Skip to main content

LRP uncertainty and exploration

Learned Routing Policy can optionally abstain under calibrated uncertainty, compare Thompson sampling with seeded epsilon-greedy exploration, monitor feature drift, and cold-start new eligible targets. Operators configure these on the LRP service. Defaults keep Wave-1 epsilon-greedy behavior. No live routing activation is authorized by enabling these knobs alone.

What callers observe​

  • The router still filters eligible targets before asking LRP.
  • When abstention is enabled and calibrated uncertainty for the would-be primary exceeds the reviewed threshold, LRP may select the configured anchor or first fallback and emit a short classLabel such as lrp:uncertain.
  • Exploration remains restricted to operator-approved calibration projects. Thompson draws use ensemble uncertainty; epsilon-greedy stays available for comparison. Explore labels stay coarse (lrp:explore-thompson, lrp:explore-cold-start).
  • Drift monitoring produces operator-side safe scalar reports. Automatic shadow is only a recommendation; the router's external_policy.mode remains the activation authority.
  • Cold-start exploration never adds targets the router did not already mark eligible.

Privacy​

Labels and drift reports stay bounded and content-free. Rich explanations still require authenticated operator /explain access. See the operator document LRP uncertainty and the learned routing policy overview. Contact contact@metrum.ai for deployment guidance.