MetaphyKing

# STELLAR — Lossless + Structured-Lossy AI Weight Compression

## Public progress note from Metaphy LLC — 2026-05-01

**Author:** Randell Logan Smith ([@MetaphyKing](https://x.com/MetaphyKing) — Metaphy LLC, [metaphysicsandcomputing.com](https://www.metaphysicsandcomputing.com))
**AI co-builder:** Anthropic's Claude Opus (sessions across 2026-04 and 2026-05)
**Date:** 2026-05-01
**Purpose:** Public, timestamped record of the conclusions and implications below, established as of this date.

> This note is deliberately scoped. The architectural details that underpin our v2 substrate are intentionally omitted because they are subject to a pending patent. Everything reported below — every ratio, every benchmark, every methodology choice — is an actual measurement against publicly-available models, reproducible from the published numbers using a standard FP32 weight pipeline.

MetaphyKing

"""
The most atomic way to train and run inference for a GPT in pure, dependency-free Python.
This file is the complete algorithm.
Everything else is just efficiency.

@karpathy
"""

import os       # os.path.exists
import math     # math.log, math.exp