# PHerc0332 spiral fit on villa main: the MSEM6 run (main/)

The best PHerc0332 spiral fit made with the fitter on the main branch of github.com/ScrollPrize/villa (commit f29334cd8),
two seeds, 2026-11-16/17. `../repro/` covers the earlier reference fits made with the frozen code tree; this folder
covers the fit on main.

| path | what |
|---|---|
| m6seam22.vcobj | the input object store (schema 22, 16.7 GB): patches, strips, tracks, fibres, anchors, outer shell |
| inputs/spiral-inputs.json | scroll frame + the paths of the umbilicus and the Lasagna manifest (edit both paths) |
| inputs/overrides_{Va,Vb}_MSEM6_s{1,2}_*.json | settings: Va = 10,000 steps from scratch, Vb = resume to 25,000 |
| inputs/overrides_{Va,Vb}_SMK22_s1_*.json | a 200 / 400 step smoke test of the same chain |
| scripts/run_main.sh | START HERE: path-free runner. `score` rescores a published result and compares it; `fit` runs Va, Vb, the 0-step held-out pass and the frozen scorer from the downloaded files (usage in its header) |
| scripts/chain_main.sh | the workstation chain the published fits were made with (workstation paths at the top; kept as the record) |
| scripts/held_ids.py, verify_held.py, holdout_uuids.json | held-out patch bookkeeping used by the chain |
| results/s{1,2}/ | Vb checkpoint, scores_trim.json, satisfied_fitted.json and the chain log of each seed |
| SHA256SUMS | sha256 of every file here |

Also needed: `../umbilicus/umbilicus.json`, the public Lasagna prediction
(vesuvius-challenge-open-data, PHerc0332/representations/predictions/lasagna/20251211183505-lasagna-20260419180421-L2/),
and for scoring `../slice_lines/`, `../anchors/` and `../eval/frozen_source_s1_f327248db2c1f694.tar.zst` (see `../repro/README.md`).

Work on a copy of the store: the first start writes fibre-patch and point-patch links into it (about an hour); keep one
cache directory per store. Published scores (seed 1 / seed 2): f025 56.97 / 57.12, f05 85.81 / 86.03, held patches
70.27 / 69.97, held fibre off 23.34 / 22.74. Replayed on 2026-11-18 in an empty folder from the published files only (fresh clone at f29334cd8, `uv sync`, this bundle, the open-data Lasagna nx / ny / grad_mag zarrs):
`run_main.sh score` on results/s1 = 4,139 score leaves identical; `run_main.sh fit --prefix SMK22 --seeds 1` ran end to end, TSCORE at 200 / 400 steps f025 0.1451 / 0.1662,
held patches 0.2566 / 0.2721, held fibre off 0.7664 / 0.7341 (the same smoke on the workstation tree: 0.1450 / 0.1663, 0.2557 / 0.2712, 0.7675 / 0.7373).
The full 25,000-step fit was not repeated this way, and the replay ran on the original workstation, so another host is untested. One Python environment
(the clone's spiral-fitting/.venv) runs both the fit and the frozen scorer.
