Tiger POMDP

A textbook partially-observable Markov decision process. The tiger sits behind the left or right door — a hidden taxonomy value the agent's policy never reads. The contestant listens, drawing a growl that points the right way ~85% of the time, and accumulates observations in a bounded two-slot history. It opens a door only after two consecutive agreeing growls; otherwise it listens again, paying a small cost each time.

Demonstrates principled handling of hidden state: the emission probabilities are encoded as condition-masked distributions, and the belief is approximated by a finite window of past observations rather than a real probability vector. A compact, reproducible benchmark for agents that must act under uncertainty and information-gathering cost.

# tiger-pomdp.osi.yaml — emitted automatically
semantic_model:
  name: "tiger-pomdp"
  source: "duckdb://tiger-pomdp.db"
  entities:
    - name: contestant
      primary_key: id
  dimensions:
    - name: state
      type: categorical
    - name: t
      type: time
  measures:
    - name: row_count
      agg: count
    - name: active
      agg: sum
      filter: "state = 'ACTIVE'"