Struct BernoulliMarginalSlopeFitResult 

pub struct BernoulliMarginalSlopeFitResult {

    pub fit: UnifiedFitResult,
    pub marginalspec_resolved: TermCollectionSpec,
    pub logslopespec_resolved: TermCollectionSpec,
    pub marginal_design: TermCollectionDesign,
    pub logslope_design: TermCollectionDesign,
    pub baseline_marginal: f64,
    pub baseline_logslope: f64,
    pub z_normalization: LatentZNormalization,
    pub latent_measure: LatentMeasureKind,
    pub score_warp_runtime: Option<DeviationRuntime>,
    pub link_dev_runtime: Option<DeviationRuntime>,
    pub gaussian_frailty_sd: Option<f64>,
    pub cross_block_warnings: Vec<CrossBlockIdentifiabilityWarning>,
    pub latent_z_rank_int_calibration: Option<LatentZRankIntCalibration>,
    pub latent_z_conditional_calibration: Option<LatentZConditionalCalibration>,
}

Fields

fit: UnifiedFitResult

marginalspec_resolved: TermCollectionSpec

logslopespec_resolved: TermCollectionSpec

marginal_design: TermCollectionDesign

logslope_design: TermCollectionDesign

baseline_marginal: f64

baseline_logslope: f64

z_normalization: LatentZNormalization

latent_measure: LatentMeasureKind

score_warp_runtime: Option<DeviationRuntime>

link_dev_runtime: Option<DeviationRuntime>

gaussian_frailty_sd: Option<f64>

Learned or fixed Gaussian-shift frailty SD. None = no frailty.

cross_block_warnings: Vec<CrossBlockIdentifiabilityWarning>

Structured warnings emitted during fit-time setup when a flex block was fully aliased by its anchor union and got dropped. The fit proceeds without the dropped block (its contribution to the joint design was numerically reproducible by the anchor span, so keeping it would leave the joint Hessian rank-deficient). Empty for fits where every flex block carried independent directions.

latent_z_rank_int_calibration: Option<LatentZRankIntCalibration>

Optional weighted rank inverse-normal (Blom rankit) calibration installed at fit time when the auto latent-z normality check failed. Some(_) ⇒ the training z was transformed in place via LatentZRankIntCalibration::apply_to_training before any downstream consumer (pooled probit baseline, term-collection designs, family PIRLS loops) saw it, and the rigid kernel routes through the standard-normal closed-form path on the calibrated scale. None ⇒ no calibration was applied (training z already passed the standard-normal diagnostics, or the caller explicitly selected a non-Auto LatentMeasureSpec).

Persisted to disk so prediction applies the same monotone map via LatentZRankIntCalibration::apply_at_predict to incoming z before the standard-normal kernel runs. The public field name is latent_z_rank_int_calibration — Agent D’s persistence pipeline reads it under that exact identifier.

latent_z_conditional_calibration: Option<LatentZConditionalCalibration>

Optional conditional location-scale calibration of the latent score (#905). Some(_) ⇒ the Auto path’s conditional E[z|C]/Var(z|C) Rao gate detected PC/grouping-dependence that the pooled-marginal gate cannot see, so the training z was replaced in place by ζ = (z − m(C))/√v(C) (via LatentZConditionalCalibration::apply) before any downstream consumer saw it. Mutually exclusive with latent_z_rank_int_calibration: rank-INT fixes a pooled-marginal defect, the conditional correction fixes a conditional-shift defect that rank-INT provably cannot. Persisted so prediction rebuilds a(C) from the (reproducible) marginal design and applies the identical map.