounds (lower, upper) at 2-sigma
    pub uncertainty_bounds: (f64, f64),
    /// Prediction intervals at multiple confidence levels
    pub prediction_intervals: Vec<PredictionInterval>,
    /// Per-model contributions with reliability scores
    pub model_contributions: HashMap<String, ModelContribution>,
    /// Uncertainty decomposition into epistemic / aleatoric / disagreement
    pub uncertainty_decomposition: UncertaintyDecomposition,
    /// Weight-averaged ensemble reliability (0.0–1.0)
    pub ensemble_reliability: f64,
    /// Prediction timestamp
    pub timestamp: chrono::DateTime<chrono::Utc>,
}

/// Individual model contribution to ensemble prediction.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ModelContribution {
    /// Model's raw prediction
    pub prediction: f64,
    /// Model's confidence
    pub confidence: f64,
    /// Model's reliability score
    pub reliability: f64,
    /// Model's weight in ensemble
    pub weight: f64,
    /// Weighted contribution to final prediction
    pub weighted_contribution: f64,
    /// Model's prediction interval
    pub prediction_interval: Option<PredictionInterval>,
}

/// Prediction interval at a specific confidence level.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct PredictionInterval {
    /// Confidence level (e.g., 0.95 for 95%)
    pub confidence_level: f64,
    /// Lower bound
    pub lower_bound: f64,
    /// Upper bound
    pub upper_bound: f64,
    /// Interval width
    pub width: f64,
}

/// Uncertainty decomposition into components.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct UncertaintyDecomposition {
    /// Epistemic uncertainty (model uncertainty — reducible)
    pub epistemic: f64,
    /// Aleatoric uncertainty (data uncertainty — irreducible)
    pub aleatoric: f64,
    /// Model disagreement component
    pub disagreement: f64,
    /// Total uncertainty = sqrt(epistemic² + aleatoric² + disagreement²)
    pub total: f64,
}

/// Reliability record for a model.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ReliabilityRecord {
    pub timestamp: chrono::DateTime<chrono::Utc>,
    pub accuracy: f64,
    pub calibration: f64,
    pub confidence_reliability: f64,
    pub actual_outcome: Option<f64>,
}

/// Disagreement record between models.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct DisagreementRecord {
    pub timestamp: chrono::DateTime<chrono::Utc>,
    pub magnitude: f64,
    pub models: Vec<String>,
    pub prediction_spread: f64,
}

// ── Configuration types ───────────────────────────────────────────────────────

/// Aggregation configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AggregationConfig {
    /// Weight confidence by reliability
    pub weight_by_reliability: bool,
    /// Minimum confidence threshold
    pub min_confidence_threshold: f64,
    /// Maximum uncertainty allowed
    pub max_uncertainty_threshold: f64,
    /// Outlier detection enabled
    pub outlier_detection_enabled: bool,
    /// Outlier threshold (in standard deviations)
    pub outlier_threshold: f64,
}

/// Epistemic uncertainty configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct EpistemicConfig {
    /// Use model disagreement as proxy for epistemic uncertainty
    pub use_model_disagreement: bool,
    /// Bayesian uncertainty estimation
    pub bayesian_estimation: bool,
    /// Monte Carlo dropout samples
    pub mc_dropout_samples: usize,
}

/// Aleatoric uncertainty configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AleatoricConfig {
    /// Heteroscedastic noise modeling
    pub heteroscedastic_noise: bool,
    /// Noise variance estimation method
    pub variance_estimation_method: VarianceEstimationMethod,
    /// Historical volatility window
    pub volatility_window: usize,
}

/// Calibration parameters for prediction intervals.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CalibrationParams {
    /// Temperature scaling parameter
    pub temperature: f64,
    /// Platt scaling enabled
    pub platt_scaling: bool,
    /// Isotonic regression enabled
    pub isotonic_regression: bool,
}

// ── Enums ─────────────────────────────────────────────────────────────────────

/// Methods for combining prediction intervals.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum CombinationMethod {
    /// Weighted average of intervals
    WeightedAverage,
    /// Conservative union of intervals
    ConservativeUnion,
    /// Bayesian model averaging
    BayesianAveraging,
    /// Mixture of Gaussians
    MixtureOfGaussians,
}

