cortex: ep_toy_gaussian birth — NUTS vs EP ×3 on the #1405 collapse toy at N=50

dashboard.html

@@ -237,6 +237,7 @@
 <table class="map">
 <tr><th>Project</th><th>Next task</th><th>Awaiting</th><th>Running</th><th>Last ruling</th></tr>
 <tr><td><b>analytic_gaussian</b></td><td>Is criterion 2's mu threshold or the minimal-EP control wrong</td><td>0</td><td>0</td><td>2026-09-10</td></tr>
+<tr><td><b>ep_toy_gaussian</b></td><td>Do three EP fits agree with NUTS at N=50</td><td>0</td><td>1</td><td>-</td></tr>
 <tr><td><b>euclid_dr1_prelim</b></td><td>Do ordered Nautilus reruns reproduce the euclid catalogue values?</td><td>1</td><td>0</td><td>2026-09-10</td></tr>
 <tr><td><b>ic50_workspace</b></td><td>Does EP match the graphical joint fit end to end</td><td>0</td><td>1</td><td>-</td></tr>
 <tr><td><b>slope_hierarchy_scale</b></td><td>Does hierarchical slope recovery hold at N=25 to 50</td><td>0</td><td>1</td><td>-</td></tr>
@@ -273,6 +274,14 @@ <h3>analytic_gaussian</h3>
 <p class="muted"><a href="https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/analytic_gaussian/">tasks/analytic_gaussian/</a></p>
 </section>
 <section class="project">
+<h3>ep_toy_gaussian</h3>
+<p class="paths"><b>Local</b> <span class="pathchip">/mnt/c/Users/Jammy/Science/ep_toy_gaussian</span></p>
+<p class="paths"><b>RAL</b> <span class="pathchip">/mnt/ral/jnightin/ep_toy_gaussian</span></p>
+<p class="muted">active · ral partition · tasks: submitted 1</p>
+<div class="task"><button class="copy" data-cmd="cd /mnt/c/Users/Jammy/Science/ep_toy_gaussian &amp;&amp; hpc/sync jobs" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoCortex/blob/main/tasks/ep_toy_gaussian/nuts_vs_ep_x3_n50.md">Do three EP fits agree with NUTS at N=50</a><span class="tags"><span class="pill y">submitted</span></span><span class="facets"> — runs 342639, 342640 · wall 0:00 of 24:00 (0%)</span></p></div>
+<p class="muted"><a href="https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/ep_toy_gaussian/">tasks/ep_toy_gaussian/</a></p>
+</section>
+<section class="project">
 <h3>euclid_dr1_prelim</h3>
 <p class="paths"><b>Local</b> <span class="pathchip">/mnt/c/Users/Jammy/Science/euclid_dr1_prelim</span></p>
 <p class="paths"><b>RAL</b> <span class="pathchip">/mnt/ral/jnightin/euclid_dr1_prelim</span></p>

dashboard.md

@@ -30,6 +30,7 @@ One row per active project — the next thing, what it is holding, and when it l
 | Project | Next task | Awaiting | Running | Last ruling |
 |---|---|---:|---:|---|
 | analytic_gaussian | Is criterion 2's mu threshold or the minimal-EP control wrong | 0 | 0 | 2026-09-10 |
+| ep_toy_gaussian | Do three EP fits agree with NUTS at N=50 | 0 | 1 | - |
 | euclid_dr1_prelim | Do ordered Nautilus reruns reproduce the euclid catalogue values? | 1 | 0 | 2026-09-10 |
 | ic50_workspace | Does EP match the graphical joint fit end to end | 0 | 1 | - |
 | slope_hierarchy_scale | Does hierarchical slope recovery hold at N=25 to 50 | 0 | 1 | - |
@@ -130,6 +131,24 @@ Finish: write today's journal entry, rewrite wiki/project/state.md, and return (
 
 [tasks/analytic_gaussian/](https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/analytic_gaussian/)
 
+### ep_toy_gaussian
+
+**Local** `/mnt/c/Users/Jammy/Science/ep_toy_gaussian`
+
+**RAL** `/mnt/ral/jnightin/ep_toy_gaussian`
+
+active · ral partition · tasks: submitted 1
+
+<details><summary>📋 <a href="tasks/ep_toy_gaussian/nuts_vs_ep_x3_n50.md">Do three EP fits agree with NUTS at N=50</a> — <b>submitted</b> — runs 342639, 342640 · wall 0:00 of 24:00 (0%)</summary>
+
+```
+cd /mnt/c/Users/Jammy/Science/ep_toy_gaussian && hpc/sync jobs
+```
+
+</details>
+
+[tasks/ep_toy_gaussian/](https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/ep_toy_gaussian/)
+
 ### euclid_dr1_prelim
 
