File size: 1,748 Bytes
3bee9e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
---

license: mit
tags:
- from-scratch
- opcode
- ternary
- trit
- execution-verified
- tiny
- superseded
---

<!-- scbe-status -->
> **Status: superseded.** Canonical coding model: [scbe-coding-agent-vtc-qwen15-v1-gguf](https://hf.co/issdandavis/scbe-coding-agent-vtc-qwen15-v1-gguf). Kept as research history.

# 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 — a bijective ternary ("trit") op grammar: `join (+)`, `take (−)`, `weave (×)` over digit leaves,
prefix-notation, depth-2. No pretraining, no outside corpus.

## Honest benchmark — 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 φ-lattice solver — 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 — proven here at 98.6%.*