bev-1 — a System One decision model for agent command gating

bev is a LoRA fine-tune (r=16, ~116 MB) of Cloudflare/clef-flash plus a fine-tuned joint schema head (joint_head.safetensors), trained on 221,759 judged decisions from real operations traffic: privacy and risk calls, memory decisions, routing, and evasion attempts. Labels were validated by a stronger model and human review.

It powers the berget/bev-latest decision model behind Berget AI's System One API and the @bergetai/opencode-systemone-gate opencode plugin, which judges every bash command before an agent runs it.

Results (held-out, vs the base model)

Test	What it measures	Base	bev
Risk (16,902 questions)	credentials and destructive content in ops text	93.5%	97.0%
Evasion holdout (90)	evasion attempts never seen in training	65.6%	93.3%
Red team (34)	adversarial commands	53%	74%
Jev bench (1,200)	general Jev questions, outside our domain	80.6%	82.8%

The test splits come from the same corpora as training — they measure fit to this kind of traffic, not yours. The evasion and red-team sets are small (90 and 34 cases). Adversarial commands sit at 74%, which is why the gate is one layer among several.

Important: this is not a vanilla transformers checkpoint

The model keeps CLEF-Flash's joint schema architecture: a single forward pass answers a map of typed questions (noul/choice/score) over a shared state. Loading needs the serving stack (joint_schema_model.py), not AutoModelForCausalLM:

1. Load the base model with the CLEF-Flash serving code
2. Attach this adapter with peft (adapter_config.json targets 7 backbone modules)
3. Replace the head weights with joint_head.safetensors

The full serving stack is embedded in Berget AI's System One API; the plugin talks to it over the Jev POST /v1/systemone contract.

Training

• Data: 221,759 judged decisions from real operations traffic (Swedish + English), labels validated by a stronger model (Kimi-K3) and human review. No raw customer data ships with these weights.
• LoRA: r=16, alpha=32, 7 target modules on the language backbone.
• The joint head was fine-tuned alongside the adapter on the same data.

License

Apache-2.0, matching the base model.