Deliberately excluded from the visual redesign (#1124) and its PR; grouped here because the fixes are entangled.
For regression/hierarchical models, each draw's geometry currently derives from different trials of the simulator metadata: metadata[\"boundary\"] is the last trial's (a scratch buffer in ssm-simulators), while metadata[\"trajectory\"]/t/z are trial 0's. The uncertainty bands added in #1124 are per-element honest, but the boundary ribbon and drift cone describe different trials. Intercept-only models are unaffected.
Scope:
- θ-reduction: one θ vector per (chain, draw) — per-draw trial-mean by default, plus an
obs= parameter to select a specific trial. Single-row no-noise simulator calls remove the mixed-trial problem by construction (and cut per-draw no-noise cost by ~n_trials).
- Deterministic reference geometry: the "mean" cartoon currently simulates one random unseeded trial of the trial-wise posterior-mean matrix (
plot_func_model / plot_func_model_n) — replace with the same reduction. Also makes the histogram baseline stable across calls.
- Seeded RNG end-to-end: one
np.random.default_rng(random_state) instead of the np.random.seed + global randint pattern; documented draw order so histograms, geometry, and trajectories are jointly reproducible.
max_t/t_s ownership: max_t is never passed to the simulator (defaults to 20 s → ~75% of every boundary polyline is outside the axes); t_s is also assigned in multiple branches. Compute once, pass everywhere.- Trajectory θ-sourcing: trajectories draw θ from
theta_mean whenever it exists, so they carry no posterior uncertainty even in samples mode; the fallback samples chain/draw/obs independently, composing θ vectors that exist in no draw.
The statistical caveat to document with the fix: geometry is nonlinear in θ, so the plug-in curve at the reduced mean is not the mean of the per-draw geometries and need not sit mid-band.
Deliberately excluded from the visual redesign (#1124) and its PR; grouped here because the fixes are entangled.
For regression/hierarchical models, each draw's geometry currently derives from different trials of the simulator metadata:
metadata[\"boundary\"]is the last trial's (a scratch buffer in ssm-simulators), whilemetadata[\"trajectory\"]/t/zare trial 0's. The uncertainty bands added in #1124 are per-element honest, but the boundary ribbon and drift cone describe different trials. Intercept-only models are unaffected.Scope:
obs=parameter to select a specific trial. Single-row no-noise simulator calls remove the mixed-trial problem by construction (and cut per-draw no-noise cost by ~n_trials).plot_func_model/plot_func_model_n) — replace with the same reduction. Also makes the histogram baseline stable across calls.np.random.default_rng(random_state)instead of thenp.random.seed+ globalrandintpattern; documented draw order so histograms, geometry, and trajectories are jointly reproducible.max_t/t_sownership:max_tis never passed to the simulator (defaults to 20 s → ~75% of every boundary polyline is outside the axes);t_sis also assigned in multiple branches. Compute once, pass everywhere.theta_meanwhenever it exists, so they carry no posterior uncertainty even in samples mode; the fallback samples chain/draw/obs independently, composing θ vectors that exist in no draw.The statistical caveat to document with the fix: geometry is nonlinear in θ, so the plug-in curve at the reduced mean is not the mean of the per-draw geometries and need not sit mid-band.