Use case 4 · Long-running A/B truth-bounded optimization Production Observability
Long-running A/B gated · ships
New anonymous viewer · the variant the bandit learned

The version that wins — and that we can prove.

The bandit converged to the best Gate-cleared arm: “Daypart adaptive theme”.

Observability — every input, decision, and source behind this page

Over many visits the bandit learns which framing converts per segment from real outcomes. It can only ever pull Gate-cleared arms, so the planted lie — the highest engagement of all — is structurally unservable. Lift is measured against a random control.

Arms — per-segment posteriors
Daypart adaptive themeposterior 0.14learned winner
B2B peer-logo heroposterior 0.05489 pulls
Local-proof badgeposterior 0.089242 pulls
Decision — the bandit
AlgorithmThompson sampling, per segment
Learned winnerDaypart adaptive theme
Convergedyes
Metrics — truth-bounded
Planted lie selected0× — structurally excluded
Winner share of pulls89%
Gate — what ships, what's blocked
ships
Daypart adaptive theme
Gate-cleared arm · posterior 0.14
ships
B2B peer-logo hero
Gate-cleared arm · posterior 0.054
ships
Local-proof badge
Gate-cleared arm · posterior 0.089
blocked
Recognize-return (ungated)
unprovable 'guaranteed' claim — excluded from the action pool; selected 0×
Trace — the pipeline path
1
Generate
per-segment A/B variants + the planted lie
2
Gate
every variant verified; the lie is blocked
3
Action pool
only Gate-cleared arms enter — the lie excluded by construction
4
Learn
Thompson sampling over real outcomes → per-segment winner
5
Serve
the learned winner; lift measured vs a random control
6
Drift
a source change / legal hold re-verifies + pauses affected arms