/// Methods for variance estimation.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum VarianceEstimationMethod {
    /// Rolling window variance
    RollingWindow,
    /// EWMA variance
    EWMA,
    /// GARCH modeling
    GARCH,
    /// Realized volatility
    RealizedVolatility,
}

// ── ConfidenceAggregator implementation ───────────────────────────────────────

impl ConfidenceAggregator {
    /// Create a new confidence aggregator with the given config.
    pub fn new(config: AggregationConfig) -> Self {
        let uncertainty_quantifier =
            UncertaintyQuantifier::new(EpistemicConfig::default(), AleatoricConfig::default());
        let reliability_scorer = ReliabilityScorer::new(0.95_f64, 10_usize);
        let interval_combiner = IntervalCombiner::new(
            CombinationMethod::WeightedAverage,
            vec![0.68_f64, 0.95_f64, 0.99_f64],
            CalibrationParams::default(),
        );

        Self {
            uncertainty_quantifier,
            reliability_scorer,
            interval_combiner,
            config,
        }
    }

    /// Aggregate predictions with full uncertainty quantification.
    ///
    /// Returns an [`EnsemblePredictionWithUncertainty`] containing the central
    /// prediction, confidence, uncertainty bounds, per-model contributions, and
    /// a full epistemic / aleatoric / disagreement decomposition.
    pub fn aggregate_with_uncertainty(
        &mut self,
        predictions: HashMap<String, ModelPrediction>,
        weights: &HashMap<String, f64>,
    ) -> Result<EnsemblePredictionWithUncertainty> {
        debug!(
            "Aggregating {} predictions with uncertainty quantification",
            predictions.len()
        );

        if predictions.is_empty() {
            anyhow::bail!("Cannot aggregate empty predictions");
        }

        // Filter outliers if enabled
        let filtered_predictions = if self.config.outlier_detection_enabled {
            self.filter_outliers(predictions)?
        } else {
            predictions
        };

        // Calculate model reliability scores
        let reliability_scores = self.calculate_reliability_scores(&filtered_predictions);

        // Compute ensemble prediction
        let (ensemble_prediction, model_contributions) =
            self.compute_weighted_prediction(&filtered_predictions, weights, &reliability_scores)?;

        // Quantify uncertainty components
        let uncertainty_decomposition = self
            .uncertainty_quantifier
            .quantify_uncertainty(&filtered_predictions, weights)?;

        // Calculate prediction intervals
        let prediction_intervals = self
            .interval_combiner
            .combine_intervals(&filtered_predictions, weights)?;

        // Calculate uncertainty bounds (2-sigma)
        let uncertainty_bounds =
            self.calculate_uncertainty_bounds(ensemble_prediction, &uncertainty_decomposition);

        // Calculate ensemble reliability
        let ensemble_reliability =
            self.calculate_ensemble_reliability(&reliability_scores, weights);

        // Compute overall confidence
        let confidence =
            self.compute_ensemble_confidence(&uncertainty_decomposition, ensemble_reliability);

        // Update disagreement tracking
        self.update_disagreement_tracking(&filtered_predictions);

        Ok(EnsemblePredictionWithUncertainty {
            prediction: ensemble_prediction,
            confidence,
            uncertainty_bounds,
            prediction_intervals,
            model_contributions,
            uncertainty_decomposition,
            ensemble_reliability,
            timestamp: chrono::Utc::now(),
        })
    }

    /// Update reliability history for a model.
    pub fn update_reliability(&mut self, model_name: String, record: ReliabilityRecord) {
        self.reliability_scorer
            .update_reliability(model_name, record);
    }

    /// Get current model reliability scores.
    pub fn get_reliability_scores(&self) -> HashMap<String, f64> {
        self.reliability_scorer.get_current_scores()
    }

    // ── private helpers ───────────────────────────────────────────────────

    fn calculate_reliability_scores(
        &self,
        predictions: &HashMap<String, ModelPrediction>,
    ) -> HashMap<String, f64> {
        predictions
            .keys()
            .map(|name| {
                let reliability = self.reliability_scorer.get_model_reliability(name);
                (name.clone(), reliability)
            })
            .collect()
    }

    fn compute_weighted_prediction(
        &self,
        predictions: &HashMap<String, ModelPrediction>,
        weights: &HashMap<String, f64>,
        reliability_scores: &HashMap<String, f64>,
    ) -> Result<(f64, HashMap<String, ModelContribution>)> {
        let mut weighted_sum = 0.0_f64;
        let mut total_weight = 0.0_f64;
        let mut model_contributions = HashMap::new();

