autofit_profiling

Where the time in one PyAutoFit search.fit goes, and the expectation-propagation / graphical-model baselines. Producer page for the PyAutoPulse instance fit (catalogue.json, profiling-summary@2). Measurements are declared, not judged.

gaussian_x1/nautilus/search_fit

af.Nautilus search.fit (PyAutoFit, checkpointing and output on) · wall 28.41 s (median of 3) · 7501 likelihood calls · cpu/numpy · PyAutoFit 2026.10.12.1 · measured 2026-10-14T09:19:45Z · result row

component	seconds	share of wall
nn_training	11.87	43.4 %
checkpoint_io	7.09	24.5 %
bound_geometry	5.60	20.4 %
visualisation	1.16	4.1 %
sampler_bookkeeping	0.67	2.3 %
samples_output	0.49	1.7 %
likelihood	0.32	1.1 %
other	0.29	1.0 %
proposal_sampling	0.22	0.8 %

Findings-page shares (sampler_internal_shares, information only): nn_training 51.7 % (Δ -4.7 pp), bound_geometry 25.2 % (Δ -3.4 pp), likelihood 14.3 % (Δ -13.1 pp) — outside ±10 pp.

gaussian_x1/nautilus/bare_sampler

bare nautilus.Sampler (the findings page's measurement; no PyAutoFit, no checkpoint) · wall 22.53 s (median of 3) · 16400 likelihood calls · cpu/numpy · PyAutoFit 2026.10.12.1 · measured 2026-10-14T09:19:45Z · result row

component	seconds	share of wall
nn_training	12.11	53.8 %
bound_geometry	5.80	25.3 %
likelihood	2.66	11.6 %
proposal_sampling	1.31	5.7 %
sampler_bookkeeping	0.70	3.1 %
other	0.00	0.0 %
checkpoint_io	0.00	0.0 %

Findings-page shares (shares_of_wall, reproduction gate): nn_training 51.7 % (Δ +2.1 pp), bound_geometry 25.2 % (Δ +0.1 pp), likelihood 14.3 % (Δ -2.7 pp) — within ±10 pp.

EP / graphical-model baselines (gaussian_1d_ep)

N 1-D Gaussians sharing one centre, fitted by expectation propagation (factor_graph.optimise, per-factor DynestyStatic + Laplace cavity) and by one joint DynestyStatic graphical fit. One fresh-process fit per cell, reproduced against the committed autofit_workspace_developer baselines with the runtime-drift-2x-1ms rule (producer note, not a Pulse comparison).

cell	wall s	committed s	ratio	drift	peak RSS MiB	truth logL sum	sanity (committed)
gaussian_1d_ep/ep/N10	152.9	566.9	0.27	improved	403	-472.253 (match yes)	PASS (PASS)
gaussian_1d_ep/ep/N3	37.1	251.9	0.15	improved	353	-128.397 (match yes)	PASS (PASS)
gaussian_1d_ep/ep/N30	842.8	1758.7	0.48	improved	642	-1500.561 (match yes)	FAIL (FAIL)
gaussian_1d_ep/graphical/N10	7.1	50.7	0.14	improved	446	-472.253 (match yes)	PASS (PASS)
gaussian_1d_ep/graphical/N3	6.0	22.8	0.26	improved	420	-128.397 (match yes)	PASS (PASS)
gaussian_1d_ep/graphical/N30	19.0	132.8	0.14	improved	515	-1500.561 (match yes)	FAIL (FAIL)

cpu/jax/float64 · PyAutoFit df6d218f10 · host DESKTOP-H143S82.

Analytic Gaussian benchmark (closed form vs EP vs graphical)

analytic Gaussian parity: closed form | minimal EP | autofit graphical | autofit EP · wall 130.0 s · peak RSS 332 MiB · PARITY PASS (41/41) · sigma_q95 12.1536 (banked 12.15, match yes) · result row

leg	column	cells passed	max a	max b	column s
A	minimal EP	6/6	0.000	0.000	0.0
A	autofit graphical	6/6	0.065	0.058	24.5
A	autofit EP	6/6	0.000	0.000	1.9
B	minimal EP	7/7	0.078	0.146	0.0
B	autofit graphical	7/7	0.050	0.041	44.4
B	autofit EP	7/7	0.079	0.146	56.8

Parked, not run: graphical/analytic_gaussian_priors.py (NEEDS_FIX 2026-09-30 - moments projection (PyAutoFit#1656) passes the gaussian and truncated scatter families (16/16 each) but the loggaussian family misses 11/16: log-space scatter comes back at log sigma 3.28 vs mode 1.53 / reference 1.83 - TransformedMessage moments path suspected; see PyAutoMind draft/bug/autofit/ep_moments_loggaussian_transformed_scatter.md).

Limitations

Laptop CPU only, pinned to one host (DESKTOP-H143S82, WSL2) so Pulse's same-host rule pairs every record; the epic's release-sweep pin (euclid-ral-gpu-2, an A100 node) does not fit CPU Dynesty runs. Cross-machine reproduction on RAL is a later campaign.
Wall times vary up to ±40 % between sessions on this host, so breakdown shares are the comparable quantity, not seconds; the EP / graphical / parity rows are single fresh-process fits (repetitions 1).
Component times are exclusive monkey-patched wall-clock intervals; 'other' is wall minus the instrumented total, not a measured phase.
The EP / graphical rows carry a producer-side reproduction note against the committed autofit_workspace_developer baselines (runtime-drift-2x-1ms, in each row's `reference`); it is not a Pulse comparison, because those baselines recorded no host and no record pair exists.
The N30 EP and graphical rows reproduce the committed sanity FAIL; their wall and memory are real measurements of a fit whose parameter recovery fails, not of a correct fit.
Every record is unreviewed: the producer declares measurements, the Brain's profiling conductor judges them; nothing here is a verdict.
No temporal comparisons: one baseline per setup.

Rendered from 9 result rows by build_catalogue.py; evidence updated 2026-10-14T14:45:45Z.