N-Orca — Neural-network Orchestrated Architecture Language

N-Orca is a Markdown DSL for declaring, verifying, visualizing, and executing neural network architectures. It is a domain-specific dialect of Orca — Orca describes finite-state machines, N-Orca describes typed DAGs of tensor-flow layers.

The point of the language is to give LLMs (and humans) one artifact that covers all five jobs:

Affordance	How
Author	Plain Markdown — LLMs already produce GitHub-flavored tables and headings reliably. No new tokens to memorize.
Read	## section headings give strong landmarks; one row = one layer; blockquotes describe intent next to declarations.
Visualize	Compile to Mermaid flowchart TD — any markdown viewer renders the diagram.
Explain	The verifier emits structured error codes (e.g. SHAPE_MISMATCH, UNREACHABLE_LAYER) with suggestion: fields written for an LLM to act on.
Execute	Compile to a runnable PyTorch nn.Module. The verifier guarantees the emitted code's shapes line up before you run it.

---

What it looks like

# architecture TransformerBlock

## hyperparameters

| Name    | Type  | Default |
|---------|-------|---------|
| d_model | int   | 512     |
| n_heads | int   | 8       |
| d_ff    | int   | 2048    |
| dropout | float | 0.1     |

## tensors

| Name | Shape           | Dtype   |
|------|-----------------|---------|
| x    | (B, S, d_model) | float32 |
| y    | (B, S, d_model) | float32 |

## layer x [input]
## layer attn_norm
- op: LayerNorm(d_model)
## layer attn
- op: MultiHeadAttention(d_model, n_heads, dropout)
## layer add_1
- op: Add
## layer ff_norm
- op: LayerNorm(d_model)
## layer ff
- op: FeedForward(d_model, d_ff, dropout)
## layer add_2
- op: Add
## layer y [output]

## flow

| Source     | Target    | Tensor    |
|------------|-----------|-----------|
| x          | attn_norm | x         |
| attn_norm  | attn      | x_normed  |
| attn       | add_1     | attn_out  |
| x          | add_1     | x_skip    |
| add_1      | ff_norm   | h         |
| ff_norm    | ff        | h_normed  |
| ff         | add_2     | ff_out    |
| add_1      | add_2     | h_skip    |
| add_2      | y         | y_out     |

## invariants
- output_shape: (B, S, d_model)

That's a real, verifying example. n-orca verify confirms reachability, shape consistency through every residual, parameter count, and the declared output-shape invariant — before any code runs.

---

Install

pip install n-orca                # core CLI, parser, verifier, compilers
pip install "n-orca[torch]"       # + PyTorch (compile to runnable nn.Module)
pip install "n-orca[hf]"          # + huggingface_hub (search, info, download, convert)
pip install "n-orca[mcp]"         # + MCP server (drive n-orca from Claude / any MCP client)
pip install "n-orca[all]"         # everything

Or from source:

git clone https://github.com/jascal/n-orca && cd n-orca
pip install -e ".[all]"

Python 3.10+. The core has zero runtime dependencies; huggingface_hub, torch, and mcp are optional extras and lazy-imported at the point of use.

---

Commands

# Verify a file
n-orca verify examples/transformer-block.n.orca.md
n-orca verify examples/transformer-block.n.orca.md --json
n-orca verify examples/transformer-block.n.orca.md --strict

# Compile
n-orca compile mermaid  examples/transformer-block.n.orca.md
n-orca compile pytorch  examples/transformer-block.n.orca.md
n-orca compile pytorch  examples/transformer-block.n.orca.md --out model.py

# Summarize
n-orca info examples/transformer-block.n.orca.md

Sample verify output:

Architecture: TransformerBlock
  Result: VALID
  Parameters: 3,152,384
  Depth: 7

When a check fails, you get a stable error code, a message, and a suggestion:

Architecture: ResidualBlock
  Result: INVALID
  [ERR] SHAPE_MISMATCH: layer 'add': Add inputs have mismatched shapes:
        [('B', 'C', 'H', 'W'), ('B', '2C', 'H', 'W')]
        -> check the layer's op args and incoming shapes

---

Verification pipeline

Five stages, run in order. A failure in an earlier stage stops later stages on that architecture (their preconditions wouldn't hold).

