Struct OuterGradientFdRecord 

pub struct OuterGradientFdRecord {

    pub theta: Array1<f64>,
    pub rho_dim: usize,
    pub psi_dim: usize,
    pub cost: f64,
    pub analytic_psi_gradient: Array1<f64>,
    pub finite_difference_psi_gradient: Array1<f64>,
    pub psi_steps: Array1<f64>,
    pub psi_fd_uncertainty: Array1<f64>,
    pub psi_fd_orders: Vec<usize>,
    pub psi_gram_anchor_deltas: Option<(f64, f64)>,
    pub decomposition: OuterGradientFdDecomposition,
    pub rho: Option<OuterGradientFdRhoBlock>,
    pub curvature: Option<OuterCurvatureDriftAudit>,
}

Analytic-vs-finite-difference evidence for the ψ block at one real outer seed.

theta retains the complete outer seed and rho_dim locates the ψ block in that seed. Every gradient and scalar-stencil array contains exactly psi_dim entries in ψ-local order. Smoothing-parameter ρ coordinates are deliberately excluded: the κ/geometry gates that request this record do not grade them, and each unnecessary finite-difference coordinate costs two complete inner profiles.

Fields

theta: Array1<f64>

rho_dim: usize

psi_dim: usize

cost: f64

analytic_psi_gradient: Array1<f64>

finite_difference_psi_gradient: Array1<f64>

psi_steps: Array1<f64>

psi_fd_uncertainty: Array1<f64>

Ridders’ estimate of each ψ finite difference’s OWN error, and the truncation order of the accepted extrapolant (2 is a raw stencil, 4 one Richardson stage, …).

Present because a finite difference is an estimator: without it a consumer cannot tell an analytic-gradient defect from its own oracle’s truncation, and has to grade at whatever tolerance the worst step happens to need — which is how these gates ended up at 5e-2 (#2461). f64::INFINITY marks a coordinate the ladder could not resolve; such a component says nothing about the analytic gradient.

psi_fd_orders: Vec<usize>

psi_gram_anchor_deltas: Option<(f64, f64)>

Max-abs of the #1033b psi-Gram anchor correction applied to the criterion’s VALUE at this seed, as (gram_delta, rhs_delta).

joint_hyper pins the n-free tensor to the exactly streamed statistics by adding a constant offset measured at one reference psi, then installs the derivative of the UNCORRECTED tensor. A constant removes nothing from a derivative, so a non-zero value here is the tensor’s own value error at this seed and its SLOPE error is loose in the gradient lane (#2464).

None means the correction never ran on this seed – NOT that it ran and was zero. The distinction is the whole point of the field: a probe that reports 0.0 for “never fired” is unfalsifiable, and reading an absent emission as a measured zero is exactly how this quantity was first mis-measured.

decomposition: OuterGradientFdDecomposition

The per-atom breakdown of the same comparison, when the objective’s criterion is assembled from atoms at all.

rho: Option<OuterGradientFdRhoBlock>

The SAME comparison for the ρ (log-smoothing) block, when the caller armed enable_outer_gradient_fd_capture_over_theta.

None is the default and means the ρ coordinates were not differenced — NOT that they agreed. The two blocks are separate fields rather than one θ-indexed array because grading ρ is a deliberate, expensive opt-in: each coordinate costs a full Ridders ladder of inner profiles, and the shipped κ/geometry gates that consume the ψ block do not grade ρ.

curvature: Option<OuterCurvatureDriftAudit>

The curvature the logdet_h atom is taken on, differenced as a MATRIX against the analytic drift the same evaluation published (#2765).

None when the backend cannot hand out a dense H. See OuterCurvatureDriftAudit for why a scalar-only audit cannot decide what this decides.

Trait Implementations

impl Clone for OuterGradientFdRecord

fn clone(&self) -> OuterGradientFdRecord

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for OuterGradientFdRecord

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.