CAUSAL DECISION RUNTIME · LIVE PROBE
Don’t ask the model to sound certain.
Make uncertainty operational.
Compare competing hypotheses, choose the experiment that actually separates them, update belief, then decide. The page is rebuilt by GitHub Actions from the repository’s real deterministic examples.
LOW AIRFLOW / HVAC
Which observation is worth buying?
1 · Hypotheses2 · Candidate actions3 · Observation4 · Decision
BELIEF STATE
H1Clogged filter
50%H2Weak supply fan
50%Equal priors. The runtime has not observed a discriminating outcome.
CANDIDATE ACTIONS
The language model prefers temperature. The runtime overrides that score because the modeled outcomes of the pressure test differ sharply under H1 and H2.
DECISION TRACE
- World model initialized
- Waiting for an observation
- Posterior not updated
GENERATED DURING DEPLOYMENT
The webpage carries evidence from the Python runtime—not hand-entered claims.
Every publish runs the no-key Bayesian example and seeded HiddenWorld suite. Failures remain visible because calibration and regret matter more than a polished success screenshot.
Identification—30 seeded episodes
Mean posterior—true hypothesis
Causal regret—lower is better
Mean probes—before intervention
Open GitHub Actions runtime output
Loading…