A 3D model pipeline, pushed by GPT-6.2-Astra

agent
Codex 0.167.2
model
gpt-6.2-astra
harness
Codex
started
2026-11-17 22:44 UTC
took
61 min · 340 turns

user

+0:00

Your brief is in ~/context/brief-lowmarch-full.md. Read it in full and carry it out.

agent

+0:03

I’ll read the full brief and check the workspace before starting the requested work.

  • exec_commandcat ~/context/brief-lowmarch-full.md (+1 more)

    command

    cat ~/context/brief-lowmarch-full.md
    
    pwd && ls -la && rg --files --hidden -g '!.git' -g 'AGENTS.md'

    result, 4 lines

    Script completed
    Wall time 0.1 seconds
    Output:
    [{"status":"fulfilled","value":{"chunk_id":"a7194c","wall_time_seconds":0.000004001,"exit_code":0,"original_token_count":941,"output":"You are running as GPT-6.2-Astra (gpt-6.2-astra) at effort high, in a session of your own. The owner, severian (he/him), asked to see what GPT-6.2-Astra would do with one of their lines of work, lowmarch, and to see one visual artifact you can make with what the project already has. The project's own session runs on a different model and is not to be touched; you are a separate, parallel look.\n\n## What lowmarch is\nA code-built anime character pipeline for a tactics game: VRoid bases reshaped in code, costumes, armour and hair generated in code, drawn by the project's own renderers (a painted three.js renderer is the target; the pixel renderer is abandoned). On 11-08 the owner decided to restart with one character from scratch, to find a faster, better way of working together; they haven't started yet.\n\nRead, in this order:\n1. `docs/handbook.md` in your worktree (entry point: each rule marked earned or guessed; guessed rules don't bind the restart)\n2. `~/context/reflect/lowmarch.html` (the 11-08 reflection; the box at the top is the restart plan)\n3. `~/context/lowmarch.md` (state, how to run things; numpy is in the user site, so run python3 without -I)\n4. `docs/portrait-guide.md`, `docs/vroid.md`, and whatever else in the repo you judge useful.\n\n## Your workspace\n- Your current folder is the git worktree `~/work/lowmarch` (branch `astra62`, made for you from master a97f25f). Work only here; you may commit here. Never touch any other repo.\n- Data lives in `~/.local/share/lowmarch/`: read it, don't modify it.\n- Output folder: `~/out/` (copied to the owner's review route afterwards).\n- Licences: the Yin/Yang rigs are personal-use study material (keep them out of the visual); FE8 references are study only, never traced or shown; VRoid CC0 bodies and the AvatarSample faces are fine for renders.\n\n## Rules\n- Headless only: no windows on the owner's desktop. Never run anything that drives Windows programs (no tools/winvroid, no powershell.exe, cmd.exe or anything under /mnt/c). Browser rendering via headless Chromium only (Playwright is at `~/context/sortie-node_modules/playwright`; page-side code compiled by tsx can throw \"__name is not defined\", so keep page-side code as plain .js evaluated as text). Blender only with `-b`.\n- Don't install packages. Don't run the full test suite (a test run once ate 36 GB); targeted checks only.\n- Nothing runs on the C1 machine; no ssh. No network uploads, no publishing anywhere. Never push.\n- Edit files only in your worktree and your output folder. Never push.\n\n## Deliverable\nOne self-contained HTML page, `index.html` in your output folder, with its images (or a short clip) beside it. The owner finds wordy interfaces tiring: organise, don't narrate; short labelled sections, plain words, no em dashes. At the top: \"GPT-6.2-Astra (gpt-6.2-astra), effort high\" and the date.\n1. **Your read of this line of work:** what's working, what isn't, what you'd keep and drop, and what you would do differently in the restart. Give opinions plainly; disagree with the reflection page where you do.\n2. **What you'd do next** if you ran the restart's first session: a short, concrete sequence.\n3. **The one visual:** a single visual artifact made with what the project already has (its tools, kits, renderers, models, data). Your choice of what it is; pick the thing that best shows what you'd do with this work. Show it large, then say in a few lines how it was made (the commands, wall time) and what you'd change next.\nKeep it to one visual; don't build a gallery. Aim to be done in roughly 90 minutes of work.\n\nWhen done, end your turn with: the page path, a 5-line summary, and any commits on your branch. If something only the owner can decide blocks you, say so and stop rather than guessing.\n"}},{"status":"fulfilled","value":{"chunk_id":"92889a","wall_time_seconds":0.000004821,"exit_code":1,"original_token_count":123,"output":"~/work/lowmarch\ntotal 36\ndrwx------ 7 runner runner 4096 Nov 17 22:43 .\ndrwx------ 3 runner runner 4096 Nov 17 22:43 ..\n-rw------- 1 runner runner   49 Nov 17 22:43 .envrc\ndrwx------ 7 runner runner 4096 Nov 17 22:43 .git\n-rw------- 1 runner runner  124 Nov 17 22:43 .gitignore\ndrwx------ 5 runner runner 4096 Nov 17 22:43 art\ndrwx------ 2 runner runner 4096 Nov 17 22:43 docs\ndrwx------ 5 runner runner 4096 Nov 17 22:43 review\ndrwx------ 5 runner runner 4096 Nov 17 22:43 tools\n"}}]
  • exec_commandcat docs/handbook.md

    command

    cat docs/handbook.md

    result, 230 lines

    Script completed
    Wall time 0.1 seconds
    Output:
    # Lowmarch handbook
    
    For whoever picks this up cold. Read this first; everything else is detail it links to.
    Written 2026-11-08 by the lowmarch session, at the owner's request, with no owner review yet.
    
    ## Where things stand
    
    - **What it is:** anime characters built in code for a tactics game. The owner named Unicorn Overlord as the style on
      10-08; lately they compare against the Yin and Yang rigs. They can't model or draw, and review pictures.
    - **The owner is restarting** with one character, from scratch, to work out a faster, better process. They start it.
      Don't build character work before then. The plan is the box at the top of
      `~/citadel-render/workers/handoff/reflect/lowmarch.html`; the retro below that box is background.
    - **What exists:** 14 bare T-posed bodies with VRoid faces and hair, 14 expressions each, a painted renderer
      (`review/vroid-02/index.html`). No dressed character on these bodies, no game frame, no game code.
    - **Review pages:** cast expressions [redacted: a link to a private page], hair research
      [redacted: a link to a private page]. Artifacts belong to the account that made them; if you can't
      update one, ask the lead (personal/citadel-lead-2).
    - **More detail:** `docs/map.md` (what's current or stale, where big files sit), `docs/todos-2026-11-08.md` (open
      todos sorted), `~/citadel-render/workers/handoff/lowmarch.md` (history).
    
    ## Earned and guessed
    
    Every rule below carries a mark, so old rules can't quietly bind the restart.
    
