SLAPSTACK STUDIO

spawn layers with AI · scramble · bind · grab · stack · every entity is a posterior ATOM SOUP

drag = pose clamp · wheel rotate · shift+wheel scale · shift+click = hint · erase tool: brush deletes atoms

SCENE

noise

composite field painter (depth) depth occlusion

binding vs truth— mean entropy— BP iterations— atoms in scene—

EDIT & LOOK

burn into template

backdrop

coarse1.00

mid1.00

fine1.00

opacity1.00

Erase deletes atoms under the brush (render only, until burn writes it into the layer template; Scramble/Bind rebuild the scene from templates). Undo works until the next Bind. Backdrop is presentation and survives everything. Band gains (coarse<4, mid 4–10, fine>10 c/img) scale atom colors in field mode; opacity blends the whole field over the backdrop.

LAYERS (front is highest)

library — place to add to the scene:

SPAWN A LAYER

Distill runs Stable Diffusion score distillation with a cross-attention matte (GPU Space only, untested — first run is a smoke test). Fit trains atoms to your image on CPU (~1 min; objects on plain mid-gray come out compact).

PROMPT AS INTERVENTION

Commands become pose clamps and depth moves; BP re-settles. Free-form uses Claude when this runs inside claude.ai; a local grammar otherwise.

OVERLAYS

atoms, colored by belief ownership tint entropy border glow pose belief (crosshair + covariance) semantic hints

HONEST LEDGER

Verified in this build: the BP core matches the SlapstackBet6 Python (transform 2e-16, render MSE 8.6e-8, identical binding); drag = pose clamp + re-settle; identity persists through occlusion; depth occlusion removes evidence honestly (marginals revert toward the prior, coverage releases when moved away); the semantic likelihood factor resolves the twins case (two identical stars: 0.05 acc on geometry alone, 0.99 with location priors) and a mildly wrong prior is overridden by evidence; image→layer fitting runs end to end on CPU. Bet 9: the eraser filters per-atom state (order-free, so undo is re-append), burn edits the template so erases survive rebuilds; the backdrop is pure presentation and persists across Bind/Scramble; band gains exploit render linearity (gain·color ≡ gain·contribution), verified in tests_edit.js.

Not verified: text→layer (SDS) and cross-attention matte are line-for-line adaptations of the verified Bet-5 GPU loop but have NOT been executed on a GPU — first run is a smoke test; SD attention maps flicker across seeds, which is why they enter BP as a prior with a weight, never as truth. Real-photo scene→atoms remains open. Painter compositing and the alpha matte are presentation choices, not the additive field model.