TigerPOMDP · replay

event-log playback ·

Contestant

—

Hidden state

Belief tiger-left

50%

Listens

Reward

Step

Episode: #2 — clean win #34 — the growls lied #51 — long deliberation

Two doors — listen, believe, open

◀ tiger LEFTbelief b = P(tiger-left | growls)tiger RIGHT ▶

ring (history: 2): — —

listening…

Return distribution — partial observability is bimodal

treasure +10 tiger −100 growl / belief left growl / belief right space play · ←→ step

Replays , distilled by tools/gen_viz_replay.py — each contestant's growls, running reward and opened door are reconstructed from logged update / transition rows. The belief bar is the textbook Bayesian update (p = 0.85) computed from those growls; the policy itself never sees it — it opens after two agreeing growls in the obs ring.