| --- |
| license: mit |
| tags: [from-scratch, opcode, ternary, trit, execution-verified, tiny] |
| --- |
| |
| # Binary-Trit Coder (Clay's opcode model) |
|
|
| A **from-scratch, 3.17M-param causal transformer** trained from *random init* on Issac's own opcode |
| substrate β a bijective ternary ("trit") op grammar: `join (+)`, `take (β)`, `weave (Γ)` over digit leaves, |
| prefix-notation, depth-2. No pretraining, no outside corpus. |
|
|
| ## Honest benchmark β its OWN domain, not Python |
| The metric is **execution-verify rate**: sample a program, *run it for real*, check the model's answer. |
| | Test | Score | |
| |---|---| |
| | in-distribution (depth-2, fresh) | **98.6%** | |
| | OOD generalization (depth-3, never trained) | 47.0% | |
|
|
| **This is NOT a Python code-gen model.** It has an 18-token opcode vocab; it will score ~0 on |
| LiveCodeBench/SciCode (wrong domain). Its "learn from doing" signal is the verify-rate above. |
|
|
| ## The chrysalis (capability layers, all execution-verified) |
| - variables/let-bindings (69.5% end-to-end; 100% with a calculator + show-work) |
| - ping-pong Ο-lattice solver β recovers a hidden intermediate 100% |
| - calc-offload, dark-space inference, combustion candidate-spark |
|
|
| ## Files |
| `scratch_coder.pt` (weights + vocab), `config.json`. Architecture + inference in |
| [github.com/issdandavis/loom](https://github.com/issdandavis/loom) `scratch_coder.py`. |
|
|
| *Thesis: coding is finite + bijective + execution-verifiable, so a small model from random init can MASTER |
| the mapping β proven here at 98.6%.* |
|
|
