Search results

Python gamfit.fit Surv() with no survival_likelihood routes to the broken 'location-scale' FitConfig default instead of the CLI's 'transformation' default, aborting the identifiability audit on textbook right-censored data the CLI fits

claude-bug-huntBugs found by Claude bug-hunt runsBugs found by Claude bug-hunt runs

Status: Open.

#1128 In SauersML/gam;

· claude opened on Jun 15, 2026

Gaussian s(x) REML fit is not equivariant to a small multiplicative response rescale: at a≲1e-5 the double-penalty smoothing-parameter selection over-smooths (λ̂ inflates ~40x at a=1e-6), reshaping—not just rescaling—the smooth (down-scale sibling of #1000)

claude-bug-huntBugs found by Claude bug-hunt runsBugs found by Claude bug-hunt runs

Status: Open.

#1127 In SauersML/gam;

· claude opened on Jun 15, 2026

Measure-jet smooth s(x, bs="mjs") is unfittable through the formula API / gamfit.fit (tol=1e-10) though the gam CLI (tol=1e-6) fits the identical data: REML-seeded anisotropic kappa optimizer non-convergence escalated to a fatal error instead of falling back to baseline

claude-bug-huntBugs found by Claude bug-hunt runsBugs found by Claude bug-hunt runs

Status: Open.

#1126 In SauersML/gam;

· claude opened on Jun 14, 2026

gam generate for a dispersion location-scale model (Gamma/NB/Beta/Tweedie + --predict-noise) ignores the fitted per-row precision channel: synthetic data is homoscedastic at the seed dispersion

claude-bug-huntBugs found by Claude bug-hunt runsBugs found by Claude bug-hunt runs

Status: Open.

#1125 In SauersML/gam;

· claude opened on Jun 14, 2026

Negative-Binomial generate / sample_replicates draws with seed theta=1, ignoring the estimated overdispersion theta_hat (NB sibling of Beta #770): Var=mu+mu^2 not mu+mu^2/theta_hat

claude-bug-huntBugs found by Claude bug-hunt runsBugs found by Claude bug-hunt runs

Status: Open.

#1124 In SauersML/gam;

· claude opened on Jun 14, 2026

SAE K=1 circle decoder rank-deficient (3/5) stalls past 2-min target on stage1-step0 (after #1113 NaN fix)

Status: Open.

#1117 In SauersML/gam;

· SauersML opened on Jun 14, 2026

PERF: ~17 quality tests exceed the 360s wall-clock budget — tensor/poisson/cyclic/CI-coverage/multinomial/competing-risks all time out

Status: Open.

#1082 In SauersML/gam;

· SauersML opened on Jun 13, 2026

n-independent outer loop: hyperparameter search must touch only k-dim objects (sufficient-statistic architecture)

Status: Open.

#1033 In SauersML/gam;

· SauersML opened on Jun 12, 2026

Compute-first primitive #3: multiresolution residual cascade for scattered low-d smooths at huge n

Status: Open.

#1032 In SauersML/gam;

· SauersML opened on Jun 12, 2026

ROADMAP: reconstruction parity with large linear SAEs — scale K, hybrid curved+linear dictionary, amortized routing

Status: Open.

#1026 In SauersML/gam;

· SauersML opened on Jun 11, 2026

PERF/SYSTEMS: device-resident SAE joint fit — the 1e4-1e6x hardware gap, measured and designed

Status: Open.

#1017 In SauersML/gam;

· SauersML opened on Jun 11, 2026

marginal-slope (binary + survival): severe slowdown / survival hang in gamfit 0.1.189

Status: Open.

#979 In SauersML/gam;

· SauersML opened on Jun 9, 2026