Previous demo retired
The public page no longer loads glyph detection, crops, sequence-alignment decoding, or the previous OCR weights, and no longer publishes their metrics.
ARTI / PULSE-NATIVE SPATIAL TEXT
The previous version relied on glyph detection, cropping, and discrete ordering, so it could not demonstrate ARTI’s core mechanism. It has been removed from public execution. The replacement receives only page pixels, continuous coordinates, masks, and the runtime output vocabulary.
The public page no longer loads glyph detection, crops, sequence-alignment decoding, or the previous OCR weights, and no longer publishes their metrics.
One official Pulse reads both whole-page and arbitrarily partitioned multiscale fragments, then compresses variable evidence into a K=32 workspace.
Interaction returns only after the larger model passes overfit, held-out, Pulse ablation, coordinate corruption, runtime-vocabulary permutation, and browser parity gates.
04 / MODEL CONTRACT
PAGE PIXELS→MULTISCALE FRAGMENTS
+ COORD + MASK→OFFICIAL ARTI PULSE
HALF + FOLD→K=32 WORKSPACE→RUNTIME VOCAB SLOTS
+ SPATIAL STRUCTURE05 / CAPACITY
If removing Pulse leaves a conventional OCR pipeline able to finish the task, the demo is not valid evidence for ARTI. An explicit pause is more accurate than substituting legacy results.