All issues

typeof(*)(:in) log-scale for a scalar A is -logdet(a): throws for a < 0 and misses the dimension

#680

· bvdmitri opened

on Sep 23, 2026

Sampled typeof(*)(:A)/(:in) messages weight each draw by |x|: they estimate out/in, not the BP message

#679

· bvdmitri opened

on Sep 23, 2026

typeof(-) marginal rules shift in2 by out − in1 instead of in1 − out when in1 is a point mass

#678

· bvdmitri opened

on Sep 23, 2026

typeof(+)(:in1) BLAS specialisation for two MvNormalWeightedMeanPrecision messages flips the sign of the mean

#677

· bvdmitri opened

on Sep 23, 2026

MvNormalGamma(:out) VMP rule drops tr(E[Λ] Cov μ)/2 from the rate for a non-point-mass q_μ

#676

· bvdmitri opened

on Sep 23, 2026

Wishart(:out) VMP rules use E[S]⁻¹ instead of E[S⁻¹] for a non-point-mass q_S

#675

· bvdmitri opened

on Sep 23, 2026

MvNormalWeightedMeanPrecision marginal rule returns keys m_out/m_ξ/m_Λ, breaking its free energy

#674

· bvdmitri opened

on Sep 23, 2026

MvNormalMeanCovariance VMP rules use E[Σ] instead of E[Σ⁻¹]⁻¹ (multivariate #669)

#673

· bvdmitri opened

on Sep 23, 2026

Gamma and GammaInverse average energies take E[x/θ] as E[x]/E[θ]

#672

· bvdmitri opened

on Sep 23, 2026

NormalMeanVariance VMP rules use E[v] instead of 1/E[1/v] as the message variance

#669

· bvdmitri opened

on Sep 23, 2026

CVI projection accuracy depends on the Cholesky fallback strategy (deterministic, reproducible)

#668

· bvdmitri opened

on Sep 21, 2026

cholinv is called on non-positive-definite matrices; propose a rule context to control the fallback, defaulting to NoCorrection

#667

· bvdmitri opened

on Sep 21, 2026