Halo mass recovery performance with DfL IIa: Distributions in the offset from true-to-predicted M₂₀₀ values (δM₂₀₀ ≡  M₂₀₀ (true)− M₂₀₀ (predicted)). The structure of this figure is identical to Fig. 6. In each panel a dashed red line shows the perfect regression value of δM₂₀₀ = 0, and the mean uncertainty on halo masses output from DfL is indicated by a horizontal error bar. Additionally, the bias (median offset), full 1σ dispersion (σ₁, containing 68% of the data set), 2σ dispersion (σ₂, containing 95% of the data), and the fraction of catastrophic outliers (those with |δM₂₀₀|> 1 dex) are indicated in each panel. For the cases of data and models being matched (top two rows), the pipeline recovers halo masses with essentially no bias (|b|≤0.02 dex) and an average uncertainty of ⟨σ₁⟩=0.12 dex. To test the impact of model uncertainty, in the lower two rows we show the δM₂₀₀ distributions for TNG data with an EAGLE model, and vice versa. Here an average bias of ⟨b⟩= − 0.10 dex is found by applying the EAGLE model to TNG data, and an average bias of ⟨b⟩= + 0.10 dex is found by applying the TNG model to EAGLE data. The overall random uncertainty remains essentially unchanged.