LOOM: Universal Bit-Perfect Deterministic AI Engine

"The SQLite of AI" — A Polyglot Neural Engine with Bit-Exact Reproducibility

Loom is a Deterministic Neural Virtual Machine (DNVM) engineered for absolute numerical consistency and extreme efficiency. It guarantees bitwise-identical results across all platforms, backends, and language bindings, bypassing memory bandwidth bottlenecks through polymorphic dispatch and volumetric 3D modeling.

🌐 The Polyglot Solution

Loom is designed as a universal runtime that prioritizes portability and sovereignty:

• True "Copy-Paste" Portability: Models are language-agnostic. Move weights and logic between Go, Python, C#, and WASM without translation layers.
• Write Once, Run Everywhere: A standardized format that performs identically on Browser (WASM/WebGPU), Mobile (iOS/Android), and Desktop (Linux/Windows/macOS).
• Universal Import: Direct ingestion from major frameworks—zero vendor lock-in.
• Active Edge Training: Full backpropagation enabled on-device. No "frozen brains"; Loom learns from user interaction at the edge.
• Sovereign & Private: Zero cloud dependencies. User data and model execution remain 100% local.

💎 The Bedrock Philosophy

Loom is a "Bedrock Edition" neural engine. Unlike standard frameworks that build on top of high-level abstractions, Loom is designed at the bit-level to bypass the physical memory limitations of consumer hardware.

• Cross-Platform Determinism: 0.0000000000 difference between CPU and GPU, x86 and ARM, native and browser.
• Universal Precision: Native support for 21 numerical types (FP64 to 1-bit Binary), allowing Loom to "morph" precision to match specific silicon preferences.
• Bit-Perfect Identity: Verified across hundreds of permutations with 0.000000% mathematical divergence.

🚀 The Technical Pillars (Final Form)

The project has transitioned to the Multi-numerical POLYmorphic Volumetric Tiled-tensor Dispatcher (M-POLY-VTD) core.

• Step neural mesh: A living mesh architecture with clock-cycle accurate updates and temporal feedback loops that simulate biological neural firing.
• DNA Engine: A hierarchical spatial correlation system that extracts topological "signatures" of models, enabling high-fidelity comparison and "Logic Shift" detection in 3D space.
• Tween (neural target propagation): A robust alternative to backpropagation that uses localized, gap-based Hebbian learning to bridge the difference between actual and idealized activations. We call it tween in code (tween.go); papers often use target propagation or related names.
• Bit-Packed Persistence: An idempotent serialization tunnel that achieves up to 98.4% compression, allowing extreme model sizes to fit in consumer RAM/VRAM.

📂 Project Structure

• poly/: The current-generation engine core (M-POLY-VTD). Active development.
• lucy/: Interactive harness — Poly Talk, ENTITY Talk, seven-layer CPU suite, HF download.
• planetbridging/: Planet → Loom bridging POC (v0.5.0 complete in-tree; releases after Loom 0.80).
• welvet/dart/: Flutter/Dart FFI plugin — welvet on pub.dev (flutter pub add welvet).

🛠️ Getting Started

For technical deep-dives into M-POLY-VTD, refer to the docs/ index (docs/index.md) and benchmarks within the poly/ core. Topics include deployment, GPU, layers — plus donate compute (LAN TCP, docs/donate_compute.md) and TANHI UDP layer telemetry for the SoulGlitch HUD (docs/tanhi.md).

Loom provides bit-exact reproducibility across:

• Go (Native)
• TypeScript/Node.js (@openfluke/welvet)
• Browser (WASM + WebGPU)
• Python (pip install welvet)
• Flutter/Dart (welvet FFI plugin — iOS, Android, Linux, macOS, Windows)
• C#/.NET (Welvet) - (In Development)

📊 Versioning & Roadmap

Loom uses a mathematical versioning system derived from a strictly verified checklist in poly/README.md (row counts and completion ratio are maintained there).

Current Version: 0.81.0 — CURRENT (from 0.80.0)

• Completion Ratio: 76.7% (112 / 146 checklist rows on adjustments)
• Codename: 0.81.0 "Accelerator Bridge" — experimental Intel CPU+NPU via poly/accel; Lucy [9]; Qualcomm NPU and Google TPU on roadmap.
• Status: Vendor plugin model shipped on Linux (CGO_ENABLED=1); GPU + ENTITY from 0.80 remain production paths. See docs/v081_release.md and docs/accelerators.md.
• Milestones:
  • v0.79.0 "Bedrock Validation" ✅ — See docs/bedrock_validation.md.
  • v0.80.0 "Native Ship" ✅ — See docs/v080_release.md.
• Next Target — v0.82: AccelPlanner + JSON exec; GPU backward (SwiGLU/MHA); Qualcomm/Google CABI stubs; Donate Compute live inference.

For a detailed breakdown of the roadmap and version calculation, see poly/README.md.

---

License

Apache License 2.0 - see LICENSE file for details.

---

Loom: Universal precision. Volumetric freedom. Bedrock performance.

Made with ❤️ by Openfluke