ARTI / CONCEPT

Compress evidence without flattening its structure.

ARTI is a PyTorch-first research architecture. It lets variable observations enter fixed-capacity latent workspaces while coordinates, provenance, visibility, and call-time output spaces remain explicit tensor contracts.

EDITORIAL FRAME · NOT A PROVEN OBJECT THEORY

A research lens for understanding ARTI

Observation, Perspective, and Boundary are not a proven object theory. They organize the public tensor contracts and the questions behind them. Task behavior still depends on training, data, and caller-supplied structure.

01OBSERVATION

A representation can be an observation.

Hidden tensors, glyphs, patches, and call-time vocabularies can be evidence for this computation without sharing one permanent token identity.

Literal · Runtime vocabulary · Pulse
02PERSPECTIVE

Relations can be observer-relative.

Coordinates, frame operators, and observer coordinates are supplied explicitly. ARTI does not discover sources or physical reference frames.

coord · observer frame · PhaseMixer
03BOUNDARY

Existence, readability, and emission differ.

Visibility, model context, and public emission can be modeled separately. Membrane is not encryption or an independent security boundary.

visibility · context · emit

01 / SHARED STRUCTURAL PROBLEM

One computational path across the mechanisms

ARTI does not require every model to use the whole path. Each module remains tensor-in/tensor-out; this map shows the structural problem they share.
01VARIABLE EVIDENCE

Keep call-time structure at input

hidden states · glyphs · coordinates · masks
02COMPACT WORKSPACE

Pulse / Half / Fold

[B,N,D] → [B,K,D]
03STATE & FUSION

Recall / FusionPulse

explicit state · compact multi-source workspaces
04RUNTIME VIEW

UnFold / Literal output

dynamic layout · call-time local slots

02 / ADOPTION

Two real adoption paths

PATH ASUPPORTED 1.X SURFACE

Keep the existing tokenizer

Attach ARTI to existing PyTorch hidden tensors to add structure, visibility, dynamic state, or fixed-capacity workspaces.

tokenizer → hidden tensor → ARTI
PATH BALPHA MODEL SURFACE

Use literal tensors directly

Let visible glyphs, layout, coordinates, and the current output vocabulary enter the model. These Alpha interfaces still require task-specific training.

literal input → layout → runtime vocabulary

03 / RESEARCH DISCIPLINE

Contracts, evidence, and hypotheses stay separate

Fixed shapes, value-preserving gathers, and visibility routing can be proven by interface tests. Cross-segmentation stability, shared object representations, or security robustness require trained tasks, baselines, and ablations.

Open the Claims MatrixIntegrate an existing model