Not a bigger model — the first world model OS that can say "I don't know"

01

Default output format

Every inference must emit five fields. Missing any one is grounds for refusal.

• 01Conclusion — the system's judgement on the current question
• 02Confidence — a calibrated probability or interval, not a smoothed tone
• 03Evidence chain — source and version pointer for every claim
• 04Boundary verdict — does the action sit inside permission / environment / safety / task constraints
• 05Next action — proceed / refuse / escalate

02

Moat

Boundary DSL + audit schema + bicameral inference. The trio is decoupled from the base model — portable across Llama / Qwen / Mistral / in-house bases. The scale camp's next release is, to us, an adapter retrain.

03

Anti-thesis pairing

"Bigger means more accurate" is a correlational promise whose failure mode is silent hallucination. "We know where we are not accurate" is a mechanistic promise whose failure mode is explicit refusal / abstention / escalation — and the latter is auditable, rollbackable, reproducible.