Impact of the gravity solver parameters on the SELFI posterior. Panels (a, c, e) display the summary statistics Φ for the three galaxy count fields, comparing the well-specified Model A (dashed lines) with the misspecified model (solid lines). The lower parts of these panels show the relative error between the two models, with shaded regions corresponding to a 1% error margin. Panels (b, d, f) show the corresponding SELFI posteriors for the initial matter power spectrum θ, with the well-specified and misspecified models represented by green and indigo lines, respectively. (a) Example of summary statistics for a single realisation with n_steps = 10 (solid lines) as opposed to n_steps = 20 time steps for Model A (dashed lines) in the gravitational evolution with COLA. Although the direct error remains well below 1% across all scales, the inverse error on the SELFI posterior (b) reaches 2% at the largest scales, corresponding to a deviation greater than 1σ from the well-specified Model A. (c) Summary statistics using only the final snapshot at redshift z₁ = 0.15 for the three mock galaxy populations, thereby neglecting light cone effects (solid lines). The forward error reaches 74% at the largest scales for the most distant galaxy population, and the corresponding SELFI posterior (d) significantly rejects the ground truth across all scales. (e) Summary statistics using a pure 2LPT gravity model instead of N-body simulations to define the galaxy populations (solid lines). The forward error reaches ∼7% at the smallest scale, and the ground truth is significantly rejected by the SELFI posterior (f), which deviates by more than 2σ compared to Model A at the largest scales.