antecedent_validate/

bayesian_checks.rs

//! Prior/posterior predictive checks and prior sensitivity.
//!
//! SPDX-License-Identifier: MIT OR Apache-2.0

#![allow(
    clippy::cast_precision_loss,
    clippy::cast_possible_truncation,
    clippy::needless_range_loop,
    clippy::too_many_lines,
    clippy::many_single_char_names
)]

use std::sync::Arc;

use antecedent_core::{CausalRng, ExecutionContext, KernelPolicy};
use antecedent_estimate::{
    BayesianGCompWorkspace, BayesianGComputationAte, CausalPosterior, PreparedBayesianProblem,
};
use antecedent_identify::IdentificationStatus;
use antecedent_kernels::{PosteriorReduceOp, reduce_posterior_draws, standard_normal};
use antecedent_prob::{
    ExternalPriorSource, HessianFactorization, PriorSensitivitySummary, PriorSet,
    compose_external_priors_with_alphas,
};
use antecedent_stats::GlmFamily;

use crate::common::RefutationReport;
use crate::error::ValidationError;

/// Result of a prior or posterior predictive check.
///
/// Carries **two** discriminating axes so a model cannot pass merely by getting
/// the predictive mean right: (1) location, via the mean of `mean_y` over
/// simulations, and (2) dispersion, via the mean of the per-simulation
/// cross-observation SD of predicted values. A model whose predictive mean is
/// unbiased but whose predictive spread is badly wrong (e.g. off by 5×) fails
/// on the dispersion axis even though the location axis looks fine.
#[derive(Clone, Debug)]
pub struct PredictiveCheckReport {
    /// Check kind.
    pub kind: PredictiveCheckKind,
    /// Observed summary statistic (e.g. outcome mean).
    pub observed: f64,
    /// Mean of the predictive summary across simulations.
    pub predictive_mean: f64,
    /// SD of the predictive summary.
    pub predictive_sd: f64,
    /// Two-sided tail probability of `observed` under the predictive distribution.
    pub p_value: f64,
    /// Observed dispersion statistic: sample SD of the outcome across rows.
    pub observed_dispersion: f64,
    /// Mean, across simulations, of the per-simulation cross-observation SD of
    /// predicted values (the dispersion test statistic).
    pub predictive_dispersion_mean: f64,
    /// Two-sided Monte Carlo p-value of `observed_dispersion` under the
    /// simulated dispersion-statistic distribution.
    pub dispersion_p_value: f64,
    /// Number of predictive simulations.
    pub n_sims: u32,
}

impl PredictiveCheckReport {
    /// Convert to a suite [`RefutationReport`] using a two-sided α threshold on both
    /// the location (`p_value`) and dispersion (`dispersion_p_value`) axes.
    ///
    /// A single mean-only statistic cannot distinguish a well-calibrated predictive
    /// distribution from one with the right mean but the wrong spread (U/M-shaped
    /// misspecification on variance), so both axes must clear the threshold.
    #[must_use]
    pub fn to_refutation_report(&self, original_ate: f64, alpha: f64) -> RefutationReport {
        let name = match self.kind {
            PredictiveCheckKind::Prior => "prior_predictive",
            PredictiveCheckKind::Posterior => "posterior_predictive",
        };
        let mean_ok = self.p_value.is_finite() && self.p_value >= alpha;
        let dispersion_ok = self.dispersion_p_value.is_finite() && self.dispersion_p_value >= alpha;
        let passed = mean_ok && dispersion_ok;
        let comparison = self.p_value.min(self.dispersion_p_value);
        RefutationReport {
            refuter: Arc::from(name),
            original_ate,
            refuted_ate: self.predictive_mean,
            comparison,
            informative: true,
            passed,
            failure_condition: if passed {
                None
            } else if !mean_ok && !dispersion_ok {
                Some(Arc::from(format!(
                    "predictive check failed on mean (p={} < alpha={alpha}) and dispersion \
                     (p={} < alpha={alpha})",
                    self.p_value, self.dispersion_p_value
                )))
            } else if !mean_ok {
                Some(Arc::from(format!(
                    "predictive check failed (p={} < alpha={alpha})",
                    self.p_value
                )))
            } else {
                Some(Arc::from(format!(
                    "predictive dispersion check failed (p={} < alpha={alpha}); predictive spread \
                     does not match observed spread even though the mean matches",
                    self.dispersion_p_value
                )))
            },
            replicates: self.n_sims,
        }
    }
}

/// Prior vs posterior predictive.
#[derive(Clone, Copy, Debug, Eq, PartialEq, Hash)]
pub enum PredictiveCheckKind {
    /// Simulate from the prior predictive.
    Prior,
    /// Simulate from the posterior predictive.
    Posterior,
}

/// Prior predictive check using coefficient draws from a prior (no data update)
/// vs observed outcome mean.
#[derive(Clone, Debug)]
pub struct PriorPredictiveCheck {
    /// Simulations.
    pub n_sims: u32,
    /// RNG seed.
    pub seed: u64,
    /// Mean family (inverse link applied to η before summarizing).
    pub family: GlmFamily,
}

impl Default for PriorPredictiveCheck {
    fn default() -> Self {
        Self::new()
    }
}

impl PriorPredictiveCheck {
    /// Default 200 sims, Gaussian identity.
    #[must_use]
    pub fn new() -> Self {
        Self { n_sims: 200, seed: 0, family: GlmFamily::GaussianIdentity }
    }

    /// Run against a prepared Bayesian design with a weakly informative prior.
    ///
    /// Prefer [`Self::check_with_prior`] when an analysis / composed prior is known.
    ///
    /// # Errors
    ///
    /// Empty design.
    pub fn check(
        &self,
        problem: &PreparedBayesianProblem,
        ctx: &ExecutionContext,
    ) -> Result<PredictiveCheckReport, ValidationError> {
        let p = problem.design.ncols;
        let prior = PriorSet::weakly_informative(p);
        self.check_with_prior(problem, &prior, ctx)
    }

    /// Run prior predictive check under an explicit coefficient prior.
    ///
    /// # Errors
    ///
    /// Empty design or missing Gaussian coefficient prior.
    pub fn check_with_prior(
        &self,
        problem: &PreparedBayesianProblem,
        prior: &PriorSet,
        ctx: &ExecutionContext,
    ) -> Result<PredictiveCheckReport, ValidationError> {
        let n = problem.design.nrows;
        let p = problem.design.ncols;
        if n == 0 || p == 0 {
            return Err(ValidationError::estimation_msg("empty design for PPC"));
        }
        let coef_prior = prior.gaussian_coefficients().ok_or_else(|| {
            ValidationError::estimation_msg("prior missing Gaussian coefficients for PPC")
        })?;
        if coef_prior.len() != p {
            return Err(ValidationError::estimation_msg(
                "prior coefficient dimension mismatch for PPC",
            ));
        }
        let mut rng = CausalRng::from_seed(self.seed);
        let mut mean_summaries = Vec::with_capacity(self.n_sims as usize);
        let mut disp_summaries = Vec::with_capacity(self.n_sims as usize);
        let mut beta = vec![0.0; p];
        let mut y_pred = vec![0.0; n];
        for _ in 0..self.n_sims {
            // Draw β ~ prior once per simulation, then μ_i = g^{-1}(x_i'β).
            for c in 0..p {
                beta[c] =
                    coef_prior.mean[c] + coef_prior.variance[c].sqrt() * standard_normal(&mut rng);
            }
            for r in 0..n {
                let mut eta = 0.0;
                for c in 0..p {
                    eta += problem.design.matrix[c * n + r] * beta[c];
                }
                y_pred[r] = self.family.mean_from_eta(eta);
            }
            push_mean_and_dispersion(
                &y_pred,
                ctx.kernel_policy,
                &mut mean_summaries,
                &mut disp_summaries,
            );
        }
        Ok(summarize_predictive_check(
            PredictiveCheckKind::Prior,
            &problem.design.outcome,
            ctx.kernel_policy,
            &mean_summaries,
            &disp_summaries,
            self.n_sims,
        ))
    }
}