    - **Earned:** the owner saw it and said so, or a failure was measured and the fix held.
    - **Guessed:** I reasoned it, took it from study, or only my own checks passed it. It binds nobody.
    
    An earned rule can still be overturned, as long as someone says so out loud. "On old bases" means earned on work now retired.
    
    ## Reading the owner's feedback
    
    - **Cohesion is about the whole, not a part.** "There has to be cohesion in it" (10-31) arrived about lashes, and I made
      it a lash rule. The same complaint runs all month: "how the features sit on the face" (10-07), "pieces look floating"
      (10-08), "head surgery" (10-10), "doesn't fit her body" (10-12). Read it as: this doesn't belong to this person.
    - **"Hard to explain" means stop naming causes.** "I can't explain it" (10-30) got "I think I can name the something off"
      from me, then a build. Make options that vary what might be wrong, and let them point.
    - **They see more than they can say, and they're usually right.** Asked to check thin arms and thighs, the survey data
      agreed (lesson 27). Take a marked-up screenshot literally: the orange nape line was correct (lesson 14).
    - **Images are data.** "Cursed", "scar", "surgery" say what it feels like, not what the mesh did.
    - **Sort praise by size.** Whole-thing verdicts: Saber (10-08), the bodies "look perfect" (10-12). "I like the shadows
      now" is a part. "I guess that works for now" (10-12) isn't approval. "Go for your list" is delegation, not agreement.
    - **"Where next?" and "I wonder how else" are requests for a map** and for how far this goes. Answer plainly, including
      "I don't know if this route gets there". I didn't, and I said "the shading layers are what close the gap" (10-30). Don't.
    - **Pivots are usually theirs.** If they're doubting the base, something's been wrong for a while. Say it first next time.
    - **"I'm lost" (10-30) and "I just need the sheets" (10-31)** mean fewer words. Skip ledger numbers during review.
    - **Money:** they've offered to buy tools (10-30), and declined a cost they hadn't noticed (11-02). Free path first, money
      in line one, nothing spent without a fresh go.
    
    ## Showing options fast
    
    The ladder, cheapest first. Stay as low as you can, and only build the owner's pick properly.
    
    | Change | Tool | Cost |
    |---|---|---|
    | Camera, colours, rough blocks on a stand-in | `tools/session/frame.py` | 3 options in 5 s (measured 11-08; 6 s cold, 3 s warm on 11-09) |
    | Painted renderer's page only | `tools/shots/arm_shoot.py`, on a copy of the site | 8 s a body (measured 11-08) |
    | A face or expression | `tools/shots/variant_build.py`, then shoot | roughly 1 min a body (older note, not re-measured) |
    | Any file in cage's cache key (below) | `python3 -I tools/cage.py` | about 10 min for all 14 (10-31) |
    
    **First session:**
    
    ```bash
    S=$(python3 tools/session/rows.py new first --ref PICTURE) # the session folder
    python3 tools/session/frame.py --info arm_f_slender        # height and landmarks, for placing blocks
    cp tools/session/smoke_spec.json $S/frame.json             # then edit camera and options to match the picture
    python3 tools/session/frame.py $S/frame.json $S/r1         # one PNG per option
    python3 tools/session/rows.py add "Big shapes" $S/r1/*.png --note "under 100 words"
    cw view $S/latest.png                                      # the row, beside their picture
    python3 tools/session/rows.py answer "B, shorter hair"     # timed; rows.py log prints the rounds
    ```
    
    Publish with the Artifact tool from the same path each time, so the link stays put. Use `$S/index_noref.html` unless the
    owner has said a private page may hold their picture (see "Things that don't bend"). `--blind` shuffles a row's labels.
    
    **The spec** (`tools/session/smoke_spec.json` is a working example, and a test, not a design):
    
    - `figure`: a built armory body. It's a stand-in; `look.hide: ["hair"]` drops its hair.
    - `frame`: `w`, `h`, `ground_y` (where the feet stand), `horizon`, `background`, `ground`, and `camera`: `kind` (`ortho` or
      `persp`), `fov`, `yaw`, `pitch`, `figure_h` (share of the frame's height), `figure_x`.
    - `look`, overridden per option: `skin`, `shade`, `hair`, `hair_shade`, `cloth`, `cloth_shade`, `ink`, `outline` (metres),
      `threshold`, `light`, `features` (flat eyes and lines), `hide`.
    - `options[].blocks`: `ball`, `tube` (points, radii, `flat`) and `loft` (rings), in metres from a landmark in `at`.
      +x is the model's left, +y up, +z front.
    
    `tools/session/smoke.sh` tests the kit end to end in about 7 s (9 s cold after a restart, 11-09). It needs the built armory site and network for three.js.
    The frame page is flat two-tone on purpose. It isn't the game's look, so don't judge finish on it.
    
    ## How to run
    
    "Ran" means run on 11-08 at master `c4f1015`. numpy is in user site-packages, so `-I` works for `cage.py` but breaks
    `lash_check.py`.
    
    | Command | What it does | Status |
    |---|---|---|
    | `python3 -I tools/cage.py [NAME...]` | builds armory bodies, cached | ran, one body, cached |
    | `LOWMARCH_SITE=armory python3 review/vroid-02/build.py` | builds the armory into `review/out/armory/site/` | not re-run; built 10-31 |
    | `python3 tools/lash_check.py [NAME...]` | open lash against closed lids, on meshes | ran, one body, 8 s |
    | `ROWS=0 python3 tools/shots/arm_shoot.py SITE OUT NAME...` | shoots a site copy, headless | ran, one body, 8 s |
    | `tools/shots/all14.sh OUT` | 6 expressions on all 14 bodies | copied, not re-run; wipes OUT first |
    | `python3 tools/regress.py NAME [TAG]` | geometry diff against a published version | not re-run |
    | `python3 tools/winvroid/winvroid.py bases\|samples` | drives VRoid Studio on Windows | blocked until the owner allows V.dll (#1105); takes over the PC, needs their go |
    
    - **Cage's cache key** hashes `cage.py, fit.py, proportions.py, makehuman.py, anthro.py, heads.py, vroid_portrait.py,
      collide.py`. Touching any of them rebuilds every body. `export_glb.py` isn't in the key, but its changes need the GLBs rebuilt.
    - **Browsers:** headless only. Headed windows pop up on the owner's desktop.
    - **Data** lives outside the repo, in `~/.local/share/lowmarch/`. Paths are in `docs/map.md`.
    - **The skeleton trap:** the bind pose holds the arms down while the mesh is T-posed (lesson 26). Arm bones' world
      positions don't mark the visible arm. `frame.py --info` reports arms from the skin instead.
    
    ## The rules
    
    Each links to its lesson in `docs/portrait-guide.md` (L12 is lesson 12).
    
