nanoAWM

nanoAWM trains compact, action-conditioned world models in a deterministic symbolic OS. The credible claim is bounded: it is local MiniOS evidence for consequence simulation, not external validation of production agent safety.

• Action-conditioned consequence prediction
• Held-out and genuinely held-out scorecards
• Negative action-paraphrase and cross-surface OOD results preserved
• Symbolic MiniOS scope stated in the README