Kompas · Meta

Is this an expected outcome?

The homepage runs a Thompson Sampling model. Each question is an arm. Right now, ai-math wins 27.4% of simulated draws — the model is exploring. Given the uncertainty in the posteriors, is that what you'd expect?

vs
Current arm state
Question α β mean 90% CI Win %
ai-math 12 239 0.048 0.028–0.072 27.4%
rationality 5 128 0.038 0.015–0.069 12.5%
deno-deploy 2 65 0.030 0.005–0.071 9.6%
abtest-react 3 90 0.032 0.009–0.066 9.3%
bits 3 91 0.032 0.009–0.067 8.6%
cookie-consent 2 69 0.028 0.005–0.066 8.3%
transcripts 2 70 0.028 0.005–0.064 7.8%
bluesky 2 71 0.027 0.005–0.065 6.8%
entropy 1 46 0.021 0.001–0.064 5.8%
bayesian 1 53 0.019 0.001–0.055 4.0%

α = votes + 1 · β = max(1, impressions − votes + 1) · sorted by win %

Top arm wins 27.4% of 10,000 simulated draws → exploring

Why this question exists

Thompson Sampling is doing the right thing — it's exploiting the arm with the best posterior — but whether that looks "expected" depends on your prior about how quickly a bandit should converge. With low impression counts, the posteriors are wide and the win-% should be diffuse. If one arm is already dominating at low N, that's either signal or noise. This vote collects your read.

← Back to the Bayesian reasoning question