    ### Working with the owner
    
    | Rule | Mark | Basis |
    |---|---|---|
    | Ask "is this good?" of a whole character, not a part | Earned | Saber 10-08, bodies 10-12; parts get only local praise |
    | Put a reference beside ours | Earned | "there's no muscle bulge" 10-09, L19 |
    | Blind rows for close calls | Earned | Sharpest verdict of the month, 10-30 |
    | Check the view the owner will see | Earned | L12; wrong renderer on 10-30 |
    | A fix that moves another part is a regression until compared | Earned | L13, L28 |
    | Study first, on anything new | Earned | Chin shadow took 3 builds before research (L33 to L35); sleeves and Saber jumped with study first |
    | Tuned on one face means tuned on one face: check the rest before calling a range done | Earned | Mouths broke on the other 12 (L40) |
    | Judge only in the game frame | Guessed | Retro; never tried |
    | Rough blocks before detail | Guessed | Retro; never tried |
    | One change a row, 3 or 4 options, under 100 words | Guessed | Retro; never tried |
    | Two rounds with no visible change: stop and say so | Guessed | Retro; never tried |
    
    ### Bases and bodies
    
    | Rule | Mark | Basis |
    |---|---|---|
    | VRoid whole, shaped by a MakeHuman cage, own head kept | Earned | "it looks perfect" 10-12, L25 |
    | Head grafts don't pass | Earned | "surgery" 10-10, L17, L25 |
    | Read flesh against survey percentiles at the body's height and waist | Earned | Matched the owner's eye, L27 |
    | Ranges measured on armoured figures don't fit bare bodies | Earned | L24, L27 |
    | Tie a hand to body height | Earned | L28 |
    | Don't force ratios from 2D art onto a 3D face | Earned | "too sharp" 10-12, L30 |
    | A VRoid face as built is good enough | Guessed | The verdict was only "I guess that works for now" |
    | No better shippable base than VRoid exists | Guessed | Research 11-08; candidates not all tried |
    
    ### Faces
    
    | Rule | Mark | Basis |
    |---|---|---|
    | Shadow made by bending mesh normals reads as a scar | Earned | 10-30, L33, L34 |
    | Chin shadow cast from the jaw in screen space | Earned | "I like the shadows now" 10-30, L35 |
    | Proxy face normals for a clean edge | Earned | Kept 10-30, L32 |
    | One lash size per character, open and closed | Earned | Owner's rule 10-31, L38. A sample of cohesion, not the whole of it |
    | A shut lid is one solid stroke | Earned | "the eyes look good now" 10-31, L38 |
    | Mouth inside from VRoid's texture, anime palette | Earned | Blind pick 10-30, L39 |
    | One-sided shapes for a smug look | Earned | Blind pick 10-30, L39 |
    | Age takes more than lines | Earned | "hardly reflect old men" 10-08; never solved, #768 |
    | 15% lash limit, split-line check | Guessed | My number, L38 |
    | Shading line at the head's middle, front stays lit | Guessed | My choice, L32. At 35° the far side reads flat. Test this first |
    | Force shadow on the far side of the face | Guessed | L03 on a retired mannequin; research agrees, never tried here |
    | Brows drawn through the bangs | Guessed | The owner doubted it: "not sure if eyebrows are supposed to show up on top of hair" (10-30) |
    | Hold the mouth 2 mm off the chin; teeth only above 8 to 11 mm | Guessed | L40, no verdict |
    | One expression list for every face | Guessed | Research says per character; the open mouths look alike |
    | Flat eye plane, per-angle tables | Guessed | L03. Arc System Works calls a tilted flat eye a mistake |
    | FE8 house style: purple ink, tone rules | Guessed | From the pixel era, now abandoned |
    
    ### Hair
    
    | Rule | Mark | Basis |
    |---|---|---|
    | Few clumps of different widths, not even points | Earned | "seven pointed bangs" 10-08, L14 |
    | Nape level with the bottom of the earlobe | Earned | Owner's orange line, L14 |
    | Sample-style hair beat the kit's locks | Earned | 10-08, #735 |
    | Hair must not sink into the body | Earned | 10-13. The fix (L31) passed only my replay, not the owner's eye |
    | Dark layer under the locks closes gaps | Guessed | Proposed 10-31, not built |
    | Hair keeps its own normals | Guessed | My eye, L32. Studios copy from simple shapes, so retest |
    | Back, then sides, then front | Guessed | My 10-31 report; no source confirms it |
    
    ### Clothes and armour
    
    | Rule | Mark | Basis |
    |---|---|---|
    | Fit by rays to what's underneath; nothing floats | Earned, old bases | "floating" 10-08, then Aldric, L07 |
    | An under-layer next to skin isn't skin-toned | Earned, old bases | 10-08, L10 |
    | Loose sleeves; no muscle through cloth | Earned, old bases | 10-09, L19 |
    | Check with arms out; cast waist rings on the torso only | Earned, old bases | L12, L13 |
    | Layered skirts share weights | Earned, old bases | Owner spotted the underskirt, L13 |
    | Long skirts posed, not simulated | Earned, old bases | Every free-cloth variant broke, L11; the owner never judged it in motion |
    | Garment offsets in millimetres | Guessed | L07, tuned on old bases |
    | Folds as ripples built into the garment | Guessed | L19. Studios place folds by hand for a pose |
    
    ### Rendering
    
    | Rule | Mark | Basis |
    |---|---|---|
    | The painted renderer is the target; pixels are abandoned | Earned | Owner 10-30 |
    | Flattice waits until scenes have a camera | Earned | Owner 10-30 |
    | A colour ramp per skin tone | Earned | Owner asked for the range; L36 |
    | Orthographic cameras everywhere | Guessed | My default; the owner never chose it. Their picture decides |
    | Headlight from the upper left, near-white default skin | Guessed | Never chosen |
    | Paint numbers measured off Unicorn Overlord (L10) | Guessed | Taken from paintings; the reference has since moved |
    
    ## Things that don't bend
    
    These are facts or standing instructions, not lessons.
    
    - **Yin and Yang rigs:** personal-use licence. Study only, never ship.
    - **References:** never traced, shipped or committed (`docs/study.md`). The same doc also keeps them off published pages.
      I wrote that clause and it's guessed; the owner hasn't ruled on a private page. The retro's box says "your picture on
      the left", so ask them. Until then, the published sheet leaves it out.
    - **VRoid licence:** read by hand on 11-09, in `docs/vroid.md` ("Licence, read 2026-11-09", #1443). Baked characters in a
      sold game, public pages and renders are allowed, no credit needed. A player-facing creator needs pixiv's licence. Our
      exports' `Redistribution_Prohibited` fields were set at our export and don't bind us. Open: unedited A to Z textures in
      a paid game. The owner decides what ships.
    - **Blender:** nothing gets installed into it without the owner's go. A trial plan is enough.
    - **The repo has no remote.** The owner hasn't decided, so don't add one. Commit locally.
    - **Writing:** docs terse, no em dashes. Pages the owner may share carry no employer names. Public write-ups call this
      "a 3D model pipeline".
    - **Contact** for anything outside lowmarch: personal/citadel-lead-2.
    