 **Local** `/mnt/c/Users/Jammy/Science/euclid_dr1_prelim`

projects.yaml

@@ -115,6 +115,20 @@ analytic_gaussian:
   status: active
   note: "born 2026-09-09 as the graphical-ep campaign's phase-1 keystone at ensemble scale; benchmark model with a closed-form posterior, borrowed from autofit_workspace_test scripts/graphical (PR #92 lineage); CPU-only, no dataset (every draw is simulated in memory from its seed); remote none until the human runs gh repo create PyAutoLabs/analytic_gaussian --private"
 
+ep_toy_gaussian:
+  remote: none
+  local_path: /mnt/c/Users/Jammy/Science/ep_toy_gaussian
+  ral_root: /mnt/ral/jnightin/ep_toy_gaussian
+  mirror: none
+  sync_cli: hpc/sync
+  sync_verbs: [push, pull, logs, sync, status, submit, push-submit, jobs, sacct, cancel, tail, du, check]
+  ledger: wiki/project/state.md
+  assistant: autofit_assistant
+  witness_file: results/**/*.json
+  partition: ral
+  status: active
+  note: "born 2026-09-10 as the graphical-ep campaign's confidence test on the PyAutoFit#1405 collapse toy at N=50: joint BlackJAX NUTS vs EP x3 on identical data; model and EP settings verbatim from complete/2026/07 ep_toy_diagnostic; CPU-only, data simulated in memory from its seed; remote none until the human runs gh repo create PyAutoLabs/ep_toy_gaussian --private"
+
 # --------------------------------------------------------------------------- #
 # dormant — git repos only. The non-git trees (subhalo, euclid_group, z_*) are
 # storage, not projects, and aris_PJ011646 is an assistant clone; none get a

