AOS-Shadow-v1.0

ArmadaOS Shadow Verification Agent — First Custom Model

AOS-Shadow-v1.0 is a specialized AI agent model fine-tuned from Qwen 3.5 9B for the Shadow Verification role within the ArmadaOS agent operating system. This is the first custom model ever trained by ArmadaOS, serving as a proof-of-concept for the strategy of training specialized agent models from open-source base models.

Model Description

The Shadow agent is ArmadaOS's independent verification and quality assurance system. It operates as a multi-model red-team pipeline that stress-tests work before delivery, running independent assessor, cross-reference, and adversarial red-team roles.

Key Capabilities

• Independent Verification: Reviews and validates outputs from other agents
• Red-Team Analysis: Identifies weaknesses, gaps, and failure modes in plans and deliverables
• Cross-Reference Checking: Validates claims against multiple sources
• Quality Scoring: Provides structured assessment with confidence levels (high/medium/low)
• Engine-Native Tool Use: Uses ArmadaOS Engine tools ([tool: shell], [tool: read], [tool: memory_store], etc.)

Training Details

Parameter	Value
Base Model	Qwen 3.5 9B (Vision-Language)
Training Method	SFT + DPO (Supervised Fine-Tuning → Direct Preference Optimization)
Framework	Unsloth + TRL with bf16 LoRA
Trainable Parameters	0.31% (LoRA)
SFT Examples	226 trajectory examples
DPO Pairs	130 preference pairs
SFT Final Loss	1.6885
DPO Final Loss	0.4744
Training Time	~18 minutes on RTX A6000 48GB
Output Format	16-bit safetensors (4 shards, ~18GB)
Training Data Version	v4.1 Gold Standard (100% validated)

Training Data

Training data was generated using the AOSP-315 v4.1 Teacher Agent Directive, a comprehensive specification for producing Engine-native training examples. All 226 SFT examples and 130 DPO pairs were validated through multiple Shadow v1.2 verification rounds:

• 100% JSON valid
• 0% Manus contamination (pure ArmadaOS Engine-native format)
• Uses <think> blocks for reasoning
• ChatML wrapping via pipeline
• 70/15/15 split: identity reinforcement / no-prompt / mixed

Intended Use

This model is designed to run within the ArmadaOS Engine as the Shadow Verification agent. It is a proof-of-concept demonstrating that specialized agent-level models can be trained from open-source base models for specific organizational roles.

Limitations

• This is a v1.0 proof-of-concept; production deployment requires further evaluation
• Trained on a relatively small dataset (226 SFT + 130 DPO examples)
• Optimized specifically for the ArmadaOS Engine tool format and may not generalize to other agent frameworks
• Vision capabilities from the base model have not been specifically fine-tuned

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ArmadaOS/AOS-Shadow-v1.0")
tokenizer = AutoTokenizer.from_pretrained("ArmadaOS/AOS-Shadow-v1.0")

About ArmadaOS

ArmadaOS is an agent operating system that orchestrates specialized AI agents for organizational operations. Each agent has a distinct role (Chief of Staff, Shadow, Oracle, etc.) and operates within the ArmadaOS Engine using a defined set of tools, memory systems, and delegation protocols.

• GitHub: kam-ship-it/ArmadaOS
• Organization: ArmadaOS

Citation

@misc{armadaos-shadow-v1,
  title={AOS-Shadow-v1.0: ArmadaOS Shadow Verification Agent},
  author={ArmadaOS},
  year={2026},
  url={https://huggingface.co/ArmadaOS/AOS-Shadow-v1.0}
}