Game guide · source of truth
Ledger

MODEL-01E

Jerboa reference-view cut fix

← Ledger

donestatus · MODEL
10✓ tests passed / failed
2.39Mtokens · 8 min
2evidence files · 377 kB

Evidence

From logs/evidence/MODEL-01E/. Click a thumbnail for the full image.

Images (1)

views-jerboa-neighbours.png
views-jerboa-neighbours.png · 376 kB

Logs and data (1)

jerboa-cut-before-after.txt · 1 kB · open
MODEL-01E - Bellwhisker Jerboa reference-view cut (ART-MON-053 cell col 2 row 2, 614x384 px)
Per view: share of the drawing / largest connected part share of the crop / aspect (w/h)

Before (MODEL-01C rule): valley cuts at x=461 and x=106 (106 = the thin tail just before its tuft;
the tail tapers into a deeper valley than the real side/front junction at ~300)
  side  0.06 / 1.00 / 1.55   <- tail tuft only
  front 0.66 / 0.53 / 1.47   <- side body + front view
  top   0.28 / 0.98 / 0.53
  boxes: side (33-105), front (114-444), top (426-587)

After (STUB rule: a valley cut that leaves on one side < 10 % of the drawing which is >= 90 % one part lying mostly
across the cut is skipped): cuts at x=461 and x~300
  side  0.41 / 1.00 / 1.25
  front 0.31 / 0.98 / 0.64
  top   0.28 / 0.98 / 0.53
  boxes: side (25-300), front (300-444), top (426-587)

Valley-cut survey of all 20 sheet cells (smaller side's share of the drawing, before the fix):
  real view boundaries 0.204-0.453 (beetle, fox, wolf, gecko x2, viper, scarab, tortoise, hyena, pangolin, jerboa 461)
  jerboa tail cut 0.057 (only one below STUB 0.1). Viper and hyena valleys also cut one shadow-joined part (64 % / 70 %
  across) but hold 23 % / 22 % of the drawing, so they stay.

--check over all 40 requests after the fix: min share 0.19, min largest-part share 0.98, aspect 0.23-1.60, 0 flagged.
--check on the jerboa with STUB = 0 (old behaviour): flagged "side: 6% of the drawing", "front: largest part 53%".

Tests

pass re-cut all 40 reference views + jerboa contact sheet · 8.9 s · raw log$ python3 tools/model-views.py --quiet --sheet logs/evidence/MODEL-01E/views-jerboa-neighbours.png --ids MDL-MON-bellwhisker_jerboa,MDL-ELT-bellwhisker_jerboa_elite,MDL-MON-brineback_tortoise,MDL-MON-duskfang_hyena,MDL-MON-siltspit_gecko
pass view-cut sanity check (share, one subject, aspect) for all 40 · 4.1 s · raw log$ python3 tools/model-views.py --check --quiet
pass data tests · 11.5 s · raw log$ pnpm data:test
pass data tests · 93.2 s · raw log$ pnpm data:test
pass zoen-sim cargo tests + clippy · 29.6 s · raw log$ pnpm sim:test
pass Zoen.Core dotnet build + xunit · 4.8 s · raw log$ pnpm core:test
pass docs lint + index · 1.1 s · raw log$ pnpm docs:lint
pass wiki typecheck · 8.0 s · raw log$ pnpm --filter @zoen/wiki typecheck
pass wiki production build · 13.9 s · raw log$ pnpm wiki:build
pass wiki e2e (changed specs) · 1.3 s · raw log$ pnpm --filter @zoen/wiki exec playwright test --only-changed

Notes

2026-10-10 08:53 UTC · Fix in tools/model-views.py: new STUB rule (0.1) - a valley cut is skipped when one side holds < 10% of the drawing and >= OWN 90% of it is one part lying mostly (> half) across the cut (a tail's tuft). Jerboa tail cut had 5.7% (86% of the part across); every real valley boundary of the 20 sheet cells is 20-45%, so only the jerboa changed: side (25-300) now holds body+tail, front (300-444) only the front view; the other 38 cuts are byte-identical (git status: only jerboa + elite side/front previews and prompt.txt changed). New --check mode (writes nothing): re-splits all 40, flags a view < CHECK_SHARE 10% of the drawing, a crop whose largest connected part is < CHECK_ONE 80% (two views), aspect outside 0.15-3.0, or a missing view file; measured after fix min share 0.19, min largest-part 0.98, aspect 0.23-1.60; with STUB=0 it flags the jerboa (side 6%, front 53%). Refactor: cell(r) shared by cut/check. Evidence: logs/evidence/MODEL-01E/views-jerboa-neighbours.png (jerboa, elite, tortoise, hyena, gecko), jerboa-cut-before-after.txt. verify --changed all green. Model inbox dry-run: 0 delivered.