`docs/cactus-scale-sample-2026-06-09.md`

26 lines
# Cactus seamless redo — scale-sample validation (2026-06-09)

28 samples generated (3 rounds × ~10 colorways) via Mac2 ComfyUI circular-padding
(GEN_BACKEND=comfy WALLCO_FORCE_TILEABLE=1), all is_published=FALSE. Steve capped
the run at ~25-30; stopped at 28.

## Results (cactus-sample-eval.py: 6-lens edges scan + llava vision)
| window | seam-PASS (ΔE≤5) | vision ≥70 | KEEP |
|---|---|---|---|
| Round 1 (10 tiles) | 9/10 (90%) | 10/10 | 9/10 |
| Round 2 window (14) | 11/14 (78%) | 14/14 | 11/14 |
| Round 3 window (13) | 10/13 (76%) | 13/13 | 10/13 |

vs OLD baseline (replicate, no circular padding): ~19% seam-PASS.

## Read
Circular padding holds at scale: per-window seam-PASS 76-90% (4-4.7x the old
baseline), vision-concept 100% across all 28. The rate straddles the 80%
"ready to scale" bar from the task brief — round 1 cleared it, sliding windows
land just under. Failures are seam-only (mid ΔE 5-7), all heal-candidates for
the sanctioned thin-band feather (seam-heal-feather.py heals only FAILING tiles).

## Status: bulk regen of the 2,262 purged cactus REMAINS GATED (Steve's explicit
go required — heavy overnight compute). Recommendation if scaled: generate +
auto-heal fails + re-gate, expect ~90%+ net KEEP.