MC Neuro-Symbolic Ledger World Model (research artifacts)

Files

• bidirectional_ft15000.pt — main bidirectional model (kb7: WASD/jump/attack/place, ledger-held injector). 15k steps: leak-immune contrastive margin, hard negatives, high-noise t-band, spatial localization.
• ar_stage1_1000.pt — ForgeWM/Causal-Forcing stage1 causalization (1k steps, feasibility: dense actions transfer; sparse place does not without dedicated supervision).
• probe_mc44.pt — frozen latent event probe (mine/pickup/use heads). NOTE: detects action events, not outcomes — calibrate against human grading (see eval_assets).
• eval_assets/ — counterfactual eval videos/strips/logs: FIRST_ONLY protocol, zero-prelude tests, CFG dose-response (w −2.5…+2.5, incl. "negative place removes blocks"), ledger demo passes, GT strips.

Honest numbers (human-graded)

• Event-level counterfactual causality at press instant: ON responds / OFF silent (12/12 windows, probe + eyes).
• Persistent visible placement (with action-CFG w=2.5): 3/12 strict; item type matches held (cobble→stone, dirt→dirt).
• Known failure modes: transient form, zero-action-tail drift, scene dependence, first-frame anchor dead zone (lat 0–3).