diff --git a/dashboard.html b/dashboard.html
index 3b69906..138ab6d 100644
--- a/dashboard.html
+++ b/dashboard.html
@@ -241,6 +241,7 @@
 <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</b></td><td>Does the sep1 DR1 delivery reproduce the prelim ten lenses</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>euclid_sersics</b></td><td>Core two-stage run of the 100 sersic lenses</td><td>0</td><td>1</td><td>-</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>
 <tr><td><b>subhalo_validation</b></td><td>Does rectangular_adapt false-detect a subhalo on the clumpy source</td><td>0</td><td>5</td><td>2026-09-07</td></tr>
@@ -330,6 +331,14 @@ <h3>euclid_dr1_prelim</h3>
 <p class="muted"><a href="https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/euclid_dr1_prelim/">tasks/euclid_dr1_prelim/</a></p>
 </section>
 <section class="project">
+<h3>euclid_sersics</h3>
+<p class="paths"><b>Local</b> <span class="pathchip">/mnt/c/Users/Jammy/Science/euclid_sersics</span></p>
+<p class="paths"><b>RAL</b> <span class="pathchip">/mnt/ral/jnightin/euclid_sersics</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/euclid_sersics &amp;&amp; hpc/sync jobs" aria-label="Copy the Claude command">📋</button><p><a href="https://github.com/PyAutoLabs/PyAutoCortex/blob/main/tasks/euclid_sersics/core_two_stage_100.md">Core two-stage run of the 100 sersic lenses</a><span class="tags"><span class="pill y">submitted</span></span><span class="facets"> — runs 342696 · wall 0:00 of 36:00 (0%)</span></p></div>
+<p class="muted"><a href="https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/euclid_sersics/">tasks/euclid_sersics/</a></p>
+</section>
+<section class="project">
 <h3>ic50_workspace</h3>
 <p class="paths"><b>Local</b> <span class="pathchip">/mnt/c/Users/Jammy/Science/ic50_workspace</span></p>
 <p class="paths"><b>RAL</b> <span class="pathchip">/mnt/ral/jnightin/ic50_workspace</span></p>
diff --git a/dashboard.md b/dashboard.md
index 21d5ec0..b8bfcc6 100644
--- a/dashboard.md
+++ b/dashboard.md
@@ -34,6 +34,7 @@ One row per active project — the next thing, what it is holding, and when it l
 | ep_toy_gaussian | Do three EP fits agree with NUTS at N=50 | 0 | 1 | - |
 | euclid_dr1 | Does the sep1 DR1 delivery reproduce the prelim ten lenses | 0 | 1 | - |
 | euclid_dr1_prelim | Do ordered Nautilus reruns reproduce the euclid catalogue values? | 1 | 0 | 2026-09-10 |
+| euclid_sersics | Core two-stage run of the 100 sersic lenses | 0 | 1 | - |
 | 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 | - |
 | subhalo_validation | Does rectangular_adapt false-detect a subhalo on the clumpy source | 0 | 5 | 2026-09-07 |
@@ -235,6 +236,24 @@ Finish: write today's journal entry, rewrite wiki/project/state.md, and return (
 
 [tasks/euclid_dr1_prelim/](https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/euclid_dr1_prelim/)
 
+### euclid_sersics
+
+**Local** `/mnt/c/Users/Jammy/Science/euclid_sersics`
+
+**RAL** `/mnt/ral/jnightin/euclid_sersics`
+
+active · ral partition · tasks: submitted 1
+
+<details><summary>📋 <a href="tasks/euclid_sersics/core_two_stage_100.md">Core two-stage run of the 100 sersic lenses</a> — <b>submitted</b> — runs 342696 · wall 0:00 of 36:00 (0%)</summary>
+
+```
+cd /mnt/c/Users/Jammy/Science/euclid_sersics && hpc/sync jobs
+```
+
+</details>
+
+[tasks/euclid_sersics/](https://github.com/PyAutoLabs/PyAutoCortex/tree/main/tasks/euclid_sersics/)
+
 ### ic50_workspace
 
 **Local** `/mnt/c/Users/Jammy/Science/ic50_workspace`
diff --git a/projects.yaml b/projects.yaml
index f54ef71..5552b0e 100644
--- a/projects.yaml
+++ b/projects.yaml
@@ -115,6 +115,20 @@ euclid_dr1:
   status: active
