acceptodds
Under review as a conference paper at ICLR 2027

Available Guardrails: Certifying Selective Prediction across ML Systems

Abstract

A selective predictor can be statistically valid yet practically unavailable: with finite calibration data, it may fail to certify enough traffic at the reporting granularity required for deployment. We formalize certificate availability, the probability that a frozen candidate receives a valid certificate, and use exact binomial power as a planning quantity for certified service. The central empirical result is not that one partitioning algorithm dominates. The reliability order used by our contiguous dynamic program beats the median random order only of the time, while adding structured candidates to a diverse random candidate pool contributes only mean coverage with median zero. Independent held-out comparison among diverse frozen candidates is more useful, but its gain over mean random-order balancing is still only with median zero. These findings turn the partition optimizer into a diagnostic of population headroom and redirect the deployable story toward two stronger levers. Under fixed semantic boundaries on HWU, direct constrained planning raises certified coverage at from to for DeBERTa and from to for DistilRoBERTa relative to support balancing. Separately, validity-preserving allocation of the familywise testing budget gains mean coverage from noisy planning data, increasing to with posterior-predictive availability. Population analysis confirms that substantially more certified traffic can exist in principle, while direct plug-in planning recovers only , identifying finite estimation rather than optimization as the bottleneck. Across four additional domains, the same safety, granularity, and traffic constraint recurs, although no single planner dominates. The contribution is therefore a deployment formulation and empirical diagnosis of certified availability, together with practical mechanisms for recovering certified service under finite evidence.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Safety, Risk-Control, Precision-Recall Tradeoff

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_yDxXOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

As I understand it, the paper is focusing on maximizing the fraction of model predictions that can be certified. It is focusing on the case where several subgroups of data are certified separately, but there might not be sufficient data to certify specific subgroups at a requested error level. It is therefore trying to predict whether subgroups at a specific granularity can be certified with a finite set of data to optimize the overall amount of predictions that can be certified.

Critical strengths

As I understand it, I really like the problem setting of explicitly trying to maximize the expected fraction of deployment traffic, and I think setting up and defining this problem is valuable in itself. The results are also described in a very nuanced way they include some negative results.

Critical weaknesses

===Readability=== I currently find the paper very difficult to follow, which makes it hard to properly judge the actual technical contributions. For example, the abstract in particular (but also the rest of the main text) contains a lot of unexplained jargon and evaluation metrics that are sometimes hard to put into context. The datasets used for evaluation are never motivated or explained in the main text. Overall, I think the paper would benefit from focusing more on readability, intuition-building, and exposition.

===References=== There seem to be a lot of references missing. The evaluation datasets are never referenced in the main text. Neither are used models such as ResNet or Qwen.

Questions for authors

None

Reviewer_NoNrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This is an honest paper. It starts from the notion of a "frozen candidate" (one reporting unit + one threshold) that serves n examples and incurs E errors. Suppose we certify the candidate's expected performance via testing. The authors formalize the probability that a frozen candidate receives a valid certificate and make it computable via an exact technique (Clopper–Pearson binomial inversion), turning classical test power into a planning quantity. They formulate reporting-partition test design (under a fixed order) as a dynamic program that exposes the safety–granularity–traffic frontier. Once a certificate is returned, it is a valid finite-sample guarantee of the claimed selective precision, exactly as formalized. In terms of complexity of the optimization, population optimization reveals substantial headroom (mean +0.157 coverage over support balancing), but direct plug-in recovers almost none of it (+0.005). The authors admit it is finite estimation—not optimization—that is the bottleneck. The interesting result is the negative one: a selective predictor can be statistically possible yet practically unavailable; finite calibration data may fail to certify enough traffic at the required reporting granularity.

Critical strengths

Solid notion: The certificate notion is sound: it is an exact finite-sample binomial test whose family-wise validity follows from fixed-sequence + union-bound arguments.

Clean problem formulation: The paper cleanly separates validity (“is the returned certificate trustworthy?”) from availability (“will finite data return a certificate at the required granularity?”).

Interesting ideas: The exact binomial identity, the dynamic program for contiguous partitions, the honest negative results on reliability ordering, and the identification of two stronger levers (constrained semantic planning and family-wise budget reallocation) are valuable.

Honest experimentation: Experiments are extensive and well designed. Results are inconclusive in terms of the practical applicability of the approach, but the limitations are stated clearly.

Critical weaknesses

Statistical and approximation issues: The planning objective replaces random binomial support with its mean. Appendix I shows this can be non-conservative (e.g., projected score 0.500 vs exact expected coverage 0.276). Exact integration helps only marginally in practice.

Restricted search. The dynamic program only considers contiguous merges under a frozen reliability order. Non-contiguous partitions are excluded by construction. The estimated reliability order itself is not very strong: it beats the median of random orders only 33% of the time and has lower mean coverage.