/// Posterior predictive check: resample outcome means from posterior coefficient draws.
#[derive(Clone, Debug)]
pub struct PosteriorPredictiveCheck {
    /// Number of posterior draws to use (capped by available).
    pub n_sims: u32,
    /// Mean family (inverse link applied to η before summarizing).
    pub family: GlmFamily,
}

impl Default for PosteriorPredictiveCheck {
    fn default() -> Self {
        Self::new()
    }
}

impl PosteriorPredictiveCheck {
    /// Default Gaussian identity.
    #[must_use]
    pub fn new() -> Self {
        Self { n_sims: 200, family: GlmFamily::GaussianIdentity }
    }

    /// Check using a fitted [`CausalPosterior`] that includes coefficient columns.
    ///
    /// # Errors
    ///
    /// Missing coefficients / empty draws.
    pub fn check(
        &self,
        problem: &PreparedBayesianProblem,
        posterior: &CausalPosterior,
    ) -> Result<PredictiveCheckReport, ValidationError> {
        let n = problem.design.nrows;
        let p = problem.design.ncols;
        let n_draws = posterior.draws.n_draws.min(self.n_sims as usize);
        if n_draws == 0 {
            return Err(ValidationError::estimation_msg("no posterior draws for PPC"));
        }
        let policy = KernelPolicy::default_policy();
        let mut mean_summaries = Vec::with_capacity(n_draws);
        let mut disp_summaries = Vec::with_capacity(n_draws);
        let mut y_pred = vec![0.0; n];
        let mut beta = vec![0.0; p];
        for d in 0..n_draws {
            // Coefficients depend only on (d, c); hoist the fallible draw
            // accessor out of the row loop (n×p calls per draw → p).
            for (c, slot) in beta.iter_mut().enumerate() {
                *slot = posterior.draws.get(d, c).map_err(ValidationError::from)?;
            }
            for r in 0..n {
                let mut eta = 0.0;
                for (c, &b) in beta.iter().enumerate() {
                    let x = problem.design.matrix[c * n + r];
                    eta += x * b;
                }
                y_pred[r] = self.family.mean_from_eta(eta);
            }
            push_mean_and_dispersion(&y_pred, policy, &mut mean_summaries, &mut disp_summaries);
        }
        Ok(summarize_predictive_check(
            PredictiveCheckKind::Posterior,
            &problem.design.outcome,
            policy,
            &mean_summaries,
            &disp_summaries,
            n_draws as u32,
        ))
    }
}

/// Default max relative range of effect means across the prior-sensitivity grid.
pub const DEFAULT_MAX_RELATIVE_PRIOR_RANGE: f64 = 0.5;

/// Prior sensitivity grid: isotropic scales **or** external α multipliers.
#[derive(Clone, Debug)]
pub struct PriorSensitivity {
    /// Prior scales (σ of isotropic Gaussian coefficient prior). Empty in α mode.
    pub scales: Arc<[f64]>,
    /// Multipliers on post-conflict applied alphas. Empty in isotropic scale mode.
    pub alphas: Arc<[f64]>,
    /// Fail when `(max−min) / scale` exceeds this, where `scale` is
    /// `max(|means…|, |original_ate|, ε)`.
    pub max_relative_range: f64,
}

/// Inputs for external α-multiplier prior sensitivity.
#[derive(Clone, Copy, Debug)]
pub struct ExternalAlphaSensitivity<'a> {
    /// Hydrated external sources (same order as composition).
    pub sources: &'a [ExternalPriorSource],
    /// Post-conflict applied alphas (length must match `sources`).
    pub alphas_applied: &'a [f64],
}

impl Default for PriorSensitivity {
    fn default() -> Self {
        Self::standard_grid()
    }
}

impl PriorSensitivity {
    /// Standard isotropic grid `{0.5, 1, 2, 5, 10, 20}` with [`DEFAULT_MAX_RELATIVE_PRIOR_RANGE`].
    #[must_use]
    pub fn standard_grid() -> Self {
        Self {
            scales: Arc::from(vec![0.5, 1.0, 2.0, 5.0, 10.0, 20.0]),
            alphas: Arc::from([]),
            max_relative_range: DEFAULT_MAX_RELATIVE_PRIOR_RANGE,
        }
    }

    /// Standard external-α multiplier grid `{0, 0.25, 0.5, 0.75, 1}`.
    ///
    /// Multiplier `0` is baseline-only; `1` uses full post-conflict applied alphas.
    #[must_use]
    pub fn standard_alpha_grid() -> Self {
        Self {
            scales: Arc::from([]),
            alphas: Arc::from(vec![0.0, 0.25, 0.5, 0.75, 1.0]),
            max_relative_range: DEFAULT_MAX_RELATIVE_PRIOR_RANGE,
        }
    }

    fn grid_len(&self) -> usize {
        if self.alphas.is_empty() { self.scales.len() } else { self.alphas.len() }
    }

    /// Refit Bayesian g-comp at each prior scale; return sensitivity summary.
    ///
    /// # Errors
    ///
    /// Fit failures or empty scale grid.
    pub fn evaluate(
        &self,
        estimator: &BayesianGComputationAte,
        problem: &PreparedBayesianProblem,
        identification: IdentificationStatus,
        workspace: &mut BayesianGCompWorkspace,
        ctx: &ExecutionContext,
    ) -> Result<(PriorSensitivitySummary, Vec<CausalPosterior>), ValidationError> {
        if self.scales.is_empty() {
            return Err(ValidationError::estimation_msg(
                "prior sensitivity scale grid is empty (use evaluate_external_alpha for α mode)",
            ));
        }
        let mut means = Vec::with_capacity(self.scales.len());
        let mut sds = Vec::with_capacity(self.scales.len());
        let mut posts = Vec::with_capacity(self.scales.len());
        for &scale in self.scales.iter() {
            let est = BayesianGComputationAte {
                prior_scale: scale,
                n_draws: estimator.n_draws.min(200),
                seed: estimator.seed,
                backend: estimator.backend,
                likelihood: estimator.likelihood,
                overlap: estimator.overlap,
                prior: None,
            };
            let post = est.fit(problem, identification, workspace, ctx).map_err(|e| {
                ValidationError::estimation_msg(format!("prior sensitivity fit failed: {e}"))
            })?;
            let eq = post.effect_column().ok_or_else(|| {
                ValidationError::estimation_msg("missing effect column in sensitivity fit")
            })?;
            means.push(post.summaries.mean[eq]);
            sds.push(post.summaries.sd[eq]);
            posts.push(post);
        }
        Ok((
            PriorSensitivitySummary {
                prior_scales: Arc::clone(&self.scales),
                alphas: Arc::from([]),
                effect_means: Arc::from(means),
                effect_sds: Arc::from(sds),
            },
            posts,
        ))
    }

    /// Refit at each α-multiplier on post-conflict applied alphas (external prior bank).
    ///
    /// For multiplier `m`, composed alphas are `m * alphas_applied[k]` (clamped to `[0, 1]`).
    ///
    /// # Errors
    ///
    /// Empty α grid, length mismatch, compose failures, or fit failures.
    pub fn evaluate_external_alpha(
        &self,
        estimator: &BayesianGComputationAte,
        problem: &PreparedBayesianProblem,
        identification: IdentificationStatus,
        workspace: &mut BayesianGCompWorkspace,
        ctx: &ExecutionContext,
        external: ExternalAlphaSensitivity<'_>,
    ) -> Result<(PriorSensitivitySummary, Vec<CausalPosterior>), ValidationError> {
        if self.alphas.is_empty() {
            return Err(ValidationError::estimation_msg("prior sensitivity alpha grid is empty"));
        }
        if external.sources.len() != external.alphas_applied.len() {
            return Err(ValidationError::estimation_msg(
                "evaluate_external_alpha: sources / alphas_applied length mismatch",
            ));
        }
        let n_coef = problem.design.ncols;
        let baseline = PriorSet::weakly_informative(n_coef);
        let requested: Vec<f64> = external.sources.iter().map(|s| s.weight.alpha).collect();
        let mut means = Vec::with_capacity(self.alphas.len());
        let mut sds = Vec::with_capacity(self.alphas.len());
        let mut posts = Vec::with_capacity(self.alphas.len());
        for &mult in self.alphas.iter() {
            if !mult.is_finite() || !(0.0..=1.0).contains(&mult) {
                return Err(ValidationError::estimation_msg(
                    "prior sensitivity alpha multiplier must be finite and in [0, 1]",
                ));
            }
            let scaled: Vec<f64> =
                external.alphas_applied.iter().map(|&a| (a * mult).clamp(0.0, 1.0)).collect();
            let composed = compose_external_priors_with_alphas(
                external.sources,
                &requested,
                &scaled,
                &baseline,
            )
            .map_err(|e| {
                ValidationError::estimation_msg(format!("prior sensitivity compose failed: {e}"))
            })?;
            let est = BayesianGComputationAte {
                prior_scale: estimator.prior_scale,
                n_draws: estimator.n_draws.min(200),
                seed: estimator.seed,
                backend: estimator.backend,
                likelihood: estimator.likelihood,
                overlap: estimator.overlap,
                prior: Some(composed.prior),
            };
            let post = est.fit(problem, identification, workspace, ctx).map_err(|e| {
                ValidationError::estimation_msg(format!("prior sensitivity α fit failed: {e}"))
            })?;
            let eq = post.effect_column().ok_or_else(|| {
                ValidationError::estimation_msg("missing effect column in α sensitivity fit")
            })?;
            means.push(post.summaries.mean[eq]);
            sds.push(post.summaries.sd[eq]);
            posts.push(post);
        }
        Ok((
            PriorSensitivitySummary {
                prior_scales: Arc::from([]),
                alphas: Arc::clone(&self.alphas),
                effect_means: Arc::from(means),
                effect_sds: Arc::from(sds),
            },
            posts,
        ))
    }

    /// Convert sensitivity range into a refutation-style report.
    ///
    /// Passes when the relative range of effect means is finite and
    /// `≤ max_relative_range`.
    #[must_use]
    pub fn to_report(
        &self,
        summary: &PriorSensitivitySummary,
        original_ate: f64,
    ) -> RefutationReport {
        let min = summary.effect_means.iter().copied().fold(f64::INFINITY, f64::min);
        let max = summary.effect_means.iter().copied().fold(f64::NEG_INFINITY, f64::max);
        let range = max - min;
        let denom = summary
            .effect_means
            .iter()
            .copied()
            .map(f64::abs)
            .fold(original_ate.abs(), f64::max)
            .max(1e-8);
        let relative = range / denom;
        let passed = relative.is_finite() && relative <= self.max_relative_range;
        let kind =
            if summary.alphas.is_empty() { "prior_sensitivity" } else { "prior_sensitivity_alpha" };
        RefutationReport {
            refuter: Arc::from(kind),
            original_ate,
            refuted_ate: summary.effect_means.last().copied().unwrap_or(original_ate),
            comparison: relative,
            informative: true,
            passed,
            failure_condition: if passed {
                None
            } else {
                Some(Arc::from(format!(
                    "prior sensitivity relative range {relative} exceeds max {}",
                    self.max_relative_range
                )))
            },
            replicates: u32::try_from(self.grid_len()).unwrap_or(u32::MAX),
        }
    }
}

fn summarize_check(
    kind: PredictiveCheckKind,
    observed: f64,
    summaries: &[f64],
    n_sims: u32,
) -> PredictiveCheckReport {
    let policy = KernelPolicy::default_policy();
    let mean = reduce_posterior_draws(summaries, PosteriorReduceOp::Mean, &policy).unwrap_or(0.0);
    let sd = reduce_posterior_draws(summaries, PosteriorReduceOp::Std, &policy).unwrap_or(0.0);
    let n = summaries.len() as f64;
    // (1 + count) / (1 + n) form (Davison & Hinkley): an exact-zero Monte Carlo
    // p-value is never valid evidence with a finite sample, so both tails are
    // bounded below by 1/(n+1) and the two-sided p-value by 2/(n+1).
    let below = summaries.iter().filter(|&&x| x <= observed).count() as f64;
    let above = summaries.iter().filter(|&&x| x >= observed).count() as f64;
    let p_lower = (1.0 + below) / (1.0 + n);
    let p_upper = (1.0 + above) / (1.0 + n);
    let p = (2.0 * p_lower.min(p_upper)).min(1.0);
    PredictiveCheckReport {
        kind,
        observed,
        predictive_mean: mean,
        predictive_sd: sd,
        p_value: p,
        // Populated by [`summarize_predictive_check`], the two-axis wrapper around this
        // function; callers of `summarize_check` directly (unit tests exercising the
        // Monte Carlo p-value formula) only care about the location axis above.
        observed_dispersion: 0.0,
        predictive_dispersion_mean: 0.0,
        dispersion_p_value: 1.0,
        n_sims,
    }
}

/// Push the per-simulation location (mean) and dispersion (cross-observation SD)
/// summary statistics for one simulated/predicted row vector `y_pred`.
///
/// Dispersion is the sample SD of `y_pred` *across observations within a single
/// simulation* — how spread the predicted values are over the design — which is
/// exactly the axis a mean-only PPC statistic is blind to (see D3 / defect C:
/// a model with an unbiased predictive mean but a badly wrong predictive spread
/// must be caught here, not just on the mean axis).
fn push_mean_and_dispersion(
    y_pred: &[f64],
    policy: KernelPolicy,
    mean_summaries: &mut Vec<f64>,
    disp_summaries: &mut Vec<f64>,
) {
    let n = y_pred.len().max(1) as f64;
    let mean_y = y_pred.iter().sum::<f64>() / n;
    let sd_y = reduce_posterior_draws(y_pred, PosteriorReduceOp::Std, &policy).unwrap_or(0.0);
    mean_summaries.push(mean_y);
    disp_summaries.push(sd_y);
}

/// Two-axis predictive check: combines a location (mean) Monte Carlo check with a
/// dispersion (cross-observation SD) Monte Carlo check via [`summarize_check`], and
/// merges both into a single [`PredictiveCheckReport`].
fn summarize_predictive_check(
    kind: PredictiveCheckKind,
    outcome: &[f64],
    policy: KernelPolicy,
    mean_summaries: &[f64],
    disp_summaries: &[f64],
    n_sims: u32,
) -> PredictiveCheckReport {
    let n = outcome.len().max(1) as f64;
    let observed_mean = outcome.iter().sum::<f64>() / n;
    let observed_dispersion =
        reduce_posterior_draws(outcome, PosteriorReduceOp::Std, &policy).unwrap_or(0.0);
    let mean_report = summarize_check(kind, observed_mean, mean_summaries, n_sims);
    let disp_report = summarize_check(kind, observed_dispersion, disp_summaries, n_sims);
    PredictiveCheckReport {
        kind,
        observed: mean_report.observed,
        predictive_mean: mean_report.predictive_mean,
        predictive_sd: mean_report.predictive_sd,
        p_value: mean_report.p_value,
        observed_dispersion,
        predictive_dispersion_mean: disp_report.predictive_mean,
        dispersion_p_value: disp_report.p_value,
        n_sims,
    }
}

/// Attach prior sensitivity onto a [`CausalPosterior`].
#[must_use]
pub fn with_prior_sensitivity(
    mut posterior: CausalPosterior,
    summary: PriorSensitivitySummary,
) -> CausalPosterior {
    posterior.prior_sensitivity = Some(summary);
    posterior
}

#[cfg(test)]
mod tests {
    use super::*;
    use antecedent_core::{
        AverageEffectQuery, CausalSchemaBuilder, MeasurementSpec, RoleHint, SmallRoleSet,
        ValueType, VariableId,
    };
    use antecedent_data::{
        Float64Column, OwnedColumn, OwnedColumnarStorage, TabularData, ValidityBitmap,
    };
    use antecedent_estimate::{BayesianBackendKind, BayesianGComputationAte};
    use antecedent_expr::{ExprId, IdentifiedEstimand};
    use antecedent_identify::IdentificationStatus;
    use antecedent_prob::{ExternalPriorWeight, GaussianCoefficientPrior, PriorSpec};

    fn toy() -> (TabularData, IdentifiedEstimand, AverageEffectQuery) {
        let n = 60usize;
        let mut b = CausalSchemaBuilder::new();
        b.add_variable(
            "t",
            ValueType::Continuous,
            SmallRoleSet::from_hint(RoleHint::TreatmentCandidate),
            None,
            None,
            MeasurementSpec::default(),
        )
        .unwrap();
        b.add_variable(
            "y",
            ValueType::Continuous,
            SmallRoleSet::from_hint(RoleHint::OutcomeCandidate),
            None,
            None,
            MeasurementSpec::default(),
        )
        .unwrap();
        b.add_variable(
            "z",
            ValueType::Continuous,
            SmallRoleSet::from_hint(RoleHint::Context),
            None,
            None,
            MeasurementSpec::default(),
        )
        .unwrap();
        let schema = b.build().unwrap();
        let t: Vec<f64> = (0..n).map(|i| (i % 2) as f64).collect();
        let z: Vec<f64> = (0..n).map(|i| i as f64 * 0.05).collect();
        let y: Vec<f64> = (0..n).map(|i| 1.0 + 2.0 * t[i] + 0.3 * z[i]).collect();
        let cols = vec![
            OwnedColumn::Float64(
                Float64Column::new(
                    VariableId::from_raw(0),
                    Arc::from(t),
                    ValidityBitmap::all_valid(n),
                )
                .unwrap(),
            ),
            OwnedColumn::Float64(
                Float64Column::new(
                    VariableId::from_raw(1),
                    Arc::from(y),
                    ValidityBitmap::all_valid(n),
                )
                .unwrap(),
            ),
            OwnedColumn::Float64(
                Float64Column::new(
                    VariableId::from_raw(2),
                    Arc::from(z),
                    ValidityBitmap::all_valid(n),
                )
                .unwrap(),
            ),
        ];
        let storage = OwnedColumnarStorage::try_new(schema, cols, None, None).unwrap();
        let estimand = IdentifiedEstimand::backdoor(
            "backdoor.adjustment",
            Arc::from([VariableId::from_raw(2)]),
            ExprId::from_raw(0),
        );
        let query =
            AverageEffectQuery::binary_ate(VariableId::from_raw(0), VariableId::from_raw(1));
        (TabularData::new(storage), estimand, query)
    }

    #[test]
    fn prior_and_posterior_ppc_run() {
        let (data, estimand, query) = toy();
        let bayes = BayesianGComputationAte {
            backend: BayesianBackendKind::ConjugateGaussian,
            n_draws: 100,
            seed: 2,
            prior_scale: 10.0,
            ..BayesianGComputationAte::new()
        };
        let prep = bayes.prepare(&data, &estimand, &query).unwrap();
        let ctx = ExecutionContext::for_tests(1);
        let prior_rep = PriorPredictiveCheck { n_sims: 50, seed: 3, ..PriorPredictiveCheck::new() }
            .check(&prep, &ctx)
            .unwrap();
        assert_eq!(prior_rep.kind, PredictiveCheckKind::Prior);
        assert!(prior_rep.p_value.is_finite());

        let mut ws = BayesianGCompWorkspace::default();
        let post = bayes
            .fit(&prep, IdentificationStatus::NonparametricallyIdentified, &mut ws, &ctx)
            .unwrap();
        let post_rep = PosteriorPredictiveCheck { n_sims: 50, ..PosteriorPredictiveCheck::new() }
            .check(&prep, &post)
            .unwrap();
        assert_eq!(post_rep.kind, PredictiveCheckKind::Posterior);
    }

    #[test]
    fn summarize_check_observed_outside_range_never_reports_zero() {
        // Simulated draws are tightly clustered; an observation far outside the
        // range on either side must not collapse the Monte Carlo p-value to
        // exactly 0 (D1: below/n or above/n hitting 0 or 1 exactly).
        let n = 200usize;
        let summaries: Vec<f64> = (0..n).map(|i| i as f64 / n as f64).collect(); // [0, 1)
        let min_p = 2.0 / (n as f64 + 1.0);

        let low = summarize_check(PredictiveCheckKind::Posterior, -10.0, &summaries, n as u32);
        assert!(low.p_value > 0.0, "p_value must be strictly positive, got {}", low.p_value);
        assert!(low.p_value >= min_p, "p_value {} below the 2/(n+1) floor {min_p}", low.p_value);

        let high = summarize_check(PredictiveCheckKind::Posterior, 10.0, &summaries, n as u32);
        assert!(high.p_value > 0.0, "p_value must be strictly positive, got {}", high.p_value);
        assert!(high.p_value >= min_p, "p_value {} below the 2/(n+1) floor {min_p}", high.p_value);
    }

    #[test]
    fn summarize_check_observed_near_centre_gives_high_p_value() {
        // Sanity check that the corrected formula still behaves as expected in
        // the ordinary case: an observation near the middle of the simulated
        // distribution should give a p-value near 1, not just "not exactly 0".
        let n = 200usize;
        let summaries: Vec<f64> = (0..n).map(|i| i as f64 / n as f64).collect(); // [0, 1)
        let centre = summarize_check(PredictiveCheckKind::Posterior, 0.5, &summaries, n as u32);
        assert!(centre.p_value > 0.9, "expected p_value near 1, got {}", centre.p_value);
    }

    #[test]
    fn prior_sensitivity_grid() {
        let (data, estimand, query) = toy();
        let bayes = BayesianGComputationAte {
            backend: BayesianBackendKind::ConjugateGaussian,
            n_draws: 80,
            seed: 4,
            ..BayesianGComputationAte::new()
        };
        let prep = bayes.prepare(&data, &estimand, &query).unwrap();
        let mut ws = BayesianGCompWorkspace::default();
        let ctx = ExecutionContext::for_tests(1);
        let sens = PriorSensitivity {
            scales: Arc::from(vec![1.0, 10.0, 50.0]),
            alphas: Arc::from([]),
            max_relative_range: DEFAULT_MAX_RELATIVE_PRIOR_RANGE,
        };
        let (summary, posts) = sens
            .evaluate(
                &bayes,
                &prep,
                IdentificationStatus::NonparametricallyIdentified,
                &mut ws,
                &ctx,
            )
            .unwrap();
        assert_eq!(summary.prior_scales.len(), 3);
        assert!(summary.alphas.is_empty());
        assert_eq!(posts.len(), 3);
        let rep =
            sens.to_report(&summary, posts[0].summaries.mean[posts[0].effect_column().unwrap()]);
        assert!(rep.passed);
    }

    #[test]
    fn prior_sensitivity_external_alpha_pulls_toward_source() {
        let fixture: serde_json::Value = serde_json::from_str(include_str!(
            "../../../conformance/validate/bayesian_checks/expected.json"
        ))
        .unwrap();
        assert!(
            fixture["contracts"]["prior_sensitivity_full_trust_moves_toward_source"]
                .as_bool()
                .unwrap()
        );

        let (data, estimand, query) = toy();
        // Data ATE ≈ 2; bank a tight prior with treatment coef mean = 8.
        let bayes = BayesianGComputationAte {
            backend: BayesianBackendKind::ConjugateGaussian,
            n_draws: 120,
            seed: 7,
            ..BayesianGComputationAte::new()
        };
        let prep = bayes.prepare(&data, &estimand, &query).unwrap();
        let n = prep.design.ncols;
        let t_col = prep.design.treatment_column().expect("treatment column");
        let mut mean = vec![0.0; n];
        mean[t_col] = 8.0;
        let mut source_prior = PriorSet::new();
        source_prior.push(PriorSpec::GaussianCoefficients(GaussianCoefficientPrior {
            mean: Arc::from(mean),
            variance: Arc::from(vec![0.05; n]),
        }));
        let sources = [ExternalPriorSource {
            id: Arc::from("survey_a"),
            prior: source_prior,
            weight: ExternalPriorWeight::power(1.0).unwrap(),
            ess: None,
        }];
        let alphas_applied = [1.0_f64];
        let mut ws = BayesianGCompWorkspace::default();
        let ctx = ExecutionContext::for_tests(1);
        let sens = PriorSensitivity::standard_alpha_grid();
        let (summary, _) = sens
            .evaluate_external_alpha(
                &bayes,
                &prep,
                IdentificationStatus::NonparametricallyIdentified,
                &mut ws,
                &ctx,
                ExternalAlphaSensitivity { sources: &sources, alphas_applied: &alphas_applied },
            )
            .unwrap();
        assert_eq!(summary.alphas.len(), 5);
        assert!(summary.prior_scales.is_empty());
        assert!(summary.effect_means.iter().all(|m| m.is_finite()));
        let m0 = summary.effect_means[0];
        let m1 = *summary.effect_means.last().unwrap();
        // Full trust (m=1) should sit closer to the banked treatment mean than baseline (m=0).
        assert!(
            (m1 - 8.0).abs() < (m0 - 8.0).abs(),
            "m=1 mean {m1} should be closer to 8 than m=0 mean {m0}"
        );
        let rep = sens.to_report(&summary, m1);
        assert_eq!(rep.refuter.as_ref(), "prior_sensitivity_alpha");
        assert!(rep.informative);
        assert!(rep.comparison.is_finite() && rep.comparison > 0.0);
    }

    #[test]
    fn ppc_catches_variance_misspecification_mean_ok() {
        // Defect C regression: construct data whose outcome mean is (nearly)
        // constant across covariates but has large residual spread (SD ~5), so a
        // model whose predictive draws imply a small cross-observation SD gets the
        // *mean* right (mean-only PPC would pass) while badly understating the
        // *dispersion* of the outcome. The added dispersion axis must catch this
        // even though the location axis does not.
        let n = 400usize;
        let mut b = CausalSchemaBuilder::new();
        b.add_variable(
            "t",
            ValueType::Continuous,
            SmallRoleSet::from_hint(RoleHint::TreatmentCandidate),
            None,
            None,
            MeasurementSpec::default(),
        )
        .unwrap();
        b.add_variable(
            "y",
            ValueType::Continuous,
            SmallRoleSet::from_hint(RoleHint::OutcomeCandidate),
            None,
            None,
            MeasurementSpec::default(),
        )
        .unwrap();
        b.add_variable(
            "z",
            ValueType::Continuous,
            SmallRoleSet::from_hint(RoleHint::Context),
            None,
            None,
            MeasurementSpec::default(),
        )
        .unwrap();
        let schema = b.build().unwrap();
        // t, z are both (near-)constant so the fitted/prior mean function has almost
        // no cross-observation spread; y has a large, fixed residual (+/- 5) on top of
        // a tiny signal, so E[y] is well predicted but SD(y) is not.
        let mut rng = CausalRng::from_seed(99);
        let t: Vec<f64> = (0..n).map(|_| 0.0).collect();
        let z: Vec<f64> = (0..n).map(|_| 0.0).collect();
        let y: Vec<f64> = (0..n).map(|_| 5.0 * standard_normal(&mut rng)).collect();
        let cols = vec![
            OwnedColumn::Float64(
                Float64Column::new(
                    VariableId::from_raw(0),
                    Arc::from(t),
                    ValidityBitmap::all_valid(n),
                )
                .unwrap(),
            ),
            OwnedColumn::Float64(
                Float64Column::new(
                    VariableId::from_raw(1),
                    Arc::from(y),
                    ValidityBitmap::all_valid(n),
                )
                .unwrap(),
            ),
            OwnedColumn::Float64(
                Float64Column::new(
                    VariableId::from_raw(2),
                    Arc::from(z),
                    ValidityBitmap::all_valid(n),
                )
                .unwrap(),
            ),
        ];
        let storage = OwnedColumnarStorage::try_new(schema, cols, None, None).unwrap();
        let estimand = IdentifiedEstimand::backdoor(
            "backdoor.adjustment",
            Arc::from([VariableId::from_raw(2)]),
            ExprId::from_raw(0),
        );
        let query =
            AverageEffectQuery::binary_ate(VariableId::from_raw(0), VariableId::from_raw(1));
        let data = TabularData::new(storage);

        let bayes = BayesianGComputationAte {
            backend: BayesianBackendKind::ConjugateGaussian,
            n_draws: 300,
            seed: 5,
            // Tight prior around 0: the mean function is confidently ~0 everywhere,
            // matching E[y]~0, but says nothing spreads out — the mean axis passes.
            prior_scale: 0.05,
            ..BayesianGComputationAte::new()
        };
        let prep = bayes.prepare(&data, &estimand, &query).unwrap();
        let ctx = ExecutionContext::for_tests(1);

        let rep = PriorPredictiveCheck { n_sims: 300, seed: 6, ..PriorPredictiveCheck::new() }
            .check(&prep, &ctx)
            .unwrap();

        assert!(
            rep.p_value >= 0.05,
            "expected the mean axis to look fine (unbiased predictive mean), got p={}",
            rep.p_value
        );
        assert!(
            rep.dispersion_p_value < 0.05,
            "expected the dispersion axis to catch the 5x variance mismatch, got \
             dispersion_p_value={}",
            rep.dispersion_p_value
        );
        let refuted = rep.to_refutation_report(0.0, 0.05);
        assert!(
            !refuted.passed,
            "predictive check must fail overall when dispersion is badly wrong even \
             though the mean matches"
        );
    }

    #[test]
    fn sbc_to_report_gates_on_uniformity_not_just_mean() {
        // Defect B regression: a U-shaped rank histogram (ranks piled at the two
        // extremes, overdispersed posterior) is symmetric about the middle bin so its
        // mean rank fraction lands squarely in [0.35, 0.65] — a mean-only gate would
        // pass it. The χ² uniformity statistic must catch it.
        let n_reps = 200u32;
        let n_draws = 100usize;
        // Every replicate's rank is pinned at one of the two extremes (bins 0 and 9),
        // alternating — a textbook U shape.
        let ranks: Vec<u32> =
            (0..n_reps).map(|i| if i % 2 == 0 { 0 } else { n_draws as u32 }).collect();
        let n_d = n_draws as f64;
        let fracs: Vec<f64> = ranks.iter().map(|&r| f64::from(r) / n_d).collect();
        let mean_rank_frac = fracs.iter().sum::<f64>() / fracs.len() as f64;
        assert!(
            (0.35..=0.65).contains(&mean_rank_frac),
            "fixture sanity: mean rank frac {mean_rank_frac} should look fine on its own"
        );
        let bins = 10usize;
        let mut counts = vec![0.0; bins];
        for &r in &ranks {
            let b = ((u64::from(r) * bins as u64) / (n_draws as u64)).min(bins as u64 - 1) as usize;
            counts[b] += 1.0;
        }
        let expected = f64::from(n_reps) / bins as f64;
        let mut chi2 = 0.0;
        for c in counts {
            let d = c - expected;
            chi2 += d * d / expected.max(1.0);
        }
        assert!(
            chi2 > SBC_CHI2_CRITICAL_9DF_P99,
            "fixture sanity: U-shaped ranks should trip the χ² statistic, got {chi2}"
        );
        let report = SbcReport { ranks: Arc::from(ranks), mean_rank_frac, uniformity_stat: chi2 };
        let sbc = SimulationBasedCalibration { n_reps, n_draws, seed: 0 };
        let rep = sbc.to_report(&report, 2.0);
        assert!(
            !rep.passed,
            "SBC must fail on a U-shaped (non-uniform) rank distribution even though \
             mean_rank_frac={mean_rank_frac:.3} is in [0.35, 0.65]"
        );
    }

    /// Calibration-gate tests for defect A: [`SimulationBasedCalibration`] and
    /// [`PosteriorCalibrationOnSyntheticScm`] are fully implemented but (prior to this
    /// change) had zero call sites outside the `lib.rs` re-export — the Bayesian
    /// posterior had no coverage check at all. `#[ignore]`d for the same reason as
    /// `antecedent-estimate/src/calibration_coverage.rs`: many refits per test, too
    /// slow for every `cargo test`. Run via `cargo test -p antecedent-validate
    /// <name> -- --ignored --nocapture`.
    mod calibration_gate {
        use super::*;

        fn coverage_band(n_reps: u32, level: f64) -> (f64, f64) {
            let se = (level * (1.0 - level) / f64::from(n_reps)).sqrt();
            let lo = (level - 4.0 * se).max(0.5);
            let hi = (level + 4.0 * se).min(1.0);
            (lo, hi)
        }

        #[test]
        #[ignore = "calibration: run via scripts/gate_calibration.sh"]
        fn sbc_conjugate_gaussian_ranks_are_uniform() {
            let (data, estimand, query) = toy();
            let bayes = BayesianGComputationAte {
                backend: BayesianBackendKind::ConjugateGaussian,
                n_draws: 300,
                seed: 11,
                prior_scale: 5.0,
                ..BayesianGComputationAte::new()
            };
            let prep = bayes.prepare(&data, &estimand, &query).unwrap();
            let mut ws = BayesianGCompWorkspace::default();
            let ctx = ExecutionContext::for_tests(1);
            let sbc = SimulationBasedCalibration { n_reps: 200, n_draws: 300, seed: 42 };
            let report = sbc
                .check(
                    &bayes,
                    &prep,
                    IdentificationStatus::NonparametricallyIdentified,
                    &mut ws,
                    &ctx,
                )
                .unwrap();
            let rep = sbc.to_report(&report, 2.0);
            assert!(
                rep.passed,
                "SBC should pass for a correctly specified conjugate Gaussian model: \
                 mean_rank_frac={:.3} chi2={:.3}",
                report.mean_rank_frac, report.uniformity_stat
            );
            assert!(
                (0.35..=0.65).contains(&report.mean_rank_frac),
                "mean_rank_frac={:.3} outside [0.35, 0.65]",
                report.mean_rank_frac
            );
            assert!(
                report.uniformity_stat < SBC_CHI2_CRITICAL_9DF_P99,
                "chi2={:.3} exceeds critical value {SBC_CHI2_CRITICAL_9DF_P99:.3}",
                report.uniformity_stat
            );
        }

        #[test]
        #[ignore = "calibration: run via scripts/gate_calibration.sh"]
        fn posterior_calibration_synthetic_scm_nominal_90_coverage() {
            let (data, estimand, query) = toy();
            let bayes = BayesianGComputationAte {
                backend: BayesianBackendKind::ConjugateGaussian,
                n_draws: 300,
                seed: 21,
                prior_scale: 5.0,
                ..BayesianGComputationAte::new()
            };
            let prep = bayes.prepare(&data, &estimand, &query).unwrap();
            let mut ws = BayesianGCompWorkspace::default();
            let ctx = ExecutionContext::for_tests(1);
            let calib = PosteriorCalibrationOnSyntheticScm {
                n_reps: 200,
                n_draws: 300,
                level: 0.9,
                seed: 77,
            };
            let report = calib
                .check(
                    &bayes,
                    &prep,
                    IdentificationStatus::NonparametricallyIdentified,
                    &mut ws,
                    &ctx,
                )
                .unwrap();
            let (lo, hi) = coverage_band(report.n_reps, 0.9);
            assert!(
                report.coverage >= lo && report.coverage <= hi,
                "nominal 90% credible-interval coverage={:.3} outside [{:.3}, {:.3}] \
                 ({} reps); mean_abs_error={:.3}",
                report.coverage,
                lo,
                hi,
                report.n_reps,
                report.mean_abs_error
            );
        }
    }
}

