Available Guardrails: Certifying Selective Prediction Across ML Systems

Forum topic · 小凯 · 2026-09-22

Summary

This post introduces an arXiv paper (2609.22048) by Parivesh Priye, Yufeng Wang, Haibin Ling, and Michael Chaykowsky on certifying selective prediction systems. A selective predictor acts as a safety gate, returning outputs only when predictions are sufficiently trustworthy. The key challenge is not whether granted certificates are valid, but whether finite calibration data can produce them at all: as gates become safer or more fine-grained, some reporting units (tools, policy labels, patient subgroups) lack enough evidence to certify. The authors make this notion of availability computable via exact-binomial inversion and formulate reporting-partition selection as a dynamic program exposing trade-offs among safety, granularity, and served traffic. An informed planner achieves 0.157 higher average coverage than support-balanced selection, while a naive estimator recovers only 0.005, highlighting finite-sample recovery as the core challenge. Planning on one partition and selecting on another improves coverage by 0.060, reproduced in 59 of 60 model effects. Experiments span LLM tool calling, content moderation, lesion classification, and recommendation tasks.

Paper Overview

Field: Machine Learning Authors: Parivesh Priye, Yufeng Wang, Haibin Ling, Michael Chaykowsky Published: 2026-09-18 arXiv: 2609.22048

Abstract (English)

A selective predictor acts as a safety gate: it returns an output only when the prediction appears sufficiently trustworthy. Deployments increasingly require this reliability to be certified at a target precision for every reporting unit of interest, such as a tool, policy label, or patient subgroup. The main difficulty is often not whether a granted certificate is valid, but whether finite calibration data can produce one at all. As the gate becomes safer or more fine-grained, some units may receive too little evidence to certify.

The authors make this notion of availability computable through classical exact-binomial inversion and formulate reporting-partition selection, under a fixed group order, as a dynamic program that exposes the trade-off among safety, granularity, and served traffic.

Key Findings

The resulting frontier reveals substantial population opportunity almost erased by finite-sample estimation: an informed, truthful planner gains 0.157 higher average coverage than support-balanced selection, while a naive estimator recovers only 0.005 — making recovery from finite data the central challenge.
Building candidate partitions on one planning partition and selecting on another partially bridges this gap, improving average coverage by 0.060 over support-balanced selection, with the direction reproduced in 59 of 60 model effects across three intent-routing datasets and two architectures.
A complementary validity-preserving lever — redistributing family-wise error budgets across reporting units — recovers additional coverage under both population-scale and noisy estimation.
The same frontier recurs as a predictor-specific ceiling across LLM tool calling, content moderation, lesion classification, and recommendation scenarios.

Conclusion

Certified availability is a plannable deployment resource: it determines when safety gates can be certified, at what granularity, and how much traffic they serve.

--- *Auto-collected on 2026-09-22*