tasks/ep_toy_gaussian/nuts_vs_ep_x3_n50.md

@@ -0,0 +1,109 @@
+# Ep_toy_gaussian — do three EP fits agree with NUTS on the #1405 collapse toy at N=50
+
+Project: ep_toy_gaussian
+Summary: Do three EP fits agree with NUTS at N=50
+State: submitted
+Gates:
+Witness: over one 50-dataset draw (seed 42), NUTS is healthy and every one of 3 EP repeats on identical data recovers the NUTS parent mean and sigma within the pre-registered thresholds in `## Witness`, committed as `results/n50_seed42/{summary.json, comparison.md}`
+Budget: 24:00
+Runs: 342639, 342640
+Ruling:
+Review-minutes: 20
+Epic: graphical-ep
+Filed: 2026-09-10
+
+## Question
+
+The `graphical-ep` campaign has three independent lines of evidence that graphical and EP are
+sound after the D1-D6 wave: the cosmology run (`slope_hierarchy_scale`, N=25), IC50 (33/33 at
+N=5) and the analytic Gaussian ensemble (200 seeds, ruling R-20260910-04). What none of them is,
+is the toy on which the hierarchical-parameter collapse was *first* found. This task is the last
+confidence test before the campaign scales up: run **that** model, at 50 datasets, with a joint
+**JAX + NUTS** reference, and fit it with EP **three times on identical data**. Do all three EP
+fits agree with each other and with NUTS?
+
+### What the earlier "collapse" runs actually were
+
+The "the toy collapsed" memory is a composite of four different runs, none of which is the run
+this task performs. Recorded here so nobody re-derives it:
+
+| run | what it was | collapse seen |
+|---|---|---|
+| `PyAutoMind/complete/2026/07/ep_scale_collapse_assets/ep_toy_diagnostic.py` + `EP_TOY_FINDINGS.md` (2026-07-21, PyAutoFit `f83f2f493`) | **The original.** HowToFit `gaussian_x1__hierarchical` toy: N=5 Gaussians, `centre_i ~ N(50, 10)`, `normalization=0.5`, `sigma=5` fixed, `centre` prior `TruncatedGaussian(50, 20, 0, 100)`, `HierarchicalFactor(GaussianPrior, mean=TG(50,10,0,100), sigma=TG(10,5,0,100))`, per-factor `DynestyStatic(nlive=100, sample="rwalk")`, `LaplaceOptimiser`, `EPHistory(kl_tol=0.05)`, `max_steps=20`. Joint reference = **Dynesty** on `global_prior_model`, not NUTS. | 30 identical runs: 70% RECOVER / 7% COLLAPSE (`scatter=0.0030 +/- 0.0000`, mean -> 54.4) / 23% CRASH. The joint fit was stable every time. |
+| `PyAutoMind/complete/2026/09/ep-scale-collapse-leg2-assets/{toy.py,run_once.py,classify.py,results_baseline.txt}` | Self-contained rebuild of the same model (data regenerated in memory; `TOY_N`, `TOY_SEED`, `TOY_MAX_STEPS`, `TOY_OPT`, `TOY_UPDATER`), noise `TOY_NOISE=0.05`. | 20 seeds at N=5: 13 COLLAPSE / 7 RECOVER (`scatter=0.0000 err=1e-11`) — harsher than the original. Three-state classifier PATHOLOGICAL / BIASED-TIGHT / RECOVER. |
+| `autofit_workspace_test/scripts/graphical/analytic_gaussian_collapse.py` | The closed-form referee's collapse configuration (N=5, 5 seeds, `max_steps=20`). | pre-fix FAIL 3/5 (STALE `10.00 +/- 3.69`, BIASED-TIGHT `7.78 +/- 0.51`); post #1558/#1560/#1562 RECOVER 5/5. |
+| `Science/z_projects_complete/slope_hierarchy` (lensing, N=5) | The only **NUTS-vs-EP** hierarchical comparison to date. | NUTS sigma `0.143 [0.117, 0.185]` vs EP `0.026 +/- 0.00001` — 4x low, errors ~1000x too tight. |
+| "50 datasets" | `autofit_workspace/scripts/simulators/simulators_sample.py:29` and `Science/concr/simulators/toy.py` — the **fixed-centre** low-SNR sample, shared-centre EP, no hierarchy. | not a collapse run |
+
+**Nothing has ever run this toy with a NUTS joint fit, at N=50, or with repeated EP fits on
+identical data.** The project is new in exactly those three respects, and keeps the original
+model and EP settings verbatim.
+
+### The fixes that might have cured it
+
+Everything below landed in PyAutoFit after the 2026-07-21 original run (local source checkout
+`PyAutoFit` main `81927cce9`, version string still 2026.8.17.1 — cite the SHA, not the version):
+**#1465** (collapse guard), **#1558** (D1 id-0 gradient), **#1560** (D4 message limits),
+**#1562** (D2/D3 fd-Hessian + skip-not-write), **#1572 / #1573 / #1574 / #1576** (the review wave,
+per-variable stale tracking), **#1578 / #1580** (stale-mask fixed point), plus the 2026-08-05
+`InitializerException` no-abort commit, which is the 23% CRASH mode of the original run. The RAL
+PyAuto mirror was verified at PyAutoFit `66f9f8d5d` (contains `08207bad0`, #1580) on 2026-09-09;
+re-verify before submitting.
+
+### What is new here
+
+1. A **joint NUTS reference** (`af.BlackJAXNUTS`, `use_jax=True`) instead of Dynesty — the first
+   time this toy has had a gradient-based joint posterior to be judged against.
+2. **N=50**, ten times the original's N=5, which is where the campaign wants to scale to.
+3. **Three EP repeats on identical data**, differing only in RNG — so EP's run-to-run spread is
+   measured directly rather than inferred across seeds. A pass on criteria 2-5 is the evidence
+   that the phase-2 / analytic-Gaussian fixes cured #1405 on its own toy; a failure isolates it
+   to the Dynesty-per-factor EP path, since the analytic ensemble used Laplace factors.
+
+## Witness
+
+Pre-registered before submission and mirrored machine-readably as `WITNESS` in
+`scripts/compare.py`. Reference = the NUTS posterior on `n50_seed42`; truth = the parent
+`(50, 10)` and the realised sample statistics of the 50 drawn centres. **A missed criterion is a
+finding to rule on, never a number to relax so a run can pass**; changing one is its own task,
+with the reason in the ruling.
+
+| # | 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` (the witness file) + `comparison.md`, committed in
+`ep_toy_gaussian`, carrying every criterion's measured number.
+
+## Where to look
+
+- `ep_toy_gaussian` (this project's row): `/mnt/c/Users/Jammy/Science/ep_toy_gaussian` —
+  `scripts/{toy.py, one_by_one.py, graphical_nuts.py, ep_repeat.py, compare.py}`,
+  `hpc/batch_cpu/{submit_nuts, submit_ep}`, `wiki/project/state.md`
+- `results/n50_seed42/` — `nuts.json`, `one_by_one.json`, `ep_r{0,1,2}.json`, and the two
+  deliverables `summary.json` + `comparison.md`. The local smoke sample is `n5_smoke`
+- `PyAutoMind/complete/2026/07/ep_scale_collapse_assets/` — the original run: `ep_toy_diagnostic.py`
+  and `EP_TOY_FINDINGS.md`, the source of the model and the EP settings used verbatim here
+- `PyAutoMind/complete/2026/09/ep-scale-collapse-leg2-assets/` — `toy.py` (borrowed for
+  `scripts/toy.py`), `run_once.py`, `classify.py` (the three-state labels), `results_baseline.txt`
+- `PyAutoMind/draft/research/graphical_ep/ep_campaign.md` — the campaign ledger; this task is
+  phase 1's confidence test
+- Cortex `tasks/analytic_gaussian/ensemble_parity.md` (ruling R-20260910-04) — the 200-seed
+  closed-form ensemble whose fixes this task tests on the original toy
+- `Science/z_projects_complete/slope_hierarchy` — the only prior NUTS-vs-EP hierarchical comparison
+- PyAutoFit#1405 — the collapse basin, still open pending the moment-matching cure
+
+## Runs
+
+- 342639: submitted — ral — submitted 2026-09-10 — wall 0:00 — one_by_one + BlackJAX NUTS, partition ral
+- 342640: submitted — ral — submitted 2026-09-10 — wall 0:00 — EP x3 on identical data, array 0-2
+
+## Ruling
+
+(none)