/// MCMC chain diagnostics gate (ESS / R-hat / divergences).
///
/// Applicable only when the posterior was produced by an MCMC backend
/// (`InferenceDiagnostics::factorization == Mcmc`).
#[derive(Clone, Copy, Debug)]
pub struct McmcDiagnosticsCheck {
    /// Maximum acceptable split-Ř.
    pub max_rhat: f64,
    /// Minimum acceptable bulk ESS.
    pub min_ess: f64,
    /// Maximum acceptable divergence count.
    pub max_divergences: u32,
}

impl Default for McmcDiagnosticsCheck {
    fn default() -> Self {
        Self { max_rhat: 1.05, min_ess: 10.0, max_divergences: u32::MAX / 4 }
    }
}

impl McmcDiagnosticsCheck {
    /// Construct with defaults.
    #[must_use]
    pub fn new() -> Self {
        Self::default()
    }

    /// Evaluate against a fitted posterior's diagnostics.
    ///
    /// Returns `None` when the posterior is not MCMC (caller should emit `NotApplicable`).
    #[must_use]
    pub fn check(&self, posterior: &CausalPosterior) -> Option<RefutationReport> {
        let d = &posterior.diagnostics;
        if d.factorization != HessianFactorization::Mcmc {
            return None;
        }
        let rhat = d.rhat_max.unwrap_or(f64::INFINITY);
        let ess = d.ess_bulk_min.unwrap_or(0.0);
        let divs = d.n_divergences.unwrap_or(u32::MAX);
        let passed = rhat.is_finite()
            && rhat <= self.max_rhat
            && ess >= self.min_ess
            && divs <= self.max_divergences
            && d.allows_posterior();
        let ate = posterior
            .effect_column()
            .and_then(|c| posterior.summaries.mean.get(c).copied())
            .unwrap_or(f64::NAN);
        Some(RefutationReport {
            refuter: Arc::from("mcmc_diagnostics"),
            original_ate: ate,
            refuted_ate: ate,
            comparison: rhat,
            informative: true,
            passed,
            failure_condition: if passed {
                None
            } else {
                Some(Arc::from(format!(
                    "MCMC diagnostics failed: rhat={rhat:.4} ess={ess:.1} divergences={divs}"
                )))
            },
            replicates: d.n_chains.unwrap_or(0),
        })
    }
}

/// Simulation-based calibration ranks for a scalar posterior functional.
///
/// For each replicate: draw θ* from the prior predictive, simulate data, refit, and
/// record the rank of θ* among posterior draws of the primary effect.
#[derive(Clone, Debug)]
pub struct SimulationBasedCalibration {
    /// Number of SBC replicates.
    pub n_reps: u32,
    /// Draws per refit.
    pub n_draws: usize,
    /// RNG seed.
    pub seed: u64,
}

impl Default for SimulationBasedCalibration {
    fn default() -> Self {
        Self { n_reps: 50, n_draws: 100, seed: 0 }
    }
}

/// SBC report.
#[derive(Clone, Debug)]
pub struct SbcReport {
    /// Rank of the prior draw in each replicate (`0..=n_draws`).
    pub ranks: Arc<[u32]>,
    /// Mean rank / `n_draws` (≈ 0.5 when calibrated).
    pub mean_rank_frac: f64,
    /// Chi² uniformity diagnostic on coarse bins (lower is better).
    pub uniformity_stat: f64,
}

impl SimulationBasedCalibration {
    /// Construct.
    #[must_use]
    pub fn new(n_reps: u32) -> Self {
        Self { n_reps: n_reps.max(1), ..Self::default() }
    }

    /// Run SBC: draw θ from the prior, simulate `y` from the prior predictive under
    /// the fixed design matrix, refit the Bayesian g-computation estimator, and
    /// rank the true ATE among posterior effect draws.
    ///
    /// # Errors
    ///
    /// Fit failures.
    pub fn check(
        &self,
        estimator: &BayesianGComputationAte,
        problem: &PreparedBayesianProblem,
        identification: IdentificationStatus,
        workspace: &mut BayesianGCompWorkspace,
        ctx: &ExecutionContext,
    ) -> Result<SbcReport, ValidationError> {
        let mut rng = CausalRng::from_seed(self.seed);
        let n = problem.design.nrows;
        let p = problem.design.ncols;
        let t_col = problem
            .design
            .treatment_column()
            .ok_or_else(|| ValidationError::estimation_msg("SBC: missing treatment column"))?;
        let mut ranks = Vec::with_capacity(self.n_reps as usize);
        let mut est = estimator.clone();
        est.n_draws = self.n_draws;
        let scale = estimator.prior_scale.max(1e-6);

        for rep in 0..self.n_reps {
            let mut beta = vec![0.0; p];
            for c in 0..p {
                beta[c] = scale * standard_normal(&mut rng);
            }
            let true_effect = (problem.active - problem.control) * beta[t_col];
            let mut y_rep = vec![0.0; n];
            for r in 0..n {
                let mut eta = 0.0;
                for c in 0..p {
                    eta += problem.design.matrix[c * n + r] * beta[c];
                }
                y_rep[r] = eta + standard_normal(&mut rng);
            }
            let mut sim_problem = problem.clone();
            let mut design = sim_problem.design.clone();
            design.outcome = Arc::from(y_rep);
            sim_problem.design = design;
            est.seed = self.seed ^ (u64::from(rep).wrapping_mul(0x9E37));
            let post = est
                .fit(&sim_problem, identification, workspace, ctx)
                .map_err(|e| ValidationError::estimation_msg(format!("SBC refit failed: {e}")))?;
            let col = post
                .effect_column()
                .ok_or_else(|| ValidationError::estimation_msg("SBC: no effect column"))?;
            let draws = post
                .draws
                .column(col)
                .map_err(|e| ValidationError::estimation_msg(format!("SBC draws: {e}")))?;
            let mut rank = 0u32;
            for &d in draws {
                if d < true_effect {
                    rank += 1;
                }
            }
            ranks.push(rank);
        }

        let n_d = self.n_draws.max(1) as f64;
        let fracs: Vec<f64> = ranks.iter().map(|&r| f64::from(r) / n_d).collect();
        let mean_rank_frac =
            reduce_posterior_draws(&fracs, PosteriorReduceOp::Mean, &ctx.kernel_policy)
                .unwrap_or(0.5);
        let bins = 10usize;
        let mut counts = vec![0.0; bins];
        let n_draws_u = u64::try_from(self.n_draws.max(1)).unwrap_or(1);
        let bins_u = u64::try_from(bins).unwrap_or(1);
        for &r in &ranks {
            let b = usize::try_from(u64::from(r) * bins_u / n_draws_u).unwrap_or(0).min(bins - 1);
            counts[b] += 1.0;
        }
        let expected = f64::from(self.n_reps) / bins as f64;
        let mut chi2 = 0.0;
        for c in counts {
            let d = c - expected;
            chi2 += d * d / expected.max(1.0);
        }
        Ok(SbcReport { ranks: Arc::from(ranks), mean_rank_frac, uniformity_stat: chi2 })
    }