Severe estimation bottleneck. Population-oracle planning gains ~0.157 mean coverage over support balancing. Direct plug-in recovers only ~0.005 (median zero; can be largely negative when groups are thin). The gap is driven by noisy estimates of traffic shares, error rates, and the reliability order, all likely to occur in practice.

Limited value of structure. Adding the structured candidates (plug-in, support-balanced, equal-count) to a diverse random-order pool contributes only ~0.0027 mean coverage with median zero. Most of the recoverable gain comes from candidate diversity + independent held-out selection, not from the reliability-ordered dynamic program.

The assumption treats reliability as an ordinal value rather than a random variable with its own probability distribution.

Except for the latter one, these limitations are well acknowledged in the paper’s own conclusion and limitations section. The work is strong as a diagnostic framework and as an identification of the safety–granularity–traffic frontier; it is weak as a deployable optimizer that reliably recovers the population opportunity from finite data.

Questions for authors

(i) Your availability formula treats the true selective error r s a fixed (unknown) constant. In practice, the conditional error rate given the unit and threshold is often stochastic—because of residual dependence, unmodeled subgroup heterogeneity, or mild non-stationarity. Can a deployer propagate uncertainty about r (e.g., via a posterior or a worst-case interval) into the availability calculation so that the planned certified traffic remains a conservative or calibrated predictor of runtime performance? (ii) Again, when r is random, the certificate-availability probability becomes a random variable as well. In a production setting where a safety gate must remain online with high probability, do you recommend planning against expected availability, a high-probability lower bound, or an adaptive scheme that reallocates the familywise budget? What finite-sample guarantees survive in that regime?

Reviewer_CnJqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper starts with the motivation that reliability might need to be certified separately for different groups, e.g., tools, patients, and so on. This will allow for detecting local failures than having a global certification procedure. However, doing the certification per group will come at the cost of having to divide calibration data into disjoint subgroups, which reduces the number of available data points per group and by itself might reduce power.

Consequently, this paper provides a pre-certification procedure that uses statistical power calculations to determine how likely it is for the certification procedure to be successful. Additionally, by assuming a value for the true error rate, and given a desired error threshold and outage rate, it determines how much calibration data we would need per group. The main objective they optimise for is the coverage of groups for which we can provide certificate. To achieve this, the paper explores different methods such as choosing the grouping, merging the groups, selecting the confidence thresholds, and dividing the FWER budgets.

Critical strengths

1- I like the direction of distinguishing between obtaining a valid certificate versus exploring the probability of obtaining one as a 'pre-certificate' procedure. For instance, a safe predictor can have very little certified service when calibration data is limited or if the observed error is close to the desired threshold. Planning ahead for this before spending the calibration data budget seems like a good motivation.

2- I appreciate the authors' honest reporting of the negative results, for instance that their proposed reliability order actually beats the random order in 1/3 of the times, or that adding the structured candidate to a random pool adds a very small amount to the coverage.

Critical weaknesses

1- While I understand the motivation behind the paper, and the theory, the practical usefulness of the proposed method still remains quite unclear to me. For instance, I am thinking of when an application would allow for the reporting groups to be redesigned in the way described in the paper. If we allow for merging groups, the original problem that motivated the paper appears again: within each merged groups, there might still be subgroups that do not pass the certificate. Additionally, another thing I do not see much practical use for is when the paper fixes the number of reporting units and changes the grouping. If the original grouping was done with a particular intention in mind, e.g., based on patient age groups, then the new groupings with the same number of members does not mean anything in terms of the original grouping.

2- The novelty of the paper is quite limited. Most of the machinery used is a combination of well-known tools in statistical testing, such as exact binomial power or allocating the FWER budget to increase power. This would mean that we would need to see a significant improvement in the experiments, which we mostly are not according to the authors. The negative results are useful, but limit the scope of the publication.

Questions for authors

1- In order for me to see if there are more practical use cases for the proposed method, would you be able to provide me with a real application scenario in which having separate group guarantees is required, but the group memberships can be changed as proposed in the paper?

2- The main objective optimises for total certified coverage. This can therefore favour providing certificates for larger groups and leaving smaller ones uncertified. Given the main motivation of the paper, will this be an acceptable outcome?

3- The availability calculation in (9) depends on having the true error rate, and the objectives in (10) and (11) depend on the unknown group specific errors. However, in practice, these must be estimated from data. How can a practitioner address this in practice, and how much will it affect the proposed method?

Reviewer_Xb6FOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies when a selective predictor can obtain enough statistical certificates to serve useful amounts of traffic. It distinguishes the validity of a certificate from the probability of obtaining one with limited calibration data. The authors use exact binomial power to plan reporting groups and thresholds while keeping certification data separate. Experiments show that a planner using true population quantities has substantial potential, but estimated reliability orders and structured candidates provide little extra benefit over diverse random partitions. