WebGrid chase = paint / actions lab · not ARC % · peak BPS is lab score. Full presentation: webgrid-ugrad.html
Capture + suicide block · go-ugrad pattern · 9-cell slice for tensor lane
Live board = human / engine. Under-layers spawn multipv trees (depth 1–3) as moves happen — chess.com-style analysis under live play. Lab: no castling / en passant · promo→Q.
DINOv3 (Meta SSL) — dense visual features without labels · scale data/model ·
Gram anchoring keeps patch maps stable on long schedules · post-hoc resolution/text align.
paper ·
github.com/facebookresearch/dinov3.
Live lab: dense pixel estimation for glyph peel (patch energy + Sobel + chroma · Gram proxy).
Higher-res: up to 64² patches · 960px working plane · hybrid color scoring.
Not full DINOv3 weights in WK — hook window.__dinoInfer(ImageData) when ONNX is loaded.
Heatmap energy↑ · chroma shifts hue · anaglyph = |R−B| · hsv = hue/sat score
Heat = per-patch dense score (luma/RGB/chroma/hybrid/anaglyph/hsv). CLI: fcs glyph color chroma · ?color=chroma&heat=turbo.
Glyph path: frame → dense map → threshold tickets → fountain QR (soak).
BC ugrad-glyph-dense · color_update on change.
Rubik language solve prep · full nets live on
cube-viewer (L0–L7 AIR→DISCOURSE · 36 langs · 15 boards).
Arena posts pathway ticks to ugrad-rubik-lang for dual-seat solve.
BC ugrad-tensor-lane · ugrad-webgrid-race · ugrad-lite-arena
Shared run log (WebGrid chase + Arena) · ARC task shape: Model · Score · Actions · Replay · Published. Lab score = peak BPS (WebGrid) or eval (chess layers). ARC ref row is comparison only.
| Model | Score | Actions | Surface | Replay | Published |
|---|