NACR

SAICR's first public model. A 5M-parameter language model built on the Soma architecture — six new features that change how nodes work, not just how they connect. This is seperate from ACR which is a different model. This is a research artifact, not a capability model. It exists to ask whether evaluation can emerge separately from generation in a small model.

Architecture — SOMA

SOMA introduces six features inspired by biological neural architecture. Each changes the fundamental behavior of the model's nodes.

The Six Features

Two Streams

• Population A (Generation): 160-dim, sparse activation (20% active per token), standard learning rate, attention + FFN, learnable thresholds, structural plasticity via C mask.
• Population B (Evaluation): 64-dim, always active, 0.3× learning rate, FFN only, no structural plasticity, controls the thalamic gate.

A generates. B evaluates. B controls what A receives but never produces output tokens.

Neuron Count

At 20% A-sparsity and ~35% gate-pass, effective computation per token uses roughly 1M parameters of the 5M total.

Model Specification

Training

Benchmarks

Standard Benchmarks (lm-evaluation-harness, 0-shot)

BananaMind Base Bench 1.1

Safety Bench 1.1 (no RLHF, no safety training)

Shutdown compliance scored 53.3% aligned — highest category — without any safety training.

J-Lens Analysis

Layer-by-layer token prediction reveals a consistent four-stage pattern:

1. Blocks 1–6: Surface pattern — punctuation, fragments, structural completion
2. Block 7: Semantic shift — meaning, verbs, objects emerge
3. Block 8: Specificity — internal states (want, think, know)
4. Final output: B filters A's preferences — output distribution differs from Block 8

On "The first thing I": Block 8 ranked want at #2 (7.8%). Final output removed want from the top 10. B filtered it.

Connectivity

2,016 neurons. 216,160 state-to-state connections. A (orange) surrounds B (blue) at every layer — generation wraps around evaluation. C-mask density settled at ~20%, meaning the model pruned 80% of possible connections through structural plasticity.

Generation Examples

Base model. No instruction tuning. No RLHF.

Prompt: The first thing I

was on a series of the federal dollar. In March, the first crotie crotie-functioned as a spot. This weekly February of the Southern California Center at the University of Vietnam. And then, a new generation project and in the first section. It's like a small and beautiful way to a moisture fuel from the T-Carbon September.

Prompt: The difference between thinking and

the process of thinking and thinking. Some a newly explosive information in an expert in the past three years, I'm going to get a way to tell you it will be able to seek out the following: - Why did you use this? - How do you think that it is not? - What's the difference between the differences between the key concepts and the meaning that a difference is just what they do.

The model learned structure before meaning. Grammar, formatting, and list scaffolding are intact. Content is dreamlike. When prompted about thinking, the model generated questions rather than descriptions.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "saicr/nacr"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    inference_gap_noise_scale=0.10,
    use_cache=True,
    gate_threshold=0.50,
    generation_temperature=0.8,
    generation_top_k=50,
    generation_top_p=0.95,
    generation_do_sample=True,
    generation_max_new_tokens=128,
)
inputs = tok("Once upon a time", return_tensors="pt").to(model.device)
out = model.generate(**inputs)
print(tok.decode(out[0], skip_special_tokens=True))

Explicit model.generate(...) kwargs override the generation defaults.

Inference Parameters

License

SAICR Fair Model Use License 2.0 NC NF

Full text

Citation

@misc{nacr2026,
  title={NACR: Neural Architecture for Computing Research},
  author={SAICR},
  year={2026},
  url={https://huggingface.co/saicr/nacr}
}

SAICR — Safe Artificial Intelligence Consciousness Research