2026.08.17 — THE ENCODER GATE
One pass, ~5M params: 23.4 dB / 0.944 SSIM on held-out clips at native resolution, macro-layout at the fitted ceiling. First arm to beat the blur baseline on every metric — by 8.7 dB over the best set-generative arm. The field, not the pixels, remains the persistent state.
2026.07.31 — A BURIAL
The Voronoi arm lost every measured axis even after a steelman round, and was removed. Mark-space generative emission is retired as the primary path: every sampler composed worse macro-structure than trilinear upsampling of its own conditioning. Full post-mortem in the decision log.
SCALING — v0 → v1 → v2
CFD 0.31 → 0.20 → 0.16, monotone on both metrics with a flattening slope: model scaling saturates against a 231-window single-scene corpus. Data is the measured next axis; a 2,392-scene sky-timelapse corpus is fitting now.