Stage	What it checks	Example error codes
1 — Naming	every flow-edge endpoint resolves; one [input] and [output]; no duplicate layer names	UNKNOWN_LAYER_REFERENCE, NO_INPUT_LAYER, DUPLICATE_LAYER
2 — Structural	DAG (no cycles); every layer reachable from an input; every layer reaches an output	CYCLE_DETECTED, UNREACHABLE_LAYER, LAYER_NOT_REACHING_OUTPUT
3 — Shape	each layer's input shapes match its op's input rule; declared shape: matches inferred	SHAPE_MISMATCH, INPUT_ARITY_MISMATCH, DECLARED_SHAPE_MISMATCH
4 — Resource	param_count / depth / output_shape against ## invariants	PARAM_BUDGET_EXCEEDED, DEPTH_BUDGET_EXCEEDED, OUTPUT_SHAPE_INVARIANT
5 — Op coverage	every layer's op exists in the standard library (warning if not)	UNKNOWN_OP

See docs/verification.md for the full error catalog with examples.

---

Standard op library

Op	Maps to	Use
Linear(in, out)	nn.Linear	dense layer
LayerNorm(d), BatchNorm1d(c), BatchNorm2d(c)	matching nn.*Norm	normalization
Conv2d(ic, oc, k, s, p)	nn.Conv2d	2D conv
MaxPool2d(k, s), AvgPool2d(k, s), AdaptiveAvgPool2d(out)	matching nn.*Pool*	spatial pooling
ReLU, GELU, SiLU, Tanh, Sigmoid, Softmax(dim)	matching nn.*	activations
Dropout(p)	nn.Dropout	regularization
Embedding(n, d)	nn.Embedding	token / position embeddings
PatchEmbed(c, d, p)	nn.Conv2d + flatten + transpose	image → patch-token sequence (ViT / I-JEPA)
TubeletEmbed(c, d, t, p)	nn.Conv3d + flatten + transpose	video clip → tubelet-token sequence (V-JEPA 2)
MultiHeadAttention(d, h, dropout)	nn.MultiheadAttention (batch-first)	self-attention
FeedForward(d, d_ff, dropout)	nn.Sequential(Linear, GELU, Dropout, Linear, Dropout)	transformer FFN
Add, Mul	functional + / *	residual / gated paths
Concat(dim)	torch.cat	skip connections
Mean(dim)	tensor.mean(dim=…)	sequence / global pooling
Flatten(start_dim), Reshape(shape), Identity	matching torch ops	shape glue

Unknown ops are accepted with an UNKNOWN_OP warning and emit nn.Identity() placeholders in PyTorch output — useful for prototyping; the verifier still checks everything around them.

See docs/grammar.md for the full language specification.

---

Hugging Face Hub integration

# Search the Hub
n-orca hf search "gpt2" --limit 10
n-orca hf search "llama" --task text-generation

# Inspect a model (metadata + config.json)
n-orca hf info gpt2
n-orca hf info meta-llama/Llama-2-7b-hf --revision <commit-sha>

# Download just config.json, or the full snapshot
n-orca hf download gpt2 --config-only
n-orca hf download gpt2 --allow "*.safetensors" --local-dir ./gpt2

# Vision / video models: also grab the (video) preprocessor config
n-orca hf download facebook/vjepa2-vitl-fpc64-256 --config-only --include-processor

# Convert an HF model -> .n.orca.md (+ optional Mermaid)
n-orca hf convert gpt2 --out gpt2.n.orca.md --mermaid gpt2.mmd
n-orca hf convert meta-llama/Llama-2-7b-hf --out llama-7b.n.orca.md

# JEPA / V-JEPA 2 / LeWorldModel world models
n-orca hf convert facebook/vjepa2-vitl-fpc64-256 --out vjepa2.n.orca.md --mermaid vjepa2.mmd
n-orca hf convert quentinll/lewm-pusht --out lewm-pusht.n.orca.md

The convert command reads config.json only — no model weights or remote code are executed. It picks an adapter by matching config["model_type"] or config["architectures"] against a registry, builds the topology, and writes a verified .n.orca.md.