    ## What I'd watch for
    
    These are my own weak spots, from the retro.
    
    - I judge my own pictures badly. Treat my checks as a floor, not a verdict.
    - I reach for a cause and a build when the honest answer is "show me which".
    - I write a checker after the owner catches a fault, and it tends to measure the part, not the whole.
    - Lessons I've written don't stop me repeating a mistake. L03 said to force shade on the far side of the face, and L32
      undid it without noticing. Read the marks above before relying on a lesson.
    
  • exec_commandcat ~/context/reflect/lowmarch.html

    command

    cat ~/context/reflect/lowmarch.html

    result, 300 lines

    Script completed
    Wall time 0.0 seconds
    Output:
    Warning: truncated output (original token count: 22028)
    Total output lines: 441
    
    Warning: truncated output (original token count: 86684)
    Total output lines: 442
    
    <!doctype html>
    <html lang="en">
    <head>
    <meta charset="utf-8">
    <meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
    <meta name="claude:viewport-fit-sides" content="cover">
    <title>Lowmarch, Looking Back</title>
    <style>
    /* Layout: one reading column, tables in their own scroll boxes. A cel ramp (lit, shade, deep, ink) from the portrait guide is the only flourish; ink is purple, never black, as the house style says. */
    :root {
      --bg: #f5f3f7; --panel: #fbeee7; --panel2: #ebe7f0; --fg: #2b2336; --muted: #675f73; --rule: #d8d2e0;
      --ink: #3c2a55; --accent: #6a3f8f; --guess: #8a4b1d; --good: #2c6a4c; --hold: #8a4b1d; --no: #8f2f3c;
      --lit: #f8e1d5; --shade: #c6a89b; --deep: #73524f;
      --head: "Avenir Next", "Segoe UI Variable Display", "Segoe UI", system-ui, sans-serif;
      --body: "Segoe UI Variable Text", "Segoe UI", system-ui, -apple-system, sans-serif;
      --mono: "Cascadia Mono", ui-monospace, "SFMono-Regular", Menlo, monospace;
    }
    @media (prefers-color-scheme: dark) { :root:not([data-theme="light"]) {
      --bg: #17141c; --panel: #2a211f; --panel2: #231f2b; --fg: #ece7f1; --muted: #a79fb3; --rule: #383243;
      --ink: #d8c8ec; --accent: #c4a2e6; --guess: #e6ae7a; --good: #86cfa8; --hold: #e6ae7a; --no: #ee9aa6; color-scheme: dark } }
    :root[data-theme="dark"] {
      --bg: #17141c; --panel: #2a211f; --panel2: #231f2b; --fg: #ece7f1; --muted: #a79fb3; --rule: #383243;
      --ink: #d8c8ec; --accent: #c4a2e6; --guess: #e6ae7a; --good: #86cfa8; --hold: #e6ae7a; --no: #ee9aa6; color-scheme: dark }
    html { -webkit-text-size-adjust: 100%; }
    body { margin: 0; background: var(--bg); color: var(--fg); font: 400 16px/1.55 var(--body); }
    .wrap { max-width: 880px; margin: 0 auto; padding-block: 26px 70px;
      padding-left: max(16px, env(safe-area-inset-left, 0px)); padding-right: max(16px, env(safe-area-inset-right, 0px)); }
    .ramp { display: grid; grid-template-columns: 5fr 3fr 2fr 1fr; height: 8px; border-radius: 4px; overflow: hidden; margin-bottom: 20px; }
    .ramp i { display: block; }
    h1, h2, h3 { font-family: var(--head); color: var(--ink); text-wrap: balance; line-height: 1.15; margin: 0; }
    h1 { font-size: 2.2rem; font-weight: 700; letter-spacing: -0.01em; }
    h2 { font-size: 1.5rem; font-weight: 700; margin-top: 52px; padding-top: 16px; border-top: 2px solid var(--ink); }
    h3 { font-size: 1.08rem; font-weight: 700; margin-top: 28px; }
    p { margin: 10px 0 0; max-width: 72ch; }
    ul, ol { margin: 10px 0 0; padding-left: 1.2em; display: grid; gap: 6px; }
    li { max-width: 72ch; }
    a { color: var(--accent); text-underline-offset: 2px; }
    a:focus-visible, summary:focus-visible { outline: 2px solid var(--accent); outline-offset: 2px; border-radius: 2px; }
    q { quotes: "\201C" "\201D"; }
    .meta { color: var(--muted); font-size: 0.92rem; margin-top: 8px; }
    .jump { display: flex; flex-wrap: wrap; gap: 6px 16px; margin-top: 12px; font-size: 0.95rem; }
    .box { margin-top: 20px; background: var(--panel); border-radius: 10px; padding: 16px 18px; }
    .box h2, .box h3 { margin: 0; padding: 0; border: 0; font-size: 1.1rem; }
    .box.alt { background: var(--panel2); }
    .box ol, .box ul { gap: 8px; }
    .tally { display: grid; grid-template-columns: repeat(auto-fit, minmax(118px, 1fr)); gap: 10px 16px; margin-top: 14px; }
    .tally div { min-width: 0; border-top: 1px solid var(--rule); padding-top: 6px; }
    .tally b { display: block; font: 700 1.3rem/1.2 var(--mono); color: var(--ink); font-variant-numeric: tabular-nums; }
    .tally span { font-size: 0.82rem; color: var(--muted); line-height: 1.3; display: block; }
    figure { margin: 18px 0 0; }
    figure img { display: block; width: 100%; height: auto; border-radius: 6px; }
    figcaption { color: var(--muted); font-size: 0.88rem; margin-top: 7px; max-width: 76ch; }
    .tbl { overflow-x: auto; margin-top: 12px; }
    table { border-collapse: collapse; width: 100%; font-size: 0.92rem; }
    .tbl.w table { min-width: 680px; }
    th, td { text-align: left; vertical-align: top; padding: 8px 10px 8px 0; border-bottom: 1px solid var(--rule); }
    th { font-weight: 600; font-size: 0.74rem; letter-spacing: 0.05em; text-transform: uppercase; color: var(--muted); }
    td:first-child { font-weight: 600; color: var(--ink); }
    .two { display: grid; grid-template-columns: 3fr 2fr; gap: 6px 26px; }
    .two > div { min-width: 0; }
    .two h4 { margin: 12px 0 0; font: 600 0.74rem/1.3 var(--body); letter-spacing: 0.05em; text-transform: uppercase; color: var(--muted); }
    .two ul { font-size: 0.95rem; }
    @media (max-width: 640px) { .two { grid-template-columns: 1fr; } h1 { font-size: 1.85rem; } }
    .g { font: 500 0.68rem/1 var(--mono); color: var(--guess); border: 1px solid currentColor; border-radius: 99px; padding: 1px 6px; white-space: nowrap; vertical-align: 1px; }
    sup a { text-decoration: none; font-size: 0.72rem; padding: 0 1px; }
    .tag { display: inline-block; font: 500 0.72rem/1.4 var(--mono); padding: 0 7px; border-radius: 99px; border: 1px solid currentColor; white-space: nowrap; }
    .tag.yes { color: var(--good); } .tag.later { color: var(--hold); } .tag.no { color: var(--no); }
    .note { color: var(--muted); font-size: 0.9rem; }
    .chain { display: grid; grid-template-columns: 4.2em 1fr; gap: 2px 10px; font-size: 0.92rem; margin-top: 10px; max-width: 72ch; }
    .chain b { font: 500 0.82rem/1.55 var(--mono); color: var(--muted); }
    .src { font-size: 0.85rem; color: var(--muted); list-style: decimal; padding-left: 1.9em; gap: 3px; }
    .src li { max-width: none; overflow-wrap: anywhere; scroll-margin-top: 12px; }
    .src li:target { color: var(--fg); background: var(--panel2); }
    .start { border: 2px solid var(--ink); background: var(--bg); }
    .start h2 { font-size: 1.35rem; }
    .start h3 { margin-top: 16px; font-size: 1rem; }
    .bg-label { margin: 34px 0 0; font: 600 0.74rem/1.3 var(--body); letter-spacing: 0.05em; text-transform: uppercase; color: var(--muted); }
    </style>
    </head>
    <body>
    <div class="wrap">
      <div class="ramp" aria-hidden="true"><i style="background:var(--lit)"></i><i style="background:var(--shade)"></i><i style="background:var(--deep)"></i><i style="background:#3c2a55"></i></div>
      <h1>Lowmarch, Looking Back</h1>
      <p class="meta">From the lowmarch session, 2026-11-08. Reflection and research only: nothing was built, rendered or changed. Dates are your local dates.</p>
      <nav class="jump" aria-label="On this page"><a href="#start">First session</a><a href="#reflect">1. What happened</a><a href="#pros">2. How professionals do it</a><a href="#tools">3. Tools</a><a href="#forward">4. How to move forward</a><a href="#sources">Sources</a></nav>
    