        for (model_name, prediction) in predictions {
            let base_weight = weights.get(model_name).copied().unwrap_or(0.0_f64);
            let reliability = reliability_scores
                .get(model_name)
                .copied()
                .unwrap_or(0.5_f64);

            let final_weight = if self.config.weight_by_reliability {
                base_weight * reliability
            } else {
                base_weight
            };

            let weighted_contribution = prediction.value * final_weight;
            weighted_sum += weighted_contribution;
            total_weight += final_weight;

            model_contributions.insert(
                model_name.clone(),
                ModelContribution {
                    prediction: prediction.value,
                    confidence: prediction.confidence,
                    reliability,
                    weight: final_weight,
                    weighted_contribution,
                    prediction_interval: None,
                },
            );
        }

        if total_weight == 0.0_f64 {
            anyhow::bail!("Total weight is zero — cannot compute ensemble prediction");
        }

        let ensemble_prediction = weighted_sum / total_weight;
        Ok((ensemble_prediction, model_contributions))
    }

    fn filter_outliers(
        &self,
        predictions: HashMap<String, ModelPrediction>,
    ) -> Result<HashMap<String, ModelPrediction>> {
        if predictions.len() < 3 {
            return Ok(predictions);
        }

        let values: Vec<f64> = predictions.values().map(|p| p.value).collect();
        let mean = values.iter().sum::<f64>() / values.len() as f64;
        let variance =
            values.iter().map(|v| (v - mean).powi(2)).sum::<f64>() / values.len() as f64;
        let std_dev = variance.sqrt();
        let threshold = self.config.outlier_threshold * std_dev;

        let mut filtered = HashMap::new();
        for (model_name, prediction) in &predictions {
            if (prediction.value - mean).abs() <= threshold {
                filtered.insert(model_name.clone(), prediction.clone());
            } else {
                warn!(
                    "Filtered outlier prediction from {}: {}",
                    model_name, prediction.value
                );
            }
        }

        if filtered.is_empty() {
            warn!("All predictions filtered as outliers, using original set");
            Ok(predictions)
        } else {
            Ok(filtered)
        }
    }

    fn calculate_uncertainty_bounds(
        &self,
        prediction: f64,
        uncertainty: &UncertaintyDecomposition,
    ) -> (f64, f64) {
        let uncertainty_width = 2.0_f64 * uncertainty.total;
        (prediction - uncertainty_width, prediction + uncertainty_width)
    }

    fn calculate_ensemble_reliability(
        &self,
        reliability_scores: &HashMap<String, f64>,
        weights: &HashMap<String, f64>,
    ) -> f64 {
        let mut weighted_reliability = 0.0_f64;
        let mut total_weight = 0.0_f64;

        for (model_name, &reliability) in reliability_scores {
            if let Some(&weight) = weights.get(model_name) {
                weighted_reliability += reliability * weight;
                total_weight += weight;
            }
        }

        if total_weight > 0.0_f64 {
            weighted_reliability / total_weight
        } else {
            0.5
        }
    }

    fn compute_ensemble_confidence(
        &self,
        uncertainty: &UncertaintyDecomposition,
        reliability: f64,
    ) -> f64 {
        let uncertainty_factor = 1.0_f64 - (uncertainty.total / (1.0_f64 + uncertainty.total));
        (uncertainty_factor * reliability).clamp(0.0_f64, 1.0_f64)
    }

    fn update_disagreement_tracking(&mut self, predictions: &HashMap<String, ModelPrediction>) {
        if predictions.len() < 2 {
            return;
        }

        let values: Vec<f64> = predictions.values().map(|p| p.value).collect();
        let mean = values.iter().sum::<f64>() / values.len() as f64;
        let spread = values
            .iter()
            .map(|v| (v - mean).abs())
            .fold(0.0_f64, f64::max);
        let disagreement_magnitude = spread / mean.abs().max(1e-6);

        let record = DisagreementRecord {
            timestamp: chrono::Utc::now(),
            magnitude: disagreement_magnitude,
            models: predictions.keys().cloned().collect(),
            prediction_spread: spread,
        };

        self.uncertainty_quantifier
            .disagreement_tracker
            .add_record(record);

        if disagreement_magnitude
            > self
                .uncertainty_quantifier
                .disagreement_tracker
                .warning_threshold
        {
            warn!(
                "High model disagreement detected: {:.3}",
                disagreement_magnitude
            );
        }
    }
}

