ARC_NEURON_SMALL_AND_CORPUS_COMPILER.md

ARC-Neuron Small + Corpus Compiler v1

This package promotes the ARC stack from Tiny-only proof into the first native-growth milestone:

• unified corpus compiler across ARC stack source packages
• ARC-Neuron Small research model
• ANCF canonical wrapper
• GGUF deployment artifact
• Omnibinary learning ledger
• Arc-RAR release bundle

Corpus compiler v1

The compiler pulls constrained text surfaces from the uploaded ARC stack packages:

• ARC Language Module
• ARC Cognition Core
• ARC Core
• Cleanroom Runtime
• Omnibinary Runtime
• Arc-RAR

Outputs:

• datasets/arc_neuron_small/arc_neuron_small_sft.jsonl
• datasets/arc_neuron_small/arc_neuron_small_next_token.txt
• reports/arc_neuron_small/corpus_manifest.json

ARC-Neuron Small

The model is a byte-level decoder-only transformer research build with a larger configuration than ARC-Neuron Tiny:

• block size: 128
• layers: 3
• heads: 4
• embedding dim: 64

Artifacts:

• artifacts/gguf/ARC-Neuron-Small-0.18M-v0.1-F32.gguf
• artifacts/ancf/ARC-Neuron-Small-0.18M-v0.1.ancf
• artifacts/omnibinary/arc-neuron-small-learning-ledger.obin
• artifacts/archives/arc-neuron-small-release.arcrar.zip

Operator flow

scripts/operator/build_arc_neuron_small.sh

Boundary

This remains a native research model and not a frontier foundation model. The value of this step is to prove:

• unified corpus shaping
• native model growth inside ARC rules
• canonical ANCF + deploy GGUF pairing
• release receipts through Omnibinary and Arc-RAR