The gacha critical path, as a ladder

Gall’s Law rungs, hours calibrated against RFD 1130 and RFD 2133–2135, 2026-08-31

Each rung is a small working system that the next extends: the whole pipeline runs at every point, with fewer steps at the bottom and more at the top. A pull is a gacha item only after rung 6; the pool at rung 7 and the public page at rung 9 are the last two rungs, not parallel tracks.

The ladder

9 public hosted page + sponsor link + automation disclosure · control: link resolves, disclosure present 8 pool → roll button page dispenses one VRM per pull · control: two rolls, two distinct VRMs 7 prompt list → pool (~50 VRMs) seed-to-VRM in a batch, brokered 4090s when the desk isn't enough · control: rerun by seed matches 6 tagged rig → VRM one command, all stages · control: opens in Godot and a third-party VRM viewer 5 skinned mesh → tagged (VoxHammer + EditScore recover partition) partition-as-corruption: VoxHammer proposes toward canonical, EditScore judges, N-bounded · every vertex tagged 4 passing mesh → skinned (SkinTokens) SkinTokens skin-mode against canonical ANNY skeleton · control: EditScore on posed frames passes, bent-rig control fails 3 judged → repaired (VoxHammer) VoxHammer repair pass on fails, rejudged, N-bounded · control: a planted broken mesh reaches passing or refuses 2 mesh → judged (EditScore) score against the prompt · control: planted-good scores above threshold, planted-bad below 1 image → mesh (Pixal3D), from OmniGen2 text→image OmniGen2 text-to-image feeds Pixal3D on the desk 3090 · control: image dimensions match target; mesh geometry non-empty START PUBLIC

Every rung shows something: rung 1 is a mesh, rung 4 is a moving figure, rung 6 is a VRM you can hand a friend, rung 8 is a page a stranger clicks. A rung is done when its output feeds the next one and a control run on a known-broken input fails.

The rungs, in order

# Rung Output Negative control
0 text → image (OmniGen2) one reference image per prompt image dimensions match target
1 image → mesh (Pixal3D) one mesh per image on the desk 3090 image in, mesh out, geometry non-empty
2 mesh → judged (EditScore) a number per (prompt, mesh) planted-good scores above threshold, planted-bad below
3 judged → repaired (VoxHammer) passing mesh, or a recorded refusal a planted broken mesh reaches passing or refuses
4 passing mesh → skinned (SkinTokens) rig against canonical ANNY EditScore on posed frames passes, bent-rig control fails
5 skinned mesh → tagged (VoxHammer + EditScore recovery) 23-tag per-vertex primvar every vertex tagged (RFD 1121); planted mislabeled partition converges to canonical
6 tagged rig → VRM one command, portable file opens in Godot and a third-party VRM viewer
7 prompt list → pool ~50 VRMs, seed-reproducible rerun by seed matches; brokered rents ride RFDs 2133–2135
8 pool → roll button a page dispensing one VRM per pull two rolls, two distinct VRMs
9 public hosted page, sponsor link, disclosure link resolves, disclosure present

Ground already laid

  • image-to-mesh through the baked USD-to-Godot path
  • canonical ANNY body (317 shapes, phenotype clips, character rig)
  • MToon materials, OmniGen2 variation + judge, TokenRig forward pass
  • spend infrastructure the higher rungs lean on: spot-broker (RFD 2133), mutual-TLS FDB with continuous Tigris backup (RFD 2134), machine checks that watch both (RFD 2135)

The same ladder as a PERT network

The ladder read as a PERT chain, so the schedule shape stays legible next to the demo shape. Estimates are in focused hours, not working days: RFD 1130 measured pipeline stages in MM:SS, RFD 1016 in projected hours, and today’s own arc calibrated the smaller ones against real work — RFD 2133–2135 shipped end to end in roughly 8–10 focused hours across a day. A day of continuous 8-hour focus is not a real unit here; the honest atom is the session. te = (O + 4M + P)/6.

Rung O (h) M (h) P (h) te (h) σ (h)
0 text→image (OmniGen2) 1 2 4 2.2 0.50
1 image→mesh (Pixal3D) 1 2 4 2.2 0.50
2 EditScore judge 2 4 8 4.3 1.00
3 repair (VoxHammer) 3 6 12 6.5 1.50
4 skin-mode (SkinTokens→ANNY) research risk 6 16 48 19.7 7.00
5 partition recovery (VoxHammer + EditScore) 3 6 12 6.5 1.50
6 seed→VRM 2 3 5 3.2 0.50
7 pool ×50 2 4 8 4.3 1.00
8 roll button 3 6 12 6.5 1.50
9 ship 1 1.5 3 1.7 0.33

Sum of te: 57.1 hours, σ 7.7 (finish window 49–65 hours). At 5 focused hours a day (RFD 2135’s duty-hours note) that is roughly 10–13 working days wall-clock. Rung 4 dominates the variance — it is the one research risk on the chain and every other rung is integration. The chain is the critical path because it IS the chain; ladder rungs have no non-critical siblings by construction, which is the Gall’s Law property this shape buys.

0 h 8 16 24 32 40 48 56 ~ 1 focused day = 5 h (RFD 2135) 0 t2i 1 mesh 2 judge 3 repair 4 skin-mode rig — research riskt_e 19.7 h, σ 7.0 5 tags 6 VRM 7 pool 8 roll 9 ship

What is not on this ladder

The 10 absent attachment tags (headwear, tail, wings…) are wave 2 by RFD 1121’s own route: nothing exists to tag until a later ladder generates them, so they are off-chart by design.