osazuwa/2D_dungeon_flier_WHAM_dsep_hazard

2D Dungeon Flier WHAM — d-separation + hazard-based event timing

World model for the chest-room environment, trained on observational video only, with two supervision signals beyond next-frame prediction:

1. d-separation penalty — the 22 conditional independencies entailed by the chest-room DAG, applied as a JSD penalty on the model's own conditionals.
2. Discrete-event hazard loss — the symbolic trace is modelled as a controlled survival process: each of the 9 DAG variables has a per-frame hazard for its first realization, and realization is absorbing at rollout.

The hazard component is the difference from 2D_dungeon_flier_WHAM_dsep. Its predecessor generated episodes in which the terminal reward R frequently never realized; per-frame softmax over {0, 1, unrealized} gives a once-per-episode event no accumulating probability of occurring.

Status — partial run

Trained to step 5,500 of a planned 20,000 (28%). Stopped early for analysis. Metrics below are provisional.

arm	event (w_dsep=5.0, w_event=1.0, w_trace=0.0)
world-model params	170,960,168
effective batch	16 (4 x 4 accumulation)
hardware	1x A100 80GB, ~6.7 h

Result at step 5,500

Event timing is close to the data when an event fires:

var	model frame	real frame
V	13.1	14.0
Z	28.2	23.0
R	113.0	112.5

The open issue is that the post-chest-opening cascade (M, A, Y, R) does not always fire. See the linked analysis for the diagnosis.

Files

• wm_ckpt_event_step5500.pt — resumable (model + optimizer + scheduler + step)
• mini_wham_event_weights.pt — weights + config only
• provenance.json — source dataset/cache/tokenizer

Provenance

Trained on the contiguous observational prefix (clips 0–4999) of osazuwa/2d_dungeon_flier_video, reusing the frozen SigmaVAE+VQ tokenizer and cached latents from osazuwa/2D_dungeon_flier_WHAM. No interventional data was used.

Code: robertness/world_model_experiments, experiment 2D_chest_room_experiments/2026-08-13_mini_wham_event_timing.