      <section class="box start" id="start">
        <h2>Start here: the first session</h2>
        <p>You want to start over with one character, from scratch, so we can work out together how I should work. Nothing is built yet, and I'll wait for you to start it. Everything below this box is background: you don't need to read it or comment on it first.</p>
    
        <h3>The one thing I'd ask first</h3>
        <p><strong>Show me one picture you wish was a screenshot of our game.</strong> It can come from anything. It gives me the look, the camera, how big the character is on screen and what they're doing, without you having to put any of that into words. We'd study it only, never trace or ship it.</p>
    
        <h3>The session, about an hour</h3>
        <ol>
          <li>I say back what I see in your picture, in a few lines. You correct me.</li>
          <li>I set up that frame: same camera, same size, a plain grey figure standing in. We judge only there from now on, not on T-poses on black.</li>
          <li><strong>Row 1, big shapes:</strong> hair and outfit as rough flat-coloured blocks, 3 or 4 options. You pick one or mark it up.</li>
          <li><strong>Row 2, light and dark:</strong> skin tone and the pattern of shadow on your pick, 3 options.</li>
          <li>I write down what you picked and why, in your words. Then we spend five minutes on what was slow or annoying. That's what this session is really testing.</li>
        </ol>
        <p>Later sessions stay in the same frame: the face, then clothes, then one motion. That gets us one whole character, and a record of what each step cost.</p>
    
        <h3>How I'd show options quickly</h3>
        <ul>
          <li><strong>One link, updated in place,</strong> so you never have to hunt for a file.</li>
          <li><strong>Your picture on the left, options beside it,</strong> labelled A, B, C. Answer "B", or draw on it.</li>
          <li><strong>Each row changes one thing,</strong> with under 100 words beneath it.</li>
          <li><strong>Rough first.</strong> Blocks and colour changes should take minutes, and only your pick gets built properly. If a row will take more than 10 minutes, I'll say so before starting.</li>
          <li><strong>I time every round,</strong> from your answer to the next row. I haven't timed this way of working yet, so those first times are the number we try to beat.</li>
        </ul>
        <p class="note">I'm taking "from scratch" to mean a new character and a new way of working. The current bodies, VRoid base and renderer stay on the shelf, used only if they help. If you meant those too, tell me, because it changes the first hour.</p>
      </section>
    
      <p class="bg-label">Background: the retro</p>
    
      <section class="box">
        <h2>Short version</h2>
        <ol>
          <li><strong>The bodies worked. Nothing since has had that feeling.</strong> Your last words for a whole thing were <q>it looks perfect</q> (10-12). Since then praise has been for parts.</li>
          <li><strong>The real complaint all month is parts that don't belong together.</strong> You said it best: <q>there has to be cohesion in it</q> (10-31). I turned that into a lash rule. The word was bigger than the lash.</li>
          <li><strong>I never told you the big limit.</strong> I don't know if this route reaches what you compare it to, and I said things that sounded like I did. Studios get that look from a person's eye making many hand decisions, per angle.</li>
          <li><strong>We've never seen the current bodies dressed, moving, or in a game frame.</strong> They're 14 bare T-poses, judged on black tiles.</li>
          <li><strong>What I'd do:</strong> restart on one whole character in one game frame, and have you choose between rows early. Free. One week answers the ceiling question better than more fixes.</li>
        </ol>
      </section>
    
      <h2 id="reflect">1. What happened</h2>
      <div class="tally">
        <div><b>9</b><span>active days: 10-07 to 10-13, 10-30, 10-31</span></div>
        <div><b>13,979</b><span>lines of tool code</span></div>
        <div><b>40</b><span>lessons written, 1,440 lines</span></div>
        <div><b>32</b><span>open todos, 19 of them from the first week</span></div>
        <div><b>0</b><span>dressed characters on the current bodies</span></div>
        <div><b>0</b><span>game frames composed, 0 lines of game</span></div>
      </div>
    
