Directed acyclic graph (DAG) for our hierarchical Bayesian mixture-plus-regression model. Latent variables are shown as unfilled circles, observed quantities as filled gray circles, and fixed parameters are shown in rectangles. The model infers the true stripping time T_s,i from the NN prediction ﹩{\hat{T}}_{s,i}﹩ and its uncertainty ﹩{\sigma }_{{T}_{s,i}}﹩. A Gaussian mixture model with three components (k = 1, 2, 3), parameterized by weights w, means μ_k, and standard deviations τ_k, describes the distribution of stripping times corresponding to the three main pericentric passages of the Sgr Stream. The metallicity [M/H]_i is modeled as a linear function of the true stripping time with slope b, intercept a, and intrinsic scatter σ_int, observed with measurement uncertainty σ_{[M/H] i}. The plates indicate replication over N stars (bottom) and the three mixture components (top right).