   note: "born 2026-09-11: the actual DR1 run on the sep1 segmentation delivery (Segmentation.zip = 3 nested batch zips, 15032 tiles, every tile already in the pipeline dataset layout); born by cloning euclid_strong_lens_modeling_pipeline main with config/visualize colormap back to autoarray; first task reruns the ten euclid_dr1_prelim lenses on the new data; remote none until the human creates one"
 
+euclid_sersics:
+  remote: none
+  local_path: /mnt/c/Users/Jammy/Science/euclid_sersics
+  ral_root: /mnt/ral/jnightin/euclid_sersics
+  mirror: none
+  sync_cli: hpc/sync
+  sync_verbs: [push, pull, logs, sync, push-data-init, pull-full, status, submit, push-submit, jobs, sacct, cancel, wait-and-pull, tail, du, check, clear-logs]
+  ledger: wiki/project/state.md
+  assistant: autolens_assistant
+  witness_file: results/**/*.json
+  partition: ral
+  status: active
+  note: "born 2026-09-11: Sersic-index pile-up investigation — 100 unanimous-success lenses (euclid inspection) on the sep1 segmentation data; core two-stage CPU run first, then four sersic_lens_model variants (baseline / n∈[0.5,10] / central noise inflation / Sersic+point MGE) reuse its vis_lp; remote none until the human creates one"
+
 analytic_gaussian:
   remote: none
   local_path: /mnt/c/Users/Jammy/Science/analytic_gaussian
diff --git a/tasks/euclid_sersics/core_two_stage_100.md b/tasks/euclid_sersics/core_two_stage_100.md
new file mode 100644
index 0000000..7a1733e
--- /dev/null
+++ b/tasks/euclid_sersics/core_two_stage_100.md
@@ -0,0 +1,109 @@
+# Euclid_sersics — the core two-stage run of the 100 Sersic-sample lenses
+
+Project: euclid_sersics
+Summary: Core two-stage run of the 100 sersic lenses
+State: submitted
+Gates:
+Witness: 100/100 array tasks complete vis_lp and vis_pix within 36 h; aggregator yields 100 lens_mass rows; no task aborted
+Budget: 36:00
+Runs: 342696
+Ruling:
+Review-minutes: 25
+Epic: euclid-sersics
+Filed: 2026-09-11
+
+## Question
+
+The June `euclid` catalogue fits one `lp_linear.Sersic` per lens under `Uniform(0.8, 5.0)` with
+hard limits, and the returned Sersic index **piles up against the top of the prior**: over 2983
+lenses the median n is 4.80, 42 % sit above 4.9, and **nothing** sits at the lower edge. In the
+unanimously-inspected subset the pile-up is *worse* (median 4.84), which is the opposite of what
+a bad-fit signature does. `scripts/simulator.py` in the pipeline already names it: "Real Euclid
+fits routinely return sersic_index = 5 … an artefact".
+
+The project that settles it (`euclid_sersics`) will run **four Sersic-stage variants** —
+`baseline` (today's prior), `wide_n` (n ∈ [0.5, 10]), `central_noise` (noise inflated inside the
+seeing disk) and `sersic_point` (Sersic + point MGE) — on one fixed sample, to separate "the
+prior is the bound" from "something biases n upward" (a too-broad model PSF or an unmodelled
+nuclear point source both push n up; correct-PSF seeing pushes it *down*).
+
+**This task is not that comparison.** It is the run every one of those variants stands on: the
+**core two-stage fit of all 100 lenses**, `vis_lp` then `vis_pix`, on the route `euclid_dr1` is
+already using. It produces the `vis_lp` result the variants reuse rather than re-derive, and the
+`baseline` lens model the other three are read against. Its question is therefore an
+*engineering* one, and it is asked first because a variant comparison built on a sample that
+half-completed is worthless:
+
+**Does the 100-lens sample run clean, end to end, on the sep1 delivery?**
+
+D1 is resolved: **both stages run.** `vis_pix` is not skipped. The pixelized source is what makes
+the lens-light Sersic honest — a variant that moves n by absorbing source flux into the lens
+would be invisible on a `vis_lp`-only run.
