ARTI 1.7.0 · PyTorch · Tensor in / Tensor out
ARTI lets models select, combine, compact, and expand tensors while keeping position, provenance, visibility, and call-time output spaces explicit. Attach it to an existing tokenizer, or train tokenizer-free models that read literal glyph tensors.
pip install arti-fitimport torch
import arti.nn as ann
x = torch.randn(4, 128, 64) # [B, N, D]
mask = torch.ones(4, 128, dtype=torch.bool)
q = torch.rand(4, 128)
pulse = ann.Pulse(k=8, dim=64, hidden_dim=128)
workspace, info = pulse(x, q=q, mask=mask, return_info=True)
assert workspace.shape == (4, 8, 64)LIVE MODELS
A trained VisualScan + Pulse model. Remove complementary views, repeat a phase, or corrupt coordinates and compare the reconstruction with the fixed-capacity workspace.
Run the Pulse demoType visible diacritics, shuffle or remove entries from this call's output vocabulary, and watch the model return local slots that compose back into text.
Run dynamic vocabularyMeasure empty state, commit the reference trace, and run near and novel probes through the real read/update graphs. Fork and reset show that memory belongs to the session.
Run the memory experimentRun reference A and changed B in one click. Compare fixed K=3 output, Half, and routing before opening the real workspace; mechanism response stays separate from task performance.
Run the evidence comparisonPUBLIC API
HalfContinuously attenuate weak features without changing shape.Core→FoldCompact a variable N slots into a fixed K slots.Core→PulseLearned projection → Half → Fold into a fixed-capacity workspace.Core→RecallRead and update explicit, fixed-capacity inference state.Core→UnFoldQuery new values and learn a layout while preserving original instances.Core→FusionPulseFuse several already-formed Pulse workspaces.Alpha→Runtime vocabularySupply a new local output space as a tensor for each call.Alpha→MembraneSeparate token readability, model context, and public emission.Alpha→Phase framesProvide observer coordinates and paired rotation operators explicitly.Mixed→INTEGRATION
Keep the tokenizer and backbone. Apply Pulse, Recall, coordinates, and visibility to the model's hidden tensors.
tokens → hidden tensor → ARTI → model headPass glyphs, layout, and the call-specific output alphabet as tensors. Input and output vocabularies may differ, but the model still needs task-specific training.
glyph tensor + layout + runtime vocabulary → local slotsRELEASE STATUS
Half, Fold, Pulse, UnFold, and the supported 1.x core are in compatibility verification. FusionPulse, Web inspection, and literal/runtime-vocabulary model surfaces remain Alpha.
Read the documentation →START BUILDING