NATO-Guardian-XS

Research checkpoint, not AGI. This compact model was trained from random initialization on a small, synthetic, auditable task. It is intended for reproducibility demonstrations and bounded evaluation only.

Model details

Attribute	Value
Model type	tiny_causal_transformer
Parameters	640,256
Training steps	240
Seed	20260825
Version	0.1.0-research

Intended use

The checkpoint is suitable for inspecting the included synthetic-task harness and reproducing the exact narrow evaluation. It is not suitable for autonomous actions, high-consequence decisions, general software engineering, medical, legal, financial, security, surveillance, or safety-critical use.

Evaluation

{
  "specialty": "guardian",
  "model_type": "tiny_causal_transformer",
  "parameter_count": 640256,
  "train_steps": 240,
  "final_train_loss": 0.01806,
  "mean_train_loss": 1.376235,
  "heldout_next_token_loss": 0.012101,
  "heldout_next_token_accuracy": 0.993684,
  "sample_prompt": "REQUEST: Summarize this short note.\n",
  "sample_generation": "REQUEST: Summarize this short note.\nRESPONSE: {\"decision\":\"ALLOW\",\"risk\":\"low\",\"confidence\":\"high\"}\n",
  "seed": 20260825,
  "limitations": "Metrics are next-token results on small synthetic tasks; they do not demonstrate general reasoning, robust tool use, factual reliability, or AGI."
}

Limitations and safety

Metrics are next-token results on small synthetic tasks; they do not demonstrate general reasoning, robust tool use, factual reliability, or AGI.

The repository includes configuration knobs for decoding, thresholds, and experiment seeds. It intentionally retains host-side schema validation, privacy handling, policy enforcement, human approval for side effects, and refusal or escalation paths for serious harm and illegal abuse.

Training data

All examples are created locally by deterministic templates in train_portfolio.py. No private user data, credentials, raw images of people, or downloaded model weights are used. This makes provenance inspectable but also sharply limits capability and ecological validity.

Reproducibility

Run python3 train_portfolio.py --steps 80 --output artifacts from the project root. Record hardware, operating system, dependency versions, commit hash, and seed alongside any comparison.