// ── UncertaintyQuantifier implementation ──────────────────────────────────────

impl UncertaintyQuantifier {
    pub fn new(epistemic_config: EpistemicConfig, aleatoric_config: AleatoricConfig) -> Self {
        let disagreement_tracker = DisagreementTracker::new(1000_usize, 0.2_f64);
        Self {
            epistemic_config,
            aleatoric_config,
            disagreement_tracker,
        }
    }

    /// Quantify total uncertainty by decomposing into three components.
    pub fn quantify_uncertainty(
        &self,
        predictions: &HashMap<String, ModelPrediction>,
        weights: &HashMap<String, f64>,
    ) -> Result<UncertaintyDecomposition> {
        let epistemic = self.calculate_epistemic_uncertainty(predictions, weights);
        let aleatoric = self.calculate_aleatoric_uncertainty(predictions);
        let disagreement = self.calculate_disagreement_uncertainty();

        let total = (epistemic.powi(2) + aleatoric.powi(2) + disagreement.powi(2)).sqrt();

        Ok(UncertaintyDecomposition {
            epistemic,
            aleatoric,
            disagreement,
            total,
        })
    }

    /// Epistemic uncertainty via model disagreement (standard deviation of predictions).
    fn calculate_epistemic_uncertainty(
        &self,
        predictions: &HashMap<String, ModelPrediction>,
        _weights: &HashMap<String, f64>,
    ) -> f64 {
        if predictions.len() < 2 || !self.epistemic_config.use_model_disagreement {
            return 0.0_f64;
        }

        let values: Vec<f64> = predictions.values().map(|p| p.value).collect();
        let mean = values.iter().sum::<f64>() / values.len() as f64;
        let variance =
            values.iter().map(|v| (v - mean).powi(2)).sum::<f64>() / values.len() as f64;
        variance.sqrt()
    }

    /// Aleatoric uncertainty from average model self-reported uncertainty (1 − confidence).
    fn calculate_aleatoric_uncertainty(
        &self,
        predictions: &HashMap<String, ModelPrediction>,
    ) -> f64 {
        let uncertainties: Vec<f64> = predictions
            .values()
            .map(|p| 1.0_f64 - p.confidence)
            .collect();

        if uncertainties.is_empty() {
            return 0.0_f64;
        }

        uncertainties.iter().sum::<f64>() / uncertainties.len() as f64
    }

    /// Historical disagreement component from recent tracking window.
    fn calculate_disagreement_uncertainty(&self) -> f64 {
        self.disagreement_tracker.get_recent_disagreement()
    }
}

// ── ReliabilityScorer implementation ──────────────────────────────────────────

impl ReliabilityScorer {
    pub fn new(decay_factor: f64, min_observations: usize) -> Self {
        Self {
            reliability_history: HashMap::new(),
            decay_factor,
            min_observations,
        }
    }

    /// Record a reliability observation for a model.
    pub fn update_reliability(&mut self, model_name: String, record: ReliabilityRecord) {
        let history = self
            .reliability_history
            .entry(model_name.clone())
            .or_default();
        history.push(record);

        // Cap history at 1000 entries
        if history.len() > 1000 {
            history.remove(0);
        }

        debug!(
            "Updated reliability history for {}: {} records",
            model_name,
            history.len()
        );
    }

    /// Get current reliability scores for all tracked models.
    pub fn get_current_scores(&self) -> HashMap<String, f64> {
        self.reliability_history
            .iter()
            .map(|(name, history)| (name.clone(), self.calculate_model_reliability(history)))
            .collect()
    }

