Issues

gam CLI and gamfit Python converge to different fits (λ/EDF/REML) on byte-identical data — CLI under-smooths to a worse REML optimum, violating the #1191/#1196 single-engine contract

claude-bug-hunt

#2116

· claude opened

on Jul 3, 2026

gam predict --uncertainty silently shifts the point prediction (η̂/mean) vs plain gam predict and vs gamfit.predict — O(1/n) bias correction fires only in the uncertainty arm

claude-bug-hunt

#2115

· claude opened

on Jul 3, 2026

ResponseGeometryModel cannot be saved or loaded: gamfit.save raises TypeError (no .save/.dumps), no reconstruction in load/loads — manifold-response models have no persistence path (analogous to fixed #2078)

claude-bug-hunt

#2114

· claude opened

on Jul 3, 2026

partial_dependence standard errors use the uncorrected conditional covariance (H⁻¹), not the smoothing-corrected one every other surface reports — term bands ~20-30% too narrow, under-cover

claude-bug-hunt

#2113

· claude opened

on Jul 3, 2026

Parametric-AFT (location-scale survival) MLE spuriously fails to converge on clean data: absolute 1e-7 gradient tol applied to the O(n) summed gradient

claude-bug-hunt

#2112

· claude opened

on Jul 3, 2026

Producer: born-circle source separation on a dense product-of-circles torus needs ISA (ring-ness is degenerate)

#2111

· SauersML opened

on Jul 3, 2026

Stagewise birth fit destroys a good LSQ decoder → born atom rejected (K=1 birth-fit fragility; deferred, resurfaces at K-scale)

#2101

· SauersML opened

on Jul 3, 2026

Scale-quotient (cone-atom) deep integration: unit-B through isometry g=JᵀJ (#673), row metric (#980), decoder-norm guards on s_k

#2099

· SauersML opened

on Jul 3, 2026

ManifoldSAE facade: migrate Python dataclass state into a Rust-owned #[pyclass] (full-native model object)

#2091

· SauersML opened

on Jul 2, 2026

sae_manifold_fit can terminate the Python process on tiny IBP circle fits instead of raising an exception

#2089

· SauersML opened

on Jul 2, 2026

Red-tree triage: ~17 gam-sae test failures accumulated during the ship-fast regime — attribution matrix and fix lanes

#2087

· SauersML opened

on Jul 2, 2026

Chart angle-fidelity: EV-optimal circle fits read non-uniform angles (corr 0.77 at chart EV 0.93)

#2081

· SauersML opened

on Jul 2, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

gam CLI and gamfit Python converge to different fits (λ/EDF/REML) on byte-identical data — CLI under-smooths to a worse REML optimum, violating the #1191/#1196 single-engine contract

gam predict --uncertainty silently shifts the point prediction (η̂/mean) vs plain gam predict and vs gamfit.predict — O(1/n) bias correction fires only in the uncertainty arm

ResponseGeometryModel cannot be saved or loaded: gamfit.save raises TypeError (no .save/.dumps), no reconstruction in load/loads — manifold-response models have no persistence path (analogous to fixed #2078)

partial_dependence standard errors use the uncorrected conditional covariance (H⁻¹), not the smoothing-corrected one every other surface reports — term bands ~20-30% too narrow, under-cover

Parametric-AFT (location-scale survival) MLE spuriously fails to converge on clean data: absolute 1e-7 gradient tol applied to the O(n) summed gradient

Producer: born-circle source separation on a dense product-of-circles torus needs ISA (ring-ness is degenerate)

Stagewise birth fit destroys a good LSQ decoder → born atom rejected (K=1 birth-fit fragility; deferred, resurfaces at K-scale)

Scale-quotient (cone-atom) deep integration: unit-B through isometry g=JᵀJ (#673), row metric (#980), decoder-norm guards on s_k

ManifoldSAE facade: migrate Python dataclass state into a Rust-owned #[pyclass] (full-native model object)

sae_manifold_fit can terminate the Python process on tiny IBP circle fits instead of raising an exception

Red-tree triage: ~17 gam-sae test failures accumulated during the ship-fast regime — attribution matrix and fix lanes

Chart angle-fidelity: EV-optimal circle fits read non-uniform angles (corr 0.77 at chart EV 0.93)

Search results