prompt(research/graphical_ep): hand-fixed header + grounding for EP hierarchical slope prompt

Type docs→research, Target PyAutoLens→graphical_ep, verified concr paths,
added sampler-decision, HPC/RAL and backlog-relationship grounding.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: 2 people

active/rename_the_morning_skill_to_wake_up.md

@@ -0,0 +1,23 @@
+## wake-up-skill-rename
+- issue: https://github.com/PyAutoLabs/PyAutoBrain/issues/114 (closed)
+- completed: 2026-07-16
+- library-pr: https://github.com/PyAutoLabs/PyAutoBrain/pull/115 (merged 6caa6c4, merge commit)
+- repos: PyAutoBrain
+- summary: Renamed the /morning composition skill to /wake_up to fit the organism analogy (the organism wakes: syncs its body, checks its vitals, remembers what it was doing). Pure behaviour-preserving rename: skills/morning/ -> skills/wake_up/ (body file + SKILL.md name), /morning references updated in COMMANDS.md, AGENTS.md, AUTONOMY.md, bin/overnight_status.sh and nightly.sh comments. Deliberately untouched: Heart's morning_health/morning_status Slack webhooks (separately named automation), historical Mind records, time-of-day prose, and the SKILL.md discovery trigger "morning status/cleanup pass". Post-merge: bin/install.sh re-run refreshed the ~/.claude surfaces; stale morning symlinks pruned.
+
+## Original prompt
+
+# Rename the morning skill to wake_up (PyAutoBrain)
+
+Type: refactor
+Target: PyAutoBrain
+Repos:
+- PyAutoBrain
+Difficulty: small
+Autonomy: supervised
+Priority: normal
+Status: formalised
+
+Rename the morning skill to wake_up in @PyAutoBrain (skills/morning/ -> skills/wake_up/), adjusting all linked files accordingly (symlinks in ~/.claude/commands and ~/.claude/skills, COMMANDS.md, AGENTS.md, AUTONOMY.md, bin/overnight_status.sh, nightly.sh references) so it fits the PyAutoScientist organism analogy. The skill is proving very valuable; keep behaviour identical. Include a short assessment of what else the skill could absorb without becoming information overload.
+
+<!-- formalised by the Intake (Conception) Agent on 2026-07-16 from user-intake -->

complete/2026/07/wake-up-skill-rename.md

@@ -1,16 +1,19 @@
-# I want to use the autolens_assistant to perform Expectation Propagation
+# EP hierarchical power-law slope recovery — autolens_assistant science project
 
-Type: docs
-Target: PyAutoLens
+Type: research
+Target: graphical_ep
 Repos:
-- PyAutoLens
 - autolens_assistant
+- PyAutoFit
+- PyAutoLens
 - autolens_workspace
 Difficulty: too-large
 Autonomy: supervised
 Priority: high
 Status: formalised
 
+## Original request (verbatim)
+
 I want to use the autolens_assistant to perform Expectation Propagation (EP) analysis of a Cosmology science case.
 
 The science case is inferring the Hubble constant from time delay lensed quasars, and a example package already
@@ -40,4 +43,77 @@ The goal are:
 
 4) To test all the graphical EP diagnostics and result analysis code, especially tyhe recent EP updates we did last week.
 