      <h3>Where it stands</h3>
      <figure>
        <img src="[redacted: data-URI prefix][redacted: base64 data]…12028 tokens truncated…r than life.<sup><a href="#s11">11</a></sup> Revisions moved to pose shapes an animator dials in.<sup><a href="#s12">12</a></sup></li>
          <li>Skirts follow the legs first and sway second. Getting the skin weights right (how much each bone moves each mesh point) beats physics settings.<sup><a href="#s18">18</a><a href="#s22">22</a></sup> Colliders on legs alone fail, and a keep-inside cage works better.<sup><a href="#s19">19</a></sup> VRM now has inside colliders.<sup><a href="#s20">20</a></sup></li>
          <li>Nobody reaches zero clipping. Uma Musume adds special motions for big sleeves.<sup><a href="#s6">6</a></sup> Blue Protocol hides clashing parts.<sup><a href="#s7">7</a></sup> Dark Souls capes <q>intersect pretty much everywhere</q> and <q>nobody really notices</q>.<sup><a href="#s21">21</a></sup></li>
          <li>Real simulation is for big items like capes.<sup><a href="#s21">21</a></sup></li>
          <li>Fabric pattern shows in light and mostly vanishes in shadow.<sup><a href="#s24">24</a></sup></li>
          <li>Armour: only forum evidence found. Why simulated folds look wrong: no source says so outright.</li>
        </ul></div>
        <div><h4>For us</h4><ul>
          <li>Our folds are sine ripples that don't change with pose. A few named fold shapes tied to joint angles would suit code. <span class="g">guess</span></li>
          <li>Our stiff skirt cone is close to the opposite weighting. <span class="g">guess</span></li>
          <li>Grade clipping by what the game camera sees, not by millimetres. <span class="g">guess</span></li>
          <li>Keep our cloth sim for capes only. <span class="g">guess</span></li>
        </ul></div>
      </div>
    
      <h3>VRoid, studios, and where code falls down</h3>
      <div class="two">
        <div><h4>What's known</h4><ul>
          <li><strong>VRoid:</strong> hair guides don't survive export. Files hold merged, baked meshes. No scripting API found, and our driver is blocked by Windows until you allow it (#1105).</li>
          <li><strong>Better base:</strong> none found that ships in a closed game. MB-Lab and CharMorph require AGPL, which can block closed-source games; Ready Player Me ended in January; Character Creator is paid and realistic.<sup><a href="#s32">32</a></sup></li>
          <li><strong>Studios:</strong> modeller, texture artist, rigger, technical artist. A freelance listing prices an anime model at about 195,000 to 295,000 yen over 3 to 4 weeks.<sup><a href="#s31">31</a></sup> My estimate for a hero with rig: 3 to 8 person-weeks. <span class="g">guess</span></li>
          <li><strong>Agents judging looks:</strong> in BlenderGym the best model's error was about 10 times a person's on shape edits and 5 times on geometry. Judging which result was better, the best model agreed with people at 0.66; people agree with each other at 0.79. These are older models.<sup><a href="#s25">25</a></sup></li>
          <li><strong>Agents in Blender:</strong> good at scene work, poor at organic forms, proportion and weights, per mostly vendor reports.<sup><a href="#s29">29</a></sup></li>
          <li><strong>Code-built garments:</strong> about 64 to 72% simulate cleanly, with no fine straps and nothing on armour or anime.<sup><a href="#s27">27</a></sup> Models do far better on a small spec language: 94% vs 65% in one test.<sup><a href="#s26">26</a></sup></li>
          <li><strong>AI-to-3D anime:</strong> fused meshes or research-only weights, all trained on VRoid models.<sup><a href="#s28">28</a></sup></li>
        </ul></div>
        <div><h4>For us</h4><ul>
          <li>Keep VRoid. Keep our spec language small. <span class="g">guess</span></li>
          <li>Have code produce shadow masks, normals and line weights, not more geometry. <span class="g">guess</span></li>
          <li>You stay the eye, but earlier. One CG studio keeps fixing the block model until it reads as the character.<sup><a href="#s30">30</a></sup></li>
          <li><strong>A licence catch:</strong> our sample exports say <code>Redistribution_Prohibited</code>, <code>OnlyAuthor</code>, commercial <code>Disallow</code> inside the file. I checked two. It's probably the export dialog's default, since VRoid Hub lists samples A to C as open. Pixiv's A to Z help page has never been read; it blocks me.</li>
        </ul></div>
      </div>
    
      <h2 id="tools">3. Tools</h2>
      <p><strong>Cost: $0. Nothing here needs buying, and I'd buy nothing yet.</strong> You said on 10-30 you're open to paid extensions. No paid hair tool I found has a documented way to script it, so they only help if someone styles by hand. Versions read 11-08, except where noted.</p>
      <div class="tbl w"><table>
        <thead><tr><th style="width:20%">Free</th><th>What it does</th><th>Suits us?</th><th style="width:19%">Licence, how alive</th><th style="width:10%">Say</th></tr></thead>
        <tbody>
          <tr><td><a href="https://docs.blender.org/manual/en/latest/modeling/geometry_nodes/hair/index.html">Blender 5.2.3 hair nodes</a>, Curve to Tube</td><td>Braid, Curl, Frizz, Clump; tube meshes ready for textures</td><td>Yes. Scriptable, already installed, no install needed</td><td>GPL. LTS to Jul 2028</td><td><span class="tag yes">try</span></td></tr>
          <tr><td>VRoid Studio</td><td>Face and hair editor. Its sliders keep a face coherent, and it's the only source of mature faces</td><td>By hand, or through our driver once you allow it</td><td>Free; game use allowed per 10-09 notes. v2.14, Jun 2026</td><td><span class="tag yes">use</span></td></tr>
          <tr><td><a href="https://github.com/pixiv/three-vrm">three-vrm</a></td><td>The standard VRM renderer</td><td>Cross-check our models outside our own renderer</td><td>MIT. 3.5.5, Jul 2026</td><td><span class="tag yes">use</span></td></tr>
          <tr><td><a href="https://github.com/mikedh/trimesh">trimesh</a>, <a href="https://github.com/KhronosGroup/glTF-Validator">glTF-Validator</a></td><td>Ray and mesh checks; export file check</td><td>Yes, for clipping checks and an export gate</td><td>MIT; Apache-2.0</td><td><span class="tag yes">use</span></td></tr>
          <tr><td><a href="https://github.com/saturday06/VRM-Addon-for-Blender">VRM Add-on for Blender</a></td><td>VRM in and out of Blender, by script</td><td>Only if the Blender trial earns its place. Installing it needs your go</td><td>MIT or GPL-3. 4.10.4, 4 Nov 2026</td><td><span class="tag later">later</span></td></tr>
          <tr><td>Blender 5.2 Cloth and Hair Dynamics nodes</td><td>Node-based physics</td><td>One skirt trial, only if a rigged skirt fails. Marked experimental</td><td>GPL. In our copy</td><td><span class="tag later">later</span></td></tr>
          <tr><td><a href="https://github.com/vrm-c/vrm-specification/tree/master/specification/VRMC_springBone-1.0">VRM 1.0 capsule colliders</a>, spring limits</td><td>Fewer collision spheres, possibly less hair through bodies. We're on VRM 0.0</td><td>A half-day test. Low urgency</td><td>Open spec. Limits landed Oct 2026, not ratified</td><td><span class="tag later">later</span></td></tr>
          <tr><td>Our own shadow-map bake</td><td>A clean face shadow that moves with the light</td><td>No free Python baker found. QuickSDFTool is Unreal-only. Write it in numpy if a mask isn't enough</td><td>Ours</td><td><span class="tag later">later</span></td></tr>
          <tr><td><a href="https://github.com/maria-korosteleva/GarmentCode">GarmentCode</a></td><td>Sewing patterns from code</td><td>Ideas only. One garment averages 30k triangles, more than a whole Uma Musume character</td><td>MIT. 2.0.1, 2024</td><td><span class="tag later">ideas</span></td></tr>
        </tbody>
      </table></div>
      <p class="note"><strong>Rejected, free:</strong> official Blender MCP (running Blender by script is simpler); mcp-for-blender (four security flaws reported by a third party); UniRig (we have a rig); Abnormal (hand tool); CATS (on life support); Goo Engine (doesn't reach three.js); lilToon, Poiyomi, Unity Toon Shader (Unity only, read them for technique); MB-Lab and CharMorph (AGPL); StdGEN and CharacterGen (see above).</p>
      <div class="tbl w"><table>
        <thead><tr><th style="width:27%">Paid, all on hold</th><th style="width:22%">Cost</th><th>Why hold</th></tr></thead>
        <tbody>
          <tr><td><a href="https://superhivemarket.com/products/stylized-hair-pro">Stylized Hair PRO</a></td><td>$22, MIT (read 10-31)</td><td>The first buy if you want hair styled by hand. Anime locks, braids, curls. No scripting. A separate build per Blender version.</td></tr>
          <tr><td><a href="https://bartoszstyperek.gumroad.com/l/hairtool">Hair Tool 4</a>, <a href="https://superhivemarket.com/products/anime-hair-maker">Anime Hair Maker</a></td><td>$52; $25 to $27 (read 10-31)</td><td>Hand tools. Stylized Hair PRO comes first.</td></tr>
          <tr><td>Marvelous Designer</td><td>$39 a month or $280 a year</td><td>Hand tool for simulated clothes. I found no confirmed way to script it.</td></tr>
          <tr><td>Lightning Boy, Anime Shading Plus, Character Creator 5</td><td>$30+; $30; $299</td><td>Blender 3.4 only per its listing; Unity; realistic and hand-driven.</td></tr>
        </tbody>
      </table></div>
    
