Issues

summary().smooth_terms is empty for every model with a categorical main effect: 'y ~ factor(g) + s(x)' loses the whole per-smooth table that the byte-identical 'y ~ s(x) + group(g)' reports

claude-bug-hunt

#2787

· github-actions opened

on Aug 23, 2026

Location-scale summary().deviance is -2*loglik, not the deviance: noise_formula='1' changes the number 5-9x on an identical mean fit (binomial, where the two coincide, is unmoved)

claude-bug-hunt

#2786

· github-actions opened

on Aug 23, 2026

predict() returns mean != inverse_link(linear_predictor) in the same table: an undocumented, unswitchable O(1/n) retransformation term on mean only (up to 97% relative at n=100)

claude-bug-hunt

#2785

· github-actions opened

on Aug 23, 2026

Every shape-constrained smooth refuses predict(interval)/observation_interval/predict_conformal/diagnose at their defaults, while summary() prices its SEs off the covariance the gate says 'does not define a truncated law'

claude-bug-hunt

#2784

· github-actions opened

on Aug 23, 2026

s(x, identifiability=...) is never parsed on the 1-D path: every value including 'totally_bogus' is accepted and inert, while te/matern/thinplate honour it and reject bad tokens

claude-bug-hunt

#2783

· github-actions opened

on Aug 23, 2026

te()/ti() silently ignore documented degree= and penalty_order= on the default (cr) margin — bit-identical fits, no warning

claude-bug-hunt

#2782

· github-actions opened

on Aug 23, 2026

period= / periods= / period_start= / origin= are silently inert without periodic=: s(t, period=24) is bit-identical to s(t), no warning, seam discontinuity remains

claude-bug-hunt

#2781

· github-actions opened

on Aug 23, 2026

Model.partial_dependence crashes on any model with a string factor: the reference template is read from the ENCODED table, so the categorical sentinel comes back as "unseen level '\x00b0'"

claude-bug-hunt

#2780

· claude opened

on Aug 23, 2026

Rail-certified fit (null random-effect variance) refuses predict(interval)/observation_interval/predict_conformal/diagnose while summary() reports SEs — diagnose() has no covariance_mode escape

claude-bug-hunt

#2779

· claude opened

on Aug 23, 2026

PosteriorSamples.method badges every standard fit 'nuts' — Gaussian-identity (Laplace) and Bernoulli-logit (Polya-Gamma) included, with hardcoded rhat=1.0/ess=n_draws presented as MCMC diagnostics

claude-bug-hunt

#2778

· claude opened

on Aug 23, 2026

Model.sample() draws the conditional Vb while summary()/predict() publish the smoothing-corrected Vp — posterior SDs up to 43% below the reported SEs on the same fit

claude-bug-hunt

#2777

· claude opened

on Aug 23, 2026

Model.check() raises GamError on a non-finite covariate instead of returning it as a SchemaIssue — the documented safe_predict guard throws from inside itself

claude-bug-hunt

#2776

· claude opened

on Aug 23, 2026

Navigation Menu

All issues

summary().smooth_terms is empty for every model with a categorical main effect: 'y ~ factor(g) + s(x)' loses the whole per-smooth table that the byte-identical 'y ~ s(x) + group(g)' reports

Location-scale summary().deviance is -2*loglik, not the deviance: noise_formula='1' changes the number 5-9x on an identical mean fit (binomial, where the two coincide, is unmoved)

predict() returns mean != inverse_link(linear_predictor) in the same table: an undocumented, unswitchable O(1/n) retransformation term on mean only (up to 97% relative at n=100)

Every shape-constrained smooth refuses predict(interval)/observation_interval/predict_conformal/diagnose at their defaults, while summary() prices its SEs off the covariance the gate says 'does not define a truncated law'

s(x, identifiability=...) is never parsed on the 1-D path: every value including 'totally_bogus' is accepted and inert, while te/matern/thinplate honour it and reject bad tokens

te()/ti() silently ignore documented degree= and penalty_order= on the default (cr) margin — bit-identical fits, no warning

period= / periods= / period_start= / origin= are silently inert without periodic=: s(t, period=24) is bit-identical to s(t), no warning, seam discontinuity remains

Model.partial_dependence crashes on any model with a string factor: the reference template is read from the ENCODED table, so the categorical sentinel comes back as "unseen level '\x00b0'"

Rail-certified fit (null random-effect variance) refuses predict(interval)/observation_interval/predict_conformal/diagnose while summary() reports SEs — diagnose() has no covariance_mode escape

PosteriorSamples.method badges every standard fit 'nuts' — Gaussian-identity (Laplace) and Bernoulli-logit (Polya-Gamma) included, with hardcoded rhat=1.0/ess=n_draws presented as MCMC diagnostics

Model.sample() draws the conditional Vb while summary()/predict() publish the smoothing-corrected Vp — posterior SDs up to 43% below the reported SEs on the same fit

Model.check() raises GamError on a non-finite covariate instead of returning it as a SchemaIssue — the documented safe_predict guard throws from inside itself

Search results