Loom v0.73.0 (Alpha) — M-POLY-VTD Engine

Loom v0.73.0 — M-POLY-VTD Alpha

First release of the M-POLY-VTD (Multi-numerical POLYmorphic Volumetric
Tiled-tensor Dispatcher) architecture.

---

What is M-POLY-VTD?

A neural network engine where:

• Every layer independently holds any of 21 numerical types — from Float64
  down to Binary — switchable at runtime without retraining
• Networks are 3D spatial grids, not flat chains — layers have (z, y, x, l)
  positions with support for remote cyclic links
• Execution is systolic — a clock advances state through the mesh, enabling
  BPTT through space-time
• Training supports target propagation (gap-based, no global chain rule)
  alongside standard backprop
• GPU acceleration via WebGPU — runs on any vendor, including in the browser

---

What's included

Core engine — full forward + backward for:
Dense · RMSNorm · SwiGLU · Multi-Head Attention · Embedding · Residual
CNN1D · CNN2D · CNN3D · LSTM · RNN · KMeans · Softmax · LayerNorm

C-ABI — 144 exported symbols, usable from any language that supports FFI

Platform binaries (12 targets):

Platform	File
Linux x86_64	welvet.so
Linux arm64	welvet.so
macOS arm64	welvet.dylib
macOS x86_64	welvet.dylib
macOS Universal	welvet.dylib
Windows x86_64	welvet.dll
Android arm64	welvet.so
Android x86_64	welvet.so
iOS arm64	welvet.a
iOS sim amd64	welvet.a
iOS sim arm64	welvet.a
iOS XCFramework	Welvet.xcframework

Each binary ships with welvet.h and a cabi_verify diagnostic tool.

---

Not in this release (coming in 0.74.0)

• pip install openfluke-welvet (Python package)
• npm install @openfluke/welvet (TypeScript/WASM package)
• Windows arm64, Linux armv7, Android armv7 binaries

---

Verification

Run the included cabi_verify against your platform's binary:

./cabi_verify ./welvet.so    # Linux
./cabi_verify ./welvet.dylib # macOS
Expected: All 144 C-ABI symbols are present.

This is an Alpha release. APIs are stable within the 0.7x series but
may change before 1.0.