    /// Get reliability for a specific model (0.5 default for unknown models).
    pub fn get_model_reliability(&self, model_name: &str) -> f64 {
        self.reliability_history
            .get(model_name)
            .map(|history| self.calculate_model_reliability(history))
            .unwrap_or(0.5_f64)
    }

    fn calculate_model_reliability(&self, history: &[ReliabilityRecord]) -> f64 {
        if history.len() < self.min_observations {
            return 0.5_f64;
        }

        let mut weighted_sum = 0.0_f64;
        let mut weight_sum = 0.0_f64;
        let current_time = chrono::Utc::now();

        for (i, record) in history.iter().rev().enumerate() {
            let age = (current_time - record.timestamp).num_hours() as f64;
            let weight = self.decay_factor.powf(age / 24.0_f64);

            let reliability_score =
                (record.accuracy + record.calibration + record.confidence_reliability) / 3.0_f64;
            weighted_sum += reliability_score * weight;
            weight_sum += weight;

            if i >= 100_usize {
                break;
            }
        }

        if weight_sum > 0.0_f64 {
            (weighted_sum / weight_sum).clamp(0.0_f64, 1.0_f64)
        } else {
            0.5_f64
        }
    }
}

// ── IntervalCombiner implementation ───────────────────────────────────────────

impl IntervalCombiner {
    pub fn new(
        _combination_method: CombinationMethod,
        confidence_levels: Vec<f64>,
        _calibration_params: CalibrationParams,
    ) -> Self {
        Self { confidence_levels }
    }

    /// Compute prediction intervals at each configured confidence level.
    pub fn combine_intervals(
        &self,
        predictions: &HashMap<String, ModelPrediction>,
        weights: &HashMap<String, f64>,
    ) -> Result<Vec<PredictionInterval>> {
        let mut intervals = Vec::new();
        for &confidence_level in &self.confidence_levels {
            let interval =
                self.compute_combined_interval(predictions, weights, confidence_level)?;
            intervals.push(interval);
        }
        Ok(intervals)
    }

    fn compute_combined_interval(
        &self,
        predictions: &HashMap<String, ModelPrediction>,
        _weights: &HashMap<String, f64>,
        confidence_level: f64,
    ) -> Result<PredictionInterval> {
        let values: Vec<f64> = predictions.values().map(|p| p.value).collect();
        let mean = values.iter().sum::<f64>() / values.len() as f64;
        let variance =
            values.iter().map(|v| (v - mean).powi(2)).sum::<f64>() / values.len() as f64;
        let std_dev = variance.sqrt();

        // Normal approximation z-scores
        let z_score = match confidence_level {
            x if x >= 0.99_f64 => 2.576_f64,
            x if x >= 0.95_f64 => 1.96_f64,
            x if x >= 0.90_f64 => 1.645_f64,
            x if x >= 0.68_f64 => 1.0_f64,
            _ => 1.96_f64,
        };

        let margin = z_score * std_dev;
        let lower_bound = mean - margin;
        let upper_bound = mean + margin;

        Ok(PredictionInterval {
            confidence_level,
            lower_bound,
            upper_bound,
            width: upper_bound - lower_bound,
        })
    }
}

// ── DisagreementTracker implementation ────────────────────────────────────────

impl DisagreementTracker {
    pub fn new(max_history_length: usize, warning_threshold: f64) -> Self {
        Self {
            disagreement_history: Vec::new(),
            max_history_length,
            warning_threshold,
        }
    }

    pub fn add_record(&mut self, record: DisagreementRecord) {
        self.disagreement_history.push(record);
        if self.disagreement_history.len() > self.max_history_length {
            self.disagreement_history.remove(0);
        }
    }

    /// Exponentially-weighted average of last 20 disagreement records.
    pub fn get_recent_disagreement(&self) -> f64 {
        if self.disagreement_history.is_empty() {
            return 0.0_f64;
        }

        let mut weighted_sum = 0.0_f64;
        let mut weight_sum = 0.0_f64;

        for (i, record) in self
            .disagreement_history
            .iter()
            .rev()
            .take(20_usize)
            .enumerate()
        {
            let weight = 0.9_f64.powi(i as i32);
            weighted_sum += record.magnitude * weight;
            weight_sum += weight;
        }

        if weight_sum > 0.0_f64 {
            weighted_sum / weight_sum
        } else {
            0.0_f64
        }
    }
}

// ── Default implementations ───────────────────────────────────────────────────

impl Default for AggregationConfig {
    fn default() -> Self {
        Self {
            weight_by_reliability: true,
            min_confidence_threshold: 0.1_f64,
            max_uncertainty_threshold: 1.0_f64,
            outlier_detection_enabled: true,
            outlier_threshold: 2.5_f64,
        }
    }
}

impl Default for EpistemicConfig {
    fn default() -> Self {
        Self {
            use_model_disagreement: true,
            bayesian_estimation: true,
            mc_dropout_samples: 100_usize,
        }
    }
}

impl Default for AleatoricConfig {
    fn default() -> Self {
        Self {
            heteroscedastic_noise: true,
            variance_estimation_method: VarianceEstimationMethod::EWMA,
            volatility_window: 50_usize,
        }
    }
}

impl Default for CalibrationParams {
    fn default() -> Self {
        Self {
            temperature: 1.0_f64,
            platt_scaling: false,
            isotonic_regression: false,
        }
    }
}