    /// Convert to a refutation report.
    ///
    /// Passes only when **both** hold:
    /// - the mean rank fraction is in `[0.35, 0.65]` (catches gross location bias), and
    /// - the χ² uniformity statistic over the 10 rank bins is below
    ///   `SBC_CHI2_CRITICAL_9DF_P99` (catches symmetric-about-0.5 U/M-shaped rank
    ///   distributions — overdispersed / underdispersed posteriors — that a mean-only
    ///   band cannot see because they average out to ≈0.5).
    #[must_use]
    pub fn to_report(&self, report: &SbcReport, original_ate: f64) -> RefutationReport {
        let mean_ok = (0.35..=0.65).contains(&report.mean_rank_frac);
        let uniform_ok = report.uniformity_stat.is_finite()
            && report.uniformity_stat <= SBC_CHI2_CRITICAL_9DF_P99;
        let passed = mean_ok && uniform_ok;
        RefutationReport {
            refuter: Arc::from("sbc"),
            original_ate,
            refuted_ate: report.mean_rank_frac,
            comparison: report.uniformity_stat,
            informative: true,
            passed,
            failure_condition: if passed {
                None
            } else if !mean_ok && !uniform_ok {
                Some(Arc::from(format!(
                    "SBC mean rank frac {:.3} outside [0.35, 0.65] and χ²={:.3} exceeds critical \
                     value {SBC_CHI2_CRITICAL_9DF_P99:.3} (9 df, p=0.99)",
                    report.mean_rank_frac, report.uniformity_stat
                )))
            } else if !mean_ok {
                Some(Arc::from(format!(
                    "SBC mean rank frac {:.3} outside [0.35, 0.65]",
                    report.mean_rank_frac
                )))
            } else {
                Some(Arc::from(format!(
                    "SBC rank distribution non-uniform: χ²={:.3} exceeds critical value \
                     {SBC_CHI2_CRITICAL_9DF_P99:.3} (9 df, p=0.99); mean rank frac {:.3} looked \
                     fine but ranks are not uniformly distributed (U- or M-shaped)",
                    report.uniformity_stat, report.mean_rank_frac
                )))
            },
            replicates: self.n_reps,
        }
    }
}

/// χ² critical value at the 0.99 quantile with 9 degrees of freedom (10 rank bins − 1).
///
/// Used to gate [`SimulationBasedCalibration::to_report`]'s uniformity check: SBC ranks
/// should be uniform under a well-calibrated posterior, and this is the standard
/// one-sided χ² goodness-of-fit critical value (e.g. Talts et al. 2018 §3 use the
/// same chi-square uniformity diagnostic).
const SBC_CHI2_CRITICAL_9DF_P99: f64 = 21.666;

/// Posterior calibration on synthetic SCMs: known-ATE credible-interval coverage.
#[derive(Clone, Debug)]
pub struct PosteriorCalibrationOnSyntheticScm {
    /// Monte Carlo replicates.
    pub n_reps: u32,
    /// Draws per fit.
    pub n_draws: usize,
    /// Nominal coverage level (e.g. 0.9).
    pub level: f64,
    /// RNG seed.
    pub seed: u64,
}

impl Default for PosteriorCalibrationOnSyntheticScm {
    fn default() -> Self {
        Self { n_reps: 40, n_draws: 100, level: 0.9, seed: 0 }
    }
}

/// Report for [`PosteriorCalibrationOnSyntheticScm`].
#[derive(Clone, Debug)]
pub struct PosteriorCalibrationReport {
    /// Empirical coverage of equal-tailed credible intervals.
    pub coverage: f64,
    /// Mean absolute error of posterior means vs true ATE.
    pub mean_abs_error: f64,
    /// Replicates.
    pub n_reps: u32,
}

impl PosteriorCalibrationOnSyntheticScm {
    /// Simulate known ATEs under the design, refit, and measure CI coverage.
    ///
    /// # Errors
    ///
    /// Fit failures.
    pub fn check(
        &self,
        estimator: &BayesianGComputationAte,
        problem: &PreparedBayesianProblem,
        identification: IdentificationStatus,
        workspace: &mut BayesianGCompWorkspace,
        ctx: &ExecutionContext,
    ) -> Result<PosteriorCalibrationReport, ValidationError> {
        let mut rng = CausalRng::from_seed(self.seed);
        let n = problem.design.nrows;
        let p = problem.design.ncols;
        let t_col = problem
            .design
            .treatment_column()
            .ok_or_else(|| ValidationError::estimation_msg("calibration: missing treatment"))?;
        let mut covered = 0u32;
        let mut abs_err = 0.0;
        let mut est = estimator.clone();
        est.n_draws = self.n_draws;
        let alpha = ((1.0 - self.level) / 2.0).clamp(0.0, 0.5);

        for rep in 0..self.n_reps {
            let true_ate = standard_normal(&mut rng);
            let mut beta = vec![0.0; p];
            let diff = problem.active - problem.control;
            beta[t_col] = if diff.abs() > 1e-12 { true_ate / diff } else { true_ate };
            for c in 0..p {
                if c != t_col {
                    beta[c] = 0.5 * standard_normal(&mut rng);
                }
            }
            let mut y = vec![0.0; n];
            for r in 0..n {
                let mut eta = 0.0;
                for c in 0..p {
                    eta += problem.design.matrix[c * n + r] * beta[c];
                }
                y[r] = eta + standard_normal(&mut rng);
            }
            let mut sim = problem.clone();
            let mut design = sim.design.clone();
            design.outcome = Arc::from(y);
            sim.design = design;
            est.seed = self.seed ^ (u64::from(rep).wrapping_mul(0xC2B2));
            let post = est
                .fit(&sim, identification, workspace, ctx)
                .map_err(|e| ValidationError::estimation_msg(format!("calibration refit: {e}")))?;
            let col = post
                .effect_column()
                .ok_or_else(|| ValidationError::estimation_msg("calibration: no effect"))?;
            let mut draws = post.draws.column(col).map_err(ValidationError::from)?.to_vec();
            draws.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
            let lo = quantile_sorted(&draws, alpha);
            let hi = quantile_sorted(&draws, 1.0 - alpha);
            let mean = reduce_posterior_draws(&draws, PosteriorReduceOp::Mean, &ctx.kernel_policy)
                .unwrap_or(0.0);
            abs_err += (mean - true_ate).abs();
            if true_ate >= lo && true_ate <= hi {
                covered += 1;
            }
        }
        Ok(PosteriorCalibrationReport {
            coverage: f64::from(covered) / f64::from(self.n_reps.max(1)),
            mean_abs_error: abs_err / f64::from(self.n_reps.max(1)),
            n_reps: self.n_reps,
        })
    }
}

fn quantile_sorted(sorted: &[f64], q: f64) -> f64 {
    if sorted.is_empty() {
        return 0.0;
    }
    let max_idx = sorted.len() - 1;
    let rank = (max_idx as f64 * q.clamp(0.0, 1.0)).round();
    let idx = (0..=max_idx)
        .min_by(|&a, &b| {
            (a as f64 - rank)
                .abs()
                .partial_cmp(&(b as f64 - rank).abs())
                .unwrap_or(std::cmp::Ordering::Equal)
        })
        .unwrap_or(0);
    sorted[idx]
}