[Jammy2211]

Registers a new science project in the body map and files its first task.

What the project is

ep_toy_gaussian is the graphical-ep campaign's last confidence test before scaling up:
the original toy on which the hierarchical-parameter collapse (PyAutoFit#1405) was first
found, run at N=50, with a joint JAX + BlackJAX NUTS reference, and fitted with EP
three times on identical data. The model and every EP setting are verbatim from
PyAutoMind/complete/2026/07/ep_scale_collapse_assets/ep_toy_diagnostic.py
(DynestyStatic(nlive=100, rwalk) per factor, LaplaceOptimiser, EPHistory(kl_tol=0.05),
max_steps=20); the three repeats differ only in RNG.

Nothing has ever run this toy with a NUTS joint fit, at N=50, or with repeated EP fits on
identical data — the task file records the four distinct earlier runs the "it collapsed"
memory is a composite of, with their paths and numbers, so nobody re-derives them.

A pass on criteria 2-5 is the evidence that the D1-D6 / analytic-Gaussian fix wave
(#1465, #1558, #1560, #1562, #1572-#1576, #1578, #1580) cured #1405 on its own toy. A failure
isolates it to the Dynesty-per-factor EP path, since the analytic ensemble used Laplace factors.

CPU-only, data simulated in memory from its seed, remote: none until the human runs
gh repo create PyAutoLabs/ep_toy_gaussian --private.

The pre-registered witness

Reference = the NUTS posterior on n50_seed42; truth = the parent (50, 10) and the realised
sample statistics of the 50 drawn centres. Mirrored machine-readably as WITNESS in the
project's scripts/compare.py.

#	criterion	threshold
1	NUTS health	0 divergences; r-hat <= 1.01 and ESS >= 400 on parent mean and sigma; realised sample mean/std inside the NUTS 3 sigma interval
2	No collapse	every EP repeat: parent sigma >= 1.0 (10% of truth) and the #1465 guard silent — PATHOLOGICAL count 0/3
3	EP-NUTS parity	every repeat: |mean_EP - mean_NUTS| <= 2 sigma_NUTS(mean) and |sigma_EP - sigma_NUTS| <= 2 sigma_NUTS(sigma)
4	EP self-agreement	max pairwise spread across the 3 repeats <= 1 sigma_NUTS, for parent mean and for sigma
5	Honest widths	every repeat: sigma_EP / sigma_NUTS within [1/3, 3] on parent mean and sigma — BIASED-TIGHT count 0/3
6	Per-dataset centres	>= 95% of the 50 centres within 3 sigma_NUTS of the NUTS median, in every repeat
7	EP flag health	no factor ends any repeat with zero SUCCESS updates; no CRASH; no NaN RuntimeWarning

Deliverable: results/n50_seed42/summary.json + comparison.md, committed in the project.
A missed criterion is a finding to rule on, never a number to relax.

Contents

• projects.yaml — new active row ep_toy_gaussian, cloned from analytic_gaussian.
• tasks/ep_toy_gaussian/nuts_vs_ep_x3_n50.md — State: ready, Epic: graphical-ep,
  Budget: 24:00. Submission is a separate step; no run lines yet.
• dashboard.md / dashboard.html regenerated (cortex dashboard --check current).

python3 scripts/cortex.py check prints cortex check: OK.

Why a human merges this

projects.yaml is code, not ledger — local_path and sync_cli are paths a conductor will
execute under. python3 scripts/ledger_merge.py classify --base origin/main says
code: 1 of 4 path(s) need a human — projects.yaml, so ledger_merge.yml will correctly leave
this branch for review rather than auto-merging it.

🤖 Generated with Claude Code

https://claude.ai/code/session_018uj47GfPHkyG7rz2J2Yna3