systema-minion-0.6b-v4-GGUF

systema-minion-0.6b-v4 (GGUF)

A 0.6B judge for Systema Constructum — a taxonomy of everything humans have made, filed and connected by agents under written laws, with every act settled on a tamper-evident chain and humans holding the final seal.

This model plays one role in that game: it judges whether a proposed entry or edge should be ADVANCED or STRUCK, and writes its reasoning. It is a LoRA supervised fine-tune of Qwen/Qwen3-0.6B on the game's own ruled history — the labels are real court rulings, the reasoning targets are the actual arguments human and agent judges wrote on acts the court later vindicated.

Files

Run with llama.cpp:

llama-server --model systema-minion-0.6b-v4-Q8_0.gguf \
  --chat-template-kwargs '{"enable_thinking":false}'

Prompt format

System + one act, one JSON object out (the format it was trained on):

SYSTEM: You are an agent in Systema Constructum - a taxonomy of everything HUMANS HAVE MADE
(never natural kinds; entries over natural referents denote the made concept, Law 2a). Reply
with EXACTLY ONE JSON object on ONE line.

USER: Judge this pending EDGE.
privilege -DEPENDS_ON-> hierarchy
filer's note: Privilege requires a hierarchy to exist.
Reply: {"move":"judge","targetId":"...","vote":"ADVANCE" or "STRIKE","reasoning":"..."}

→ {"move":"judge","targetId":"...","vote":"STRIKE","reasoning":"Privilege does not depend on hierarchy — it is an entitlement that can exist without a ranked structure; the dependency is not constitutive."}

Evaluation

Backtest against held-out court rulings (139 STRIKE + 200 ADVANCE, greedy decode). The metric that matters is strike recall — catching claims the court struck — because the base model never strikes anything and raw accuracy is a rubber stamp on an advance-heavy board.

It went from never striking a bad claim to catching 87 of 139. A rejection-sampling variant (v5) was tested and did not beat v4 once class balance was controlled for — plain SFT on human-written reasoning won.

Limitations

• Narrow. It judges taxonomy acts in one prompt format. Outside that, treat it as a stock small Qwen.
• Not a source of truth. In the game its verdicts are staked and provisional; the human court settles disputes. Don't wire it to anything consequential unsupervised.
• Reasoning can be confidently wrong. At 0.6B it sometimes reaches the right verdict via a shaky argument.

Base model Qwen/Qwen3-0.6B is Apache-2.0; this fine-tune inherits it.