+
+## The sample
+
+100 lenses, chosen by a rule written down so it cannot drift, because every variant must run on
+**these exact lenses**: from `euclid`'s inspection set (2990 lenses, 1–7 human votes each, 1293
+unanimous Success) — **unanimous Success, present in both `lens_sersic.csv` and `lens_mass.csv`,
+sorted by `total_valid_votes` descending then `object_id` ascending, first 100**. That is 16
+lenses at 7/7 votes and 84 at 6/6, median June n = 4.917, 22 of them interior (n < 4.5), min
+0.859, max 4.999.
+
+The data is the September-1 segmentation delivery, streamed out of `euclid_dr1`'s
+`Segmentation.zip` into `dataset/dr1_sep1_sersics/` — 100 tile folders, 11 files each, all
+verified present, two spot-checked through `util.load_vis_dataset`.
+
+June `object_id` → sep1 tile is a **coordinate join, never a name match**: tile names were parsed
+(`RA`/`DEC` as 3 integer + 10 decimal digits of degrees) and nearest-neighboured against all
+15 032 sep1 names, a unique match required inside 3". **100/100 matched, 95 names identical, the
+5 that moved by 0.68–0.92" (max 0.921")**. `sample/lens_map.csv` in the project is that table and
+is the only legitimate join back to the June catalogue.
+
+## Witness
+
+100/100 array tasks complete vis_lp and vis_pix within 36 h; aggregator yields 100 lens_mass rows; no task aborted
+
+## Where to look
+
+- `euclid_sersics` (project row `euclid_sersics`): `output/dr1_sep1_sersics/<tile>/` once pulled
+  — one subdirectory per stage per lens; the data is `dataset/dr1_sep1_sersics/` and the route is
+  `hpc/batch_cpu/submit_initial_lens_model_two_stage` (array `0-99`, partition `ral`, 8 CPUs,
+  64 GB, 36 h)
+- `euclid_sersics` `sample/lens_map.csv` — the June `object_id` ↔ sep1 tile provenance table, and
+  `sample/lenses_100_june.csv` — the June-side list with votes, θ_E, n and R_eff
+- `euclid_sersics` `wiki/project/state.md` — this project's own ledger (the commentary; the
+  ruling of record is the Cortex's)
+- `euclid_dr1` (project row `euclid_dr1`): the same two-stage route on the same delivery, run
+  342650 — the reference for what a healthy run of this script looks like
+- `autolens_assistant/skills/euclid_prepare_data.md`, `euclid_setup_pipeline.md`,
+  `euclid_model_lens.md`, `euclid_hpc_runs.md` — how this pipeline is driven
+
+## Notes
+
+- 2026-09-11 — **the project was born for this task.** `euclid_sersics` is a fresh clone of
+  `euclid_strong_lens_modeling_pipeline` at **`dbbb2a5`** — deliberately the *same base commit* as
+  `euclid_dr1`, plus that project's two local deltas (`activate.sh`'s `AUTOLENS_ASSISTANT` export;
+  `config/visualize/general.yaml` back on the library default `colormap: autoarray`). Any
+  difference between the two projects' fits is therefore the sample or the model, never the
+  pipeline. The clone's `origin` is the pipeline repo and it has local commits that are never
+  pushed.
+- 2026-09-11 — **the submit script differs from `euclid_dr1`'s in five things and nothing else:**
+  the job name, `PROJECT_PATH`, `sample=dr1_sep1_sersics`, `--array=0-99`, and the dataset list,
+  which is read from the sample directory
+  (`mapfile -t datasets < <(ls $PROJECT_PATH/dataset/$sample/ | grep '^Tile' | sort)`) instead of
+  100 inline tile names, behind a guard that exits non-zero unless exactly 100 are found. The
+  guard matters: a partial `hpc/sync push` would otherwise shift every array index and fit the
+  wrong lens under each one, silently.
+- 2026-09-11 — **the pipeline work the variants need does not exist yet**: a `--variant` flag in
+  `scripts/sersic_lens_model.py`, a central-noise-inflation hook in `util.load_vis_dataset`, and
+  an `mge_point_model_from` for the Sersic+point variant. That is a separate PR against the
+  pipeline repo, and no variant can be submitted before it merges. This task is unblocked by none
+  of it.
+
+## Runs
+
+- 342696: submitted — ral — submitted 2026-09-11 — wall 0:00
+
+## Ruling
+
+(none)