-<!-- formalised by the Intake (Conception) Agent on 2026-07-16 from file:/tmp/claude-1000/-home-jammy-Code-PyAutoLabs/32708468-5918-4dbc-a763-583805364341/scratchpad/intake_ep_cosmology.md -->
+## Intake corrections & grounding (2026-07-16)
+
+**Corrected source paths** (verified on disk; the request has typos):
+
+- Legacy EP scripts: `/mnt/c/Users/Jammy/Science/concr/scripts/cosmology/`
+  — `ep.py`, `graphical.py`, `one_by_one.py`, `simple_plot.py`. These are the
+  "main scripts" to port: EP fit, joint graphical fit, and one-by-one fits.
+- Legacy simulator: `/mnt/c/Users/Jammy/Science/concr/simulators/cosmology.py`;
+  an existing simulated dataset lives at `concr/dataset/cosmology__time_delay`.
+- Workspace examples: `autolens_workspace/scripts/guides/modeling/advanced/`
+  — `expectation_propagation.py`, `graphical.py`, `hierarchical.py`.
+- Expect API drift in the `concr` package (old PyAutoLens); brush up against
+  the installed stack, grounding via the PyAuto API gate / `dir()` rather than
+  guessing.
+
+**Descoped goal (this task).** No H0 / time delays yet. Simulate N power-law +
+shear lenses with slopes drawn from a hierarchical (parent) distribution;
+example scripts recover the per-lens slopes and — the primary metric — the
+mean and scatter of the parent slope distribution. Keep the project layout
+H0-compatible so time-delay cosmography can be re-scoped in later.
+
+**Project home.** New science project stamped from
+`autolens_assistant/skills/start-new-project.md` (project skeleton: `scripts/`,
+`dataset/`, `results/`, `hpc/`, `wiki/project/profile.md`). Private science
+project, same pattern as PJ011646.
+
+**Goal 1 sampler — the `(SPECIFY)` decision, to settle at start_dev.**
+Candidates now first-class in PyAutoFit: multi-start JAX gradient optimizers
+(`MultiStartAdam` / ADABelief / Lion — promoted across lib + workspaces) for
+MAP-style recovery, and gradient-based posterior sampling (blackjax NUTS/HMC)
+for errors. The non-EP baseline must produce parent mean+scatter *with
+uncertainties*, so a sampling (not just optimization) path is needed for the
+comparison in goal 2.
+
+**Goal 4 context.** The "recent EP updates last week" are the diagnostics /
+result-analysis work shipped under `ep_framework_review.md` (this folder,
+execution-complete 2026-07-08, wrap-up on PyAutoFit#1330). This project is the
+end-to-end exercise of that shipped tooling on a realistic lensing case.
+
+**HPC (goal 3).** Run path is RAL via the assistant HPC link:
+
+- Cluster access: SSH alias `euclid_jump` (ProxyJump through `jump_finan`);
+  projects under `/mnt/ral/jnightin/<project>`; PyAuto venv at
+  `/mnt/ral/jnightin/PyAuto` kept in sync with local `main`s (`HPCPullPyAuto`),
+  activated via `activate.sh` (never pip-install PyAuto* into the venv).
+- Project machinery: copy the assistant's `hpc/` (batch_cpu/, batch_gpu/,
+  `sync`, `template.py`) into the new project per the start-new-project HPC
+  step; docs at `autolens_assistant/wiki/core/operations/hpc_infrastructure.md`.
+  Pipeline scripts must preserve the `template.py` HPC interface
+  (`parse_fit_args`, `--sample`/`--dataset`/`--use_cpu`/`--number_of_cores`).
+- Drive runs with `hpc/sync push-submit gpu <script>` (SLURM array, one
+  dataset per task, JAX auto-uses the GPU), then `hpc/sync jobs`/`tail`/`pull`.
+  The legacy `concr/hpc/` uses the same layout — port its `sync.conf` pattern,
+  not its stale scripts.
+- RAL `/mnt/ral` is NFS-slow: detach long remote work (nohup+setsid, poll a
+  log), never foreground-`timeout` an ssh.
+- N-lens array scaling is the design driver: one lens per SLURM array task for
+  the one-by-one/EP factor fits; the joint graphical fit needs a single
+  larger-memory task.
+
+**Relationship to existing backlog** (this folder):
+
+- `ep_scoping.md` — EP per-fit overhead / scale-up (performance): directly
+  relevant once N grows; do not duplicate its profiling work here.
+- `graphical_scoping.md` — joint-graph scale-up (performance): the joint-fit
+  baseline in goal 1/2 will hit exactly the dimensionality limits it maps.
+- `ep_framework_review.md` — execution-complete; its diagnostics are what
+  goal 4 validates.
+
+**Sizing.** too-large is right: expect phased delivery at start_dev
+(project stamp + simulator port → one-by-one/JAX-gradient baseline → EP fit +
+diagnostics → HPC scale-up), one PR per phase.
+
+<!-- formalised by the Intake (Conception) Agent on 2026-07-16 from file:/tmp/claude-1000/-home-jammy-Code-PyAutoLabs/32708468-5918-4dbc-a763-583805364341/scratchpad/intake_ep_cosmology.md; header + grounding hand-fixed at review (docs/autolens → research/graphical_ep) -->