ARTI / LIVE MODELS
Run ARTI's core mechanisms directly.
Each lab starts with an action and a visible consequence, then runs a trained model, state graph, or mechanism graph. Change the input, compare the result, and open real intermediate tensors only when you need them. Failed inference never falls back to precomputed output.
TRAINED MODEL · PULSE
Recover one glyph from four incomplete views.
Choose a character and change the observations. The model compacts a variable set of visual fragments into 16 Pulse slots, then reconstructs a 16×16 glyph.
Each row is a real slot; color encodes relative feature magnitude.
—
- Target character
- —
- Model output
- —
- Execution provider
- —
- Run time
- —
Locked held-out set · n=2048
100%held-out 8-class character accuracyThis controlled visible-glyph task is limited to O, 0, I, 1, S, 5, B, and 8. It is not general OCR, handwriting recognition, Unicode-source recovery, or a way to infer zero-width characters from screenshots.
Model and runtime details
- Input
- frames [1,4,1,8,8] · shifts [1,4,2] · frame_mask [1,4]
- Real outputs
- workspace [1,16,21] · reconstruction [1,1,16,16] · scores [1,8]
- Artifact
- arti.VisualScan · opset 18 · SHA-256 verified · ARTI commit 7e61815
- WebGPU gate
- Probing browser adapter…
- Claim boundary
- controlled eight-class visible-glyph recovery