Struct ExternalOptimResult 

pub struct ExternalOptimResult {

    pub beta: Array1<f64>,
    pub lambdas: Array1<f64>,
    pub likelihood_family: LikelihoodSpec,
    pub likelihood_scale: LikelihoodScaleMetadata,
    pub log_likelihood_normalization: LogLikelihoodNormalization,
    pub log_likelihood: f64,
    pub standard_deviation: f64,
    pub iterations: usize,
    pub finalgrad_norm: f64,
    pub outer_converged: bool,
    pub pirls_status: PirlsStatus,
    pub deviance: f64,
    pub stable_penalty_term: f64,
    pub used_device: bool,
    pub max_abs_eta: f64,
    pub constraint_kkt: Option<ConstraintKktDiagnostics>,
    pub artifacts: FitArtifacts,
    pub inference: Option<FitInference>,
    pub reml_score: f64,
    pub fitted_link: FittedLinkState,
    pub outer_cost_evals: usize,
    pub inner_pirls_solves: usize,
}

Fields

beta: Array1<f64>

lambdas: Array1<f64>

likelihood_family: LikelihoodSpec

likelihood_scale: LikelihoodScaleMetadata

log_likelihood_normalization: LogLikelihoodNormalization

log_likelihood: f64

standard_deviation: f64

Residual scale on the response scale.

Contract: Gaussian identity models store the residual standard deviation sigma here. Non-Gaussian families keep the response-scale summary used by their explicit likelihood-scale metadata.

iterations: usize

finalgrad_norm: f64

outer_converged: bool

True iff the outer optimizer reached a stationary point (gradient norm below tolerance), as reported by the optimizer itself. False when the run exhausted its iteration budget without reaching the gradient tolerance. Downstream consumers should NOT assume that a fit with outer_converged == false is unusable — it may still be the best basin reached given the budget — but they must not treat it as certified-converged either.

pirls_status: PirlsStatus

deviance: f64

stable_penalty_term: f64

Stable quadratic penalty term βᵀSβ, including any solver ridge quadratic.

used_device: bool

max_abs_eta: f64

constraint_kkt: Option<ConstraintKktDiagnostics>

artifacts: FitArtifacts

inference: Option<FitInference>

reml_score: f64

Complete REML/LAML objective value used for smoothing selection.

fitted_link: FittedLinkState

outer_cost_evals: usize

Number of outer REML cost-only evaluations executed during the fit (the count the outer optimizer’s trust-region/line-search probes drive, each paying an inner P-IRLS solve). Surfaced for regression guards on outer work (#1575); not part of the statistical contract.

inner_pirls_solves: usize

Number of actual full-n inner P-IRLS solves performed (cache-missing prepare_eval_bundlewithkey calls). This is the true cost driver the #1575 slowdown is measured in (“~150 outer cost evals each running a full n-sized P-IRLS”): unlike outer_cost_evals, it excludes single-slot cache hits and prior short-circuits and includes every solve done by the seed-grid prepass, screening, multistart, and finalize phases. Surfaced for a regression guard that pins the warm-start / parsimony-waiver / PSIS-optin economy (#1575); not part of the statistical contract.