Supported model families

Adapter	Matches	Notes
Gpt2Adapter	gpt2, gpt_neo, gpt_neox, openai-gpt	Decoder-only, pre-LN, learned positional embeddings
LlamaFamilyAdapter	llama, mistral, mixtral, qwen2, qwen3, qwen2_moe, gemma, gemma2, phi, phi3	Decoder-only, RMSNorm + RoPE + SwiGLU (approximated in v1)
BertAdapter	bert, roberta, distilbert, electra, albert	Encoder-only, post-LN; segment-type embeddings omitted in v1
EsmAdapter	esm	ESM-2 protein LM; pre-norm encoder, rotary positions left implicit
JepaAdapter	vjepa2, ijepa, jepa, leworldmodel, lewm (+ VJEPA2Model / IJepaModel / LeWM Hydra configs)	Joint-embedding world models: ViT encoder → predictor. Video tubelet (Conv3d) or image patch (Conv2d) embed; optional action conditioning + projector. Mask tokens, EMA target, SIGReg (latent regularizer) captured as verification rules

JepaAdapter normalizes two unrelated config schemas into one encoder → predictor DAG: the flat transformers config (V-JEPA 2 / I-JEPA, with pred_* predictor fields) and the nested Hydra config used by LeWorldModel (quentinll/lewm-*), which carries no model_type and is matched structurally. Both latent outputs (encoder_latents, predicted_latents) live in the encoder's embedding space, so the output_shape invariant enforces latent-dim consistency between the representation and the forecast.

Adding a new family is a ~50-line file in n_orca/hf/adapters/ — declare model_types and implement build(config, name=...) to return an Architecture AST.

Library API

from n_orca.hf import HfClient, convert

client = HfClient()
results = client.search("llama", task="text-generation", limit=5)
for r in results:
    print(r.id, r.downloads)

info = client.info("gpt2")
print(info.config["model_type"], info.config["n_layer"])

# Convert (reads config.json directly — no weights needed)
result = convert("gpt2")
result.write_markdown("gpt2.n.orca.md")
result.write_mermaid("gpt2.mmd")
print(f"params: {result.report.param_count:,}, valid: {result.report.valid}")

Pre-generated example outputs live in examples/hf-generated/: gpt2-small.n.orca.md, bert-base-uncased.n.orca.md, llama-7b.n.orca.md, tinyllama-2L.n.orca.md, vjepa2.n.orca.md (V-JEPA 2 ViT-L, 326M params), lewm-pusht.n.orca.md (action-conditioned LeWorldModel) — each with a matching .mmd diagram.

---

Examples

All examples in examples/ verify clean and round-trip through the PyTorch compiler with a real forward pass:

Example	Architecture	Params (default)
simple-mlp.n.orca.md	2-layer MLP classifier	203,530
residual-block.n.orca.md	Pre-activation ResNet-v2 block	74,112
transformer-block.n.orca.md	GPT-style pre-norm encoder block	3,152,384
tiny-vit.n.orca.md	Minimal Vision Transformer	205,834
conv-classifier.n.orca.md	Small CNN classifier	20,042
unet-stub.n.orca.md	Two-level U-Net skeleton with skip-concat	3,057

---

Tests

pip install -e ".[test,torch]"
pytest -v

tests/test_cli.py                 7 passed
tests/test_hf_adapters.py        28 passed
tests/test_hf_cli.py              4 passed
tests/test_hf_client.py           8 passed
tests/test_mermaid.py             4 passed
tests/test_mcp_server.py          5 passed
tests/test_ops.py                22 passed
tests/test_parser.py             11 passed
tests/test_pytorch.py             8 passed
tests/test_render.py              2 passed
tests/test_sae_and_world_models.py  24 passed
tests/test_verifier.py           16 passed
=========================
139 passed in 13.29s

---

Design philosophy

N-Orca borrows three load-bearing ideas from Orca and Q-Orca:

1. The spec is the artifact the LLM edits. Not a Python file with a sidecar comment; not a JSON config alongside Python. One Markdown file is the source of truth.
2. Tables, not new syntax. LLMs hallucinate the least on flat, row-oriented tables. Layers are one section; flow edges are one section; each row is one fact.
3. Verify before execute. The pipeline checks topology (DAG-ness, reachability), then shapes (the place neural network code most often breaks at runtime), then resource bounds — before the PyTorch compiler emits anything. By the time you forward() the model, the architecture has already been checked.

The net effect: an LLM can write an architecture in Markdown, get structured verifier feedback, refine, and ship runnable code — without ever editing Python.