[Jammy2211]

Why

The graphical-ep campaign's end goal 1 asks for a model that statistically demonstrates the graphical and EP source code is correct — means and errors against a closed form. Phase 1 built that model (autofit_workspace_test#91 / PR #92) and ran it at one seed as a CI gate. That form earned its keep: it found six PyAutoFit defects (D1–D6, all shipped: #1558, #1560, #1562, #1572, #1573, #1574, #1576, #1578, #1580). But one seed cannot answer a question about a distribution — its 41-cell count moved 37/41 → 38/41 between two runs of identical code, because a verdict on a tolerance edge is a coin flip.

Phases 3 and 4 became science projects that run on RAL and are scored against pre-registered witnesses (slope_hierarchy_scale, ic50_workspace). The keystone that referees both was the one arm still living as a parked CI script. This gives it the same standing.

What

• projects.yaml — new active row analytic_gaussian: partition ral, assistant autofit_assistant, remote: none until a human runs gh repo create PyAutoLabs/analytic_gaussian --private. CPU-only, and with no dataset/: every draw is simulate(seed, n_datasets) in memory, so a seed is the dataset and nothing can drift between laptop and cluster.
• tasks/analytic_gaussian/ensemble_parity.md — State submitted, Runs: 342413. A 200-seed ensemble at N=5, live on the ral partition (50 concurrent).
• Dashboard regenerated in the same commit.

The witness was pre-registered before the wave went out

Six criteria, mirrored machine-readably as WITNESS in the project's aggregate_ensemble.py, covering the a/b error distributions per column, the empirical coverage of the closed form's [q05, q95] on the scatter (with a Wilson interval, so "3/3 covered" cannot be read as certainty), the scale-collapse rate, and EP flag health.

Criterion 4 (autofit EP, leg B) is expected to miss — that is the point. The documented Laplace-on-scatter caveat (PyAutoFit#1405, autofit/graphical/README.md §3.5) becomes a measured rate, and the moment-matching cure gets a number to beat.

Evidence already in hand

First local run on the post-D1–D6 library (seed 0, N=5, autofit 2026.8.17.1):

• leg A autofit EP is now exact — mu 50.8595 ± 4.1104 against a reference of 50.8595 ± 4.1104 (a = b = 0.000), and the same on every x_i. This is the leg that parked analytic_gaussian.py NEEDS_FIX, where mu used to sit at its starting prior 50.00 ± 20.6.
• leg B autofit EP returns σ 9.3722 ± 3.5664 against 6.5667 ± 2.8832: outside the cell tolerance, inside [q05, q95] = [2.980, 12.154], not collapsed.
• Over three local seeds the referee already reads: leg A EP 3/3, leg B EP 2/3, minimal EP exact to 1.16e-14.

The RAL PyAuto mirror was verified at PyAutoFit 66f9f8d5d (contains 08207bad0, #1580) before submitting — phase 3's EP arm ran on a mirror that predated the fix it was meant to exercise, and that is not repeated here.

Human call

projects.yaml is code, so this needs a human merge. The science tree is born locally at /mnt/c/Users/Jammy/Science/analytic_gaussian (commit ea3853f) and is unpushed until gh repo create PyAutoLabs/analytic_gaussian --private.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GkvYQqRKrZU4K9siYCSd8M