Codette Llama 3.1 8B — Merged Orchestrator Base

Llama 3.1 8B Instruct with the Codette Orchestrator LoRA permanently merged into the base weights. This is the inference base for the Codette multi-perspective reasoning system — pair it with the perspective LoRA adapters for full multi-agent synthesis.

Paper: Codette: Multi-Perspective Reasoning as a Convergent Dynamical System GitHub: Raiff1982/Codette-Reasoning ORCID: 0009-0003-7005-8187

Benchmark Results (May 2026)

17-problem benchmark across 6 cognitive categories, 4-condition ablation:

Condition	Composite Score	vs. Baseline
SINGLE (baseline)	0.357	—
MULTI (6 perspectives)	0.521	+46.1%
MEMORY (+ cocoon store)	0.574	+60.8%
CODETTE (full system)	0.744	+108.8%

Cohen's d = 8.31 (large effect; d > 0.8 is large by convention)
Paired t-test: p < 0.0001
Turing naturalness: 0.245 → 0.820 (+235%) — depth–naturalness tradeoff resolved
Coherence: 0.477 → 0.700

GPQA (graduate-level science, 0-shot, 198-question diamond set):

Run	Accuracy	Environment
Base model + adapters (Kaggle cloud, June 17 2026)	27.8% (55/198)	Direct transformers+PEFT, no orchestration
Full Codette system (local server, June 6 2026)	30.8% (61/198)	Multi-agent debate + coherence tracking + cocoon memory

Baselines: random 25%, GPT-4 0-shot 39%, human expert 65%. The ~3pp gap between runs quantifies the system layer's contribution on GPQA specifically.

Model Details

Property	Value
Base Model	meta-llama/Llama-3.1-8B-Instruct
Merged Adapter	Orchestrator LoRA
Format	SafeTensors (full precision, ~16 GB)
Context Length	8192 tokens
Quantized version	codette-llama-3.1-8b-gguf

System Architecture

Query
  │
  ▼
Executive Controller (complexity routing)
  │
  ▼
Merged Orchestrator Base  ◄── this repo
  │
  ▼
LoRA Hot-Swap (newton / davinci / empathy / philosophy /
               quantum / consciousness / multi_perspective /
               systems_architecture)
  │
  ▼
Multi-Agent Debate + Semantic Tension (RC+ξ)
  │
  ▼
AEGIS Ethical Governance (6 frameworks)
  │
  ▼
Synthesized Response + Cocoon Memory

The RC+ξ (Recursive Convergence + Epistemic Tension) formalism models cognitive state evolution as a convergent dynamical system:

Ψ(t+1) = Ψ(t) + α·∇Coherence(Ψ(t)) − β·ξ(t)·∇Tension(Ψ(t))

Quick Start

With Transformers (full precision)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Raiff1982/codette-llama-3.1-8b-merged")
tokenizer = AutoTokenizer.from_pretrained("Raiff1982/codette-llama-3.1-8b-merged")

inputs = tokenizer("Explain the nature of consciousness", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With 4-bit quantization (recommended for 8–16 GB VRAM)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16)
model = AutoModelForCausalLM.from_pretrained(
    "Raiff1982/codette-llama-3.1-8b-merged",
    quantization_config=bnb, device_map="auto"
)

With perspective adapters (multi-agent mode)

from peft import PeftModel

# Apply a perspective adapter on top of the base
model = PeftModel.from_pretrained(model, "Raiff1982/codette-lora-adapters",
                                  subfolder="newton_v2")

Full local server

git clone https://github.com/Raiff1982/Codette-Reasoning
cd Codette-Reasoning
python inference/codette_server.py  # serves on :7860

Related Resources

Resource	Link
Perspective LoRA adapters	codette-lora-adapters
Quantized GGUF	codette-llama-3.1-8b-gguf
Training datasets	codette-training-data
GitHub	Raiff1982/Codette-Reasoning
Paper (preprint)	Research Square DOI
Zenodo archive	10.5281/zenodo.19480004
Kaggle AGI benchmark	RC+ Diagnostic Suite