// ── Tests ─────────────────────────────────────────────────────────────────────

#[cfg(test)]
#[allow(clippy::assertions_on_result_states)]
mod tests {
    use super::*;

    fn make_prediction(model_id: &str, value: f64, confidence: f64) -> ModelPrediction {
        ModelPrediction::new(model_id.to_owned(), value, confidence)
    }

    #[test]
    fn test_confidence_aggregator_creation() {
        let config = AggregationConfig::default();
        let aggregator = ConfidenceAggregator::new(config);
        assert!(aggregator.config.weight_by_reliability);
    }

    #[test]
    fn test_uncertainty_quantification() {
        let mut aggregator = ConfidenceAggregator::new(AggregationConfig::default());

        let mut predictions = HashMap::new();
        predictions.insert("model1".to_owned(), make_prediction("model1", 0.5, 0.8));
        predictions.insert("model2".to_owned(), make_prediction("model2", 0.6, 0.7));

        let mut weights = HashMap::new();
        weights.insert("model1".to_owned(), 0.6_f64);
        weights.insert("model2".to_owned(), 0.4_f64);

        let result = aggregator.aggregate_with_uncertainty(predictions, &weights);
        assert!(result.is_ok());

        let ensemble_pred = result.expect("checked is_ok above");
        assert!(ensemble_pred.confidence >= 0.0_f64 && ensemble_pred.confidence <= 1.0_f64);
        assert!(ensemble_pred.uncertainty_decomposition.total >= 0.0_f64);
        assert_eq!(ensemble_pred.prediction_intervals.len(), 3);
    }

    #[test]
    fn test_disagreement_tracker() {
        let mut tracker = DisagreementTracker::new(100_usize, 0.2_f64);

        let record = DisagreementRecord {
            timestamp: chrono::Utc::now(),
            magnitude: 0.15_f64,
            models: vec!["model1".to_owned(), "model2".to_owned()],
            prediction_spread: 0.1_f64,
        };

        tracker.add_record(record);
        let disagreement = tracker.get_recent_disagreement();
        assert_eq!(disagreement, 0.15_f64);
    }

    #[test]
    fn test_reliability_scorer() {
        let mut scorer = ReliabilityScorer::new(0.95_f64, 5_usize);

        let record = ReliabilityRecord {
            timestamp: chrono::Utc::now(),
            accuracy: 0.8_f64,
            calibration: 0.75_f64,
            confidence_reliability: 0.85_f64,
            actual_outcome: None,
        };

        scorer.update_reliability("test_model".to_owned(), record);
        let reliability = scorer.get_model_reliability("test_model");
        assert!((0.0_f64..=1.0_f64).contains(&reliability));
    }

    #[test]
    fn test_single_model_aggregation() {
        let mut aggregator = ConfidenceAggregator::new(AggregationConfig::default());

        let mut predictions = HashMap::new();
        predictions.insert("solo".to_owned(), make_prediction("solo", 1.23, 0.9));

        let mut weights = HashMap::new();
        weights.insert("solo".to_owned(), 1.0_f64);

        let result = aggregator
            .aggregate_with_uncertainty(predictions, &weights)
            .expect("single model should succeed");

        // Single model → prediction equals that model's value (no outlier filtering below 3)
        assert!((result.prediction - 1.23).abs() < 1e-9);
        assert!(result.confidence > 0.0);
    }

    #[test]
    fn test_empty_weights_returns_error() {
        let mut aggregator = ConfidenceAggregator::new(AggregationConfig::default());

        let mut predictions = HashMap::new();
        predictions.insert("m1".to_owned(), make_prediction("m1", 0.5, 0.8));

        // No weights → effective weight is 0.0 after reliability multiplication
        let weights: HashMap<String, f64> = HashMap::new();
        let result = aggregator.aggregate_with_uncertainty(predictions, &weights);
        assert!(result.is_err());
    }
}