Dispatch ESSAY June 18, 2026 1 MIN READ BY PRAVEEN LAVU

A converged router is a strong opinion about a world that moved

My router spent a month learning which model to trust. Then one of them got silently updated, its accuracy fell off a cliff, and it kept sending work to the dead arm because three weeks ago that was the right call.

Converged routing policies become stale when the underlying system changes, causing models to confidently route to degraded endpoints based on outdated learned preferences. Drift detection should run continuously in production to catch when past performance no longer predicts present outcomes, not after requests fail.

Key takeaways

A converged policy is a fixed opinion about reality; when endpoints update, degrade, or the input distribution shifts, the policy can become confidently wrong without any internal signal of failure.
Shorter memory across the system is not the solution because it causes estimates to fluctuate at noise; instead, implement per-arm drift detection that triggers resets only when the learned relationship breaks.
Drift detection should be integrated before requests reach users, using established techniques like ADWIN to continuously validate whether historical outcomes still predict current performance.

It was not broken. It was right about a world that no longer existed.

That is the trap with anything that learns from outcomes. We treat convergence as the finish line. The loss curve flattens, the policy stabilizes, the weights stop moving, and we exhale. But a converged policy is just a strong opinion about a fixed reality, and production has no fixed reality. Models get updated. Endpoints degrade. The prompt mix drifts. A confident estimate built on two thousand good calls does not flinch when the truth drops underneath it.

The fix is not a shorter memory everywhere, because that just makes every estimate twitch at noise. It is a tripwire per arm: one quiet check after every outcome, asking whether the past still predicts the present, ready to forget on purpose the moment it stops. The technique is old (ADWIN, 2007) and the wiring is small. The mindset shift is the hard part.

So stop treating convergence as done. A system that learned is a system that can be confidently, durably wrong the day the world it learned stops being true. Put the tripwire in the loop before the page arrives, not after.

Full writeup: Drift Detection for LLM Routing

Get the next dispatch when it drops.

SUBSCRIBE FOR THE NEXT DROP MORE ARTICLES