File size: 4,044 Bytes
aa5ed92
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
# lean-llm-starter

A deterministic Lean 4 verification harness that treats the LLM as an untrusted proposal engine and Lean 4 as the trusted kernel.

## Architecture

```text
operator intent
  -> Prolog gate
  -> Granite 4.1 / Llemma inference
  -> schema validation
  -> Lean 4 parse / verify
  -> WORM-ready audit artifacts
```

Trust split:

- `inference/` proposes proof artifacts
- `logic/` blocks unauthorized or malformed runs
- `lean4/` type-checks and verifies
- `infra/verification-loop/` pins deterministic execution
- `eval/` runs reproducible local benchmarks

## Repo Layout

```text
lean-llm-starter/
├── .github/workflows/ci.yml
├── .gitattributes
├── .gitignore
├── LICENSE
├── Makefile
├── README.md
├── docker-compose.yml
├── docs/
│   ├── HUGGINGFACE_PUBLISHING.md
│   └── REPO_MAP.md
├── eval/
│   ├── requirements.txt
│   └── run_minif2f.py
├── fixtures/
│   └── sample_input.jsonl
├── inference/
│   ├── Dockerfile
│   ├── requirements.txt
│   ├── prompt.txt
│   └── server.py
├── infra/
│   └── verification-loop/
│       ├── .env.example
│       └── docker-compose.yml
├── lean4/
│   ├── lakefile.toml
│   ├── lean-toolchain
│   ├── MiniF2F.lean
│   ├── VerifyMain.lean
│   └── src/
│       └── SovereignCorpus/
│           ├── Bridge/
│           │   ├── Granite4Parser.lean
│           │   └── Granite4Schema.lean
│           ├── Core.lean
│           └── Tactics/
│               └── PlasmaGate.lean
├── logic/
│   ├── sovereign_verification.pl
│   └── verification_loop.pl
└── hf/
    └── README.md
```

## What This Starter Includes

- Lean 4 harness skeleton
- Granite 4.1 JSON interchange schema
- Lean-side parser scaffold
- Prolog gate and retry loop scaffold
- vLLM / docker compose verification loop
- MiniF2F-style eval harness
- GitHub Actions CI scaffold
- Hugging Face publishing docs and model card template
- sample JSONL parse fixture

## What It Does Not Pretend Yet

- not fully air-gapped out of the box
- not fully air-gapped out of the box
- not a finished proof search system
- not yet shipping weights

Current verified state:

- Lean 4 project builds successfully on this machine with `C:\Users\jessi\.elan\bin\lake.exe`
- Python eval and inference files compile
- parse-mode fixture is present

This is a scaffold you can harden into:

- a GitHub repo
- a Hugging Face model or Space companion repo
- a deterministic local verification loop

## Quick Start

If `lake` is not on your PATH on Windows, use:

```powershell
C:\Users\jessi\.elan\bin\lake.exe
```

### 1. Lean side

```bash
cd lean4
C:\Users\jessi\.elan\bin\lake.exe update
C:\Users\jessi\.elan\bin\lake.exe build
```

### 2. Inference side

```bash
cd inference
docker build -t lean-llm-inference .
docker run -d -p 8080:8080 --name lean-llm-inference lean-llm-inference
```

### 3. Eval side

```bash
cd eval
pip install -r requirements.txt
python run_minif2f.py
```

### 4. Full verification loop

```bash
docker compose --env-file infra/verification-loop/.env.example up -d granite-verifier
swipl -g "verify_with_retries('theorem demo : True := by trivial', 'ED25519_SIG', 3, Result), writeln(Result), halt" logic/verification_loop.pl
```

### 5. Shortcut targets

```bash
make lean-build
make infra-up
make eval
```

### 6. Parse smoke test

```bash
cd lean4
C:\Users\jessi\.elan\bin\lake.exe exe verify -- --parse ../fixtures/sample_input.jsonl
```

## Hugging Face Readiness

This repo is prepared for later Hugging Face publication with:

- `.gitattributes` for LFS-managed weight files
- `hf/README.md` model card template
- `docs/HUGGINGFACE_PUBLISHING.md` publish checklist

Recommended publish modes:

1. code-only harness repo
2. model repo for GGUF / safetensors
3. Space repo for interactive verify loop UI