Model self-postmortem: four architecture choices that made my Three.js Minecraft lose to Fable 5.2

teortaxesTex · x · 2026-11-10

The author had the model post-mortem why Fable 5.2's Three.js folding-mansion demo crushes his own build. Four wrong forks: (1) folding 24 independent tiles with a per-depth hinge schedule instead of Fable's 6-face hinge tree with one theta (fold: 3.32s, max per-frame tile move 7.6m); (2) keeping object identity across the reality switch where Fable preserved appearance identity — same shader for land and walls, mansion renders in 326 triangles vs his 42k–147k; (3) imposing a globally consistent 'door forest' instead of local gluing + stencil clipping, which accounts for most of his 60-fold.js bug surface; (4) hand-rolled random-rolling-walk world packing producing 52% Unbound vs Fable's DFS branch-and-bound at 37.8%. He lists ordered V4 fixes and notes his failures were accidental architecture mistakes that looked like visual bugs, while Fable's limits are chosen trade-offs. Screenshots show the session ran on DeepSeek V4.1 Flash Max (5 turns, 403 steps, 99% cache hit).

Original post →

Model self-postmortem: four architecture choices that made my Three.js Minecraft lose to Fable 5.2

More from coding & agent