      <h2 id="forward">4. How to move forward</h2>
      <div class="box"><p style="margin:0"><strong>Restart on one whole character in one game frame. Everything else waits.</strong> Cost: $0, about 10 minutes of your time on most days.</p></div>
    
      <h3>What I'd change</h3>
      <ol>
        <li><strong>One whole character first:</strong> dressed, moving, in the frame the player sees. Range comes after. Saber is what won you over, and it was a whole person.</li>
        <li><strong>Big shapes before detail.</strong> You approve hair and outfit as flat shapes, plus the light and dark pattern, before any fine work.</li>
        <li><strong>You choose between rows.</strong> Two to four options beside your references, blind when it matters. You point, and I work out why.</li>
        <li><strong>Per-character work over global fixes:</strong> a tuned expression set and a designed shadow mask per character, and hair designed from a silhouette.</li>
        <li><strong>Study first, test the same day.</strong> One look at how it's done before the first build of anything new, and every piece of research ends in a test on our model.</li>
        <li><strong>I hold the map.</strong> One short page: the target, this week, what's parked. And each week one plain line: how far we are, and whether I still think this route gets there.</li>
        <li><strong>Two rounds on one complaint with no visible change, and I stop and say so.</strong></li>
      </ol>
    
      <h3>What I'd ask you, one at a time</h3>
      <div class="tbl w"><table>
        <thead><tr><th style="width:52%">Question</th><th>What I'd assume if you have no strong view</th></tr></thead>
        <tbody>
          <tr><td>What do you wish ours looked like now? Five images.</td><td>Yin and Yang for faces, Unicorn Overlord for costume. Your reference moved, and I never asked.</td></tr>
          <tr><td>What does the player actually see? Camera and size, in battle and in dialogue.</td><td>A side-view battle frame, plus one dialogue close-up.</td></tr>
          <tr><td>Who's the first character, and what are they doing?</td><td>An original swordswoman on the slender body, in the Saber armour kit, doing one attack. Saber was the high point, and the kit exists.</td></tr>
          <tr><td>What's this for: a game you'll ship, or finding how far a model can go?</td><td>Both, with the second driving. It changes what "done" means.</td></tr>
          <tr><td>Is model-only part of the point, or would a person's eye on key decisions be welcome, if free?</td><td>Model-only for week one. No money implied.</td></tr>
          <tr><td>Could you open pixiv's AvatarSample A to Z help page by hand? It blocks me, and we need it before anything ships.</td><td>Not urgent, but nobody has read it.</td></tr>
        </tbody>
      </table></div>
      <p class="note">Small ones still waiting: the old man's fangs (I'd remove them), and allowing the VRoid driver through Windows (#1105), which any mature face needs.</p>
    
      <h3>A first week</h3>
      <div class="tbl w"><table>
        <thead><tr><th style="width:9%">Day</th><th>What I do</th><th style="width:30%">What you see</th></tr></thead>
        <tbody>
          <tr><td>1</td><td>The questions above, one at a time, played back to you. No code.</td><td>One page: references, target frame, the character. About 30 minutes of your time.</td></tr>
          <tr><td>2</td><td>Put the current body in the frame on plain ground. Test the light and dark pattern: skin tone, a shadow down the side of the face, darker and simpler hair.</td><td>Three rows beside your references. Pick one.</td></tr>
          <tr><td>3</td><td>Hair and outfit as big flat shapes, three or four options. The dark layer under the hair goes in here, which closes the slice.</td><td>Options in the frame. Pick or mark up.</td></tr>
          <tr><td>4</td><td>Her face only: the 3/4 moves, a shadow mask, six expressions tuned for her.</td><td>Blind rows at 3/4 and in the frame.</td></tr>
          <tr><td>5</td><td>Dress her. Folds placed for the pose, skirt following the legs, hair and cloth motion. One idle, one attack. Clipping graded from the game camera.</td><td>A short clip and a turning view.</td></tr>
          <tr><td>6-7</td><td>Your verdict, then my plain note: how close, what's between, what more work buys. Then we decide: next character, more depth, or change route.</td><td>One page, one decision.</td></tr>
        </tbody>
      </table></div>
      <p class="note">How we'd know it worked: you'd show her to someone, or say something like 10-08 again. If you don't, the route gets questioned out loud instead of patched.</p>
    
      <h3>What I'd stop doing</h3>
      <ul>
        <li>Fixing 14 bodies before one looks right.</li>
        <li>Judging on T-poses on black.</li>
        <li>Writing new checkers for looks. Checks stay for things with a right answer: proportions, holes, clearance, regressions.</li>
        <li>Patching VRoid's painted eyes and mouths face by face without asking whether that's the right fight.</li>
        <li>Reports that end with nothing changed on screen.</li>
        <li>Ledger numbers and long replies during review. A sheet and under 100 words.</li>
        <li>Saying something closes the gap. I'll show it, and you say.</li>
        <li>Carrying 32 open todos. Close or merge the ten or so about retired characters and old dressed work, and keep the rest on the one page.</li>
      </ul>
    
      <h2 id="sources">Sources</h2>
      <p class="note">All read 2026-11-08 through a summarising fetch. I checked the main quotes from 1, 5 and 9, and the figures from 25, directly. Local evidence: <code>docs/portrait-guide.md</code>, <code>docs/vroid.md</code>, the handoff, session transcripts, and the VRM files themselves.</p>
      <ol class="src">
        <li id="s1" value="1">Arc System Works, modelling tips: <a href="https://www.docswell.com/s/ASW_Academy/5D2VY6-GG-Modeling">docswell.com</a></li>
        <li id="s2" value="2">Arc System Works, toon lines: <a href="https://www.docswell.com/s/ASW_Academy/5LVY67-GG-Toonline-Eng">docswell.com</a></li>
        <li id="s3" value="3">4Gamer, Guilty Gear Xrd part 1 (pages 1-2): <a href="https://www.4gamer.net/games/216/G021678/20140703095/">4gamer.net</a></li>
        <li id="s4" value="4">4Gamer, Guilty Gear Xrd part 2 (page 1): <a href="https://www.4gamer.net/games/216/G021678/20140714079/">4gamer.net</a></li>
        <li id="s5" value="5">Game Watch, Uma Musume models: <a href="https://game.watch.impress.co.jp/docs/kikaku/1366154.html">game.watch.impress.co.jp</a></li>
        <li id="s6" value="6">Game Watch, Uma Musume conversation scenes: <a href="https://game.watch.impress.co.jp/docs/kikaku/1366170.html">game.watch.impress.co.jp</a></li>
        <li id="s7" value="7">Game Watch, Blue Protocol, CEDEC 2020: <a href="https://game.watch.impress.co.jp/docs/news/1275034.html">game.watch.impress.co.jp</a></li>
        <li id="s9" value="9">Game Watch, Gakumas, CEDEC 2024: <a href="https://game.watch.impress.co.jp/docs/kikaku/1617554.html">game.watch.impress.co.jp</a></li>
        <li id="s10" value="10">Anime Shading Plus, face shadow map workflow: <a href="https://erichu33.github.io/ASPDocs/en/articles/face-shadow-map-creation-and-baking-workflow.html">erichu33.github.io</a></li>
        <li id="s11" value="11">Autodesk cel-look modelling tutorial: <a href="https://area.autodesk.jp/column/tutorial/celllook-character-modeling/01-basics-preparation/">part 1</a>, <a href="https://area.autodesk.jp/column/tutorial/celllook-character-modeling/02-edit-rough-model/">2</a>, <a href="https://area.autodesk.jp/column/tutorial/celllook-character-modeling/03-detailed-modeling/">3</a></li>
        <li id="s12" value="12">CGWORLD, Revisions: <a href="https://cgworld.jp/feature/201812-cgw245-rvsns.html">cgworld.jp</a></li>
        <li id="s13" value="13">CGWORLD, Arc System Works feature (partial): <a href="https://cgworld.jp/special-feature/202508-asw.html">cgworld.jp</a></li>
        <li id="s14" value="14">CHARM, a study of VRoid hair: <a href="https://arxiv.org/html/2509.21114v1">arxiv.org</a></li>
        <li id="s16" value="16">Inside Games, Gakumas budgets: <a href="https://www.inside-games.jp/article/2024/08/31/158987.html">inside-games.jp</a></li>
        <li id="s17" value="17">80.lv, long curly anime hair: <a href="https://80.lv/articles/modeling-long-curly-hair-for-a-3d-anime-style-character">80.lv</a></li>
        <li id="s18" value="18">Zenn, a hobbyist's memo on hair and skirts: <a href="https://zenn.dev/link/comments/524655cd70e5a3">zenn.dev</a></li>
        <li id="s19" value="19">Avatar makers on skirts: <a href="https://cgbox.jp/2021/11/09/howtomakeskirt/">cgbox.jp</a>, <a href="https://signyamo.blog/phys-bone_skirt/">signyamo.blog</a>, <a href="https://note.com/airgreen/n/na9eeaf74e6ed">note.com</a></li>
        <li id="s20" value="20">VRM extended colliders: <a href="https://github.com/vrm-c/vrm-specification/blob/master/specification/VRMC_springBone_extended_collider-1.0/README.md">github.com</a></li>
        <li id="s21" value="21">80.lv, cloth simulation for games (2020): <a href="https://80.lv/articles/cloth-simulation-for-games-difficulties-and-current-solutions">80.lv</a></li>
        <li id="s22" value="22">MagicaCloth, preventing penetration: <a href="https://magicasoft.jp/en/prevent-penetration-2/">magicasoft.jp</a></li>
        <li id="s24" value="24">CGWORLD, Soten no Ken: <a href="https://cgworld.jp/feature/201812-cgw245-soten.html">cgworld.jp</a></li>
        <li id="s25" value="25">BlenderGym (GPT-4o and Claude 3.5 era; photometric loss): <a href="https://arxiv.org/html/2504.01786v1">arxiv.org</a></li>
        <li id="s26" value="26">Design2GarmentCode: <a href="https://arxiv.org/html/2412.08603v2">arxiv.org</a></li>
        <li id="s27" value="27">GarmentCodeData: <a href="https://arxiv.org/html/2405.17609v3">arxiv.org</a></li>
        <li id="s28" value="28">AI-to-3D: <a href="https://github.com/hyz317/StdGEN">StdGEN</a>, <a href="https://github.com/zjp-shadow/CharacterGen">CharacterGen</a>, <a href="https://github.com/ShuhongChen/panic3d-anime-reconstruction">PAniC-3D</a></li>
        <li id="s29" value="29">Agents in Blender, practitioner and vendor reports: <a href="https://www.mindstudio.ai/blog/claude-blender-mcp-real-world-performance">mindstudio.ai</a>, <a href="https://www.3daistudio.com/3d-generator-ai-comparison-alternatives-guide/blender-mcp">3daistudio.com</a></li>
        <li id="s30" value="30">CGWORLD, Snack World (search snippet, likely page): <a href="https://cgworld.jp/regular/201805-c-olm.html">cgworld.jp</a></li>
        <li id="s31" value="31">Prices: <a href="https://www.lancers.jp/menu/detail/1238820">Lancers listing</a> (search snippet only, two price tiers)</li>
        <li id="s32" value="32">Bases: <a href="https://raw.githubusercontent.com/animate1978/MB-Lab/master/license.txt">MB-Lab licence</a>, <a href="https://raw.githubusercontent.com/Upliner/CharMorph/master/license.txt">CharMorph licence</a>, <a href="https://genies.com/blog/ready-player-me-shutdown">Ready Player Me shutdown</a>, <a href="https://www.reallusion.com/license/content.html">Reallusion</a></li>
      </ol>
    </div>
    </body>
    </html>