Add a preset

A preset is an assertion about a real model, so it has to be checkable. The twenty of them are the regression suite.

Write the document

The library is packages/core-go/presets/data: twenty-three JSON documents and an index.json listing them, embedded into the binary by go:embed. There is no builder to go through — a preset is a document in exactly the format the editor saves, and the format import writes.

If the model has a Hugging Face config.json, start there:

bun packages/cli/src/index.ts import config.json \
  --name my-model-7b \
  --out packages/core-go/presets/data/my-model-7b.json

The importer reproduces eight of the presets to the parameter. It prints the count the analysis gets, to be held against the model card, and every warning about where the document is not the model — a multi-token-prediction head left out, layers made sparse that the model keeps dense. Each warning gets fixed by hand or written into the notes; none of them are swallowed. The families it knows are gpt2, llama, mistral, mixtral, qwen2, qwen3, qwen3_moe, gemma, gemma2 and deepseek_v3, and it refuses anything else by name rather than guessing.

--out puts the document straight into the library; the name still has to go into index.json.

Otherwise start from the nearest preset rather than an empty file. Almost every decoder-only model differs from llama-3-8b.json in seven numbers and a note:

"symbols": {
  "L":   { "kind": "design", "value": 32,     "doc": "Number of transformer layers" },
  "D":   { "kind": "design", "value": 4096,   "doc": "Residual stream width (d_model)" },
  "H":   { "kind": "design", "value": 32,     "doc": "Query heads" },
  "Hkv": { "kind": "design", "value": 8,      "doc": "Key/value heads" },
  "dh":  { "kind": "design", "value": "D/H",  "doc": "Head dimension" },
  "F":   { "kind": "design", "value": 14336,  "doc": "Feed-forward hidden width" },
  "V":   { "kind": "design", "value": 128256, "doc": "Vocabulary size" }
}

The graph below them is four or five nodes, because transformer_block and the composites inside it carry the architecture. Change meta.name, meta.published, the notes, and whatever the model does differently — the normalization, the activation, the RoPE theta — and add the file to index.json.

Two presets are not transformers at all and are written out block by block: ijepa-vit-h14.json has bidirectional attention and no vocabulary, and alexnet.json is convolutional with B C H W tensors. Copy those instead when that is what you are describing.

published is the point

An import gives you meta.name and meta.family. The claim is the rest, and either way you write it yourself:

• params — what the authors report, with a source link to the config or paper.
• tolerance — only where the published figure is itself rounded ("22B active"). Say so in the notes.

If your number disagrees with the one people quote, find out why and write it down. I-JEPA's ViT-H is 630.4M, not the 632M everyone cites, because its positions are frozen sincos and it has no class token. That sentence is worth more than the number.

Check it

bun run scripts/report.ts

Your preset should say exact. Every other row must be unchanged.

Then against real PyTorch:

bun run scripts/codegen-demo.ts my-model-7b
python -m tensorcad_runtime verify out/my-model-7b/model.py

See Verify a design against PyTorch.

Write down what it says

The golden files are the specification, so a new preset needs its own:

cd packages/core-go
go run ./cmd/golden    # writes testdata/{golden,analysis,rules,codegen}/my-model-7b.json
go test ./...

Read the diff. Four new files appear and nothing else should move; a number that changed under an existing preset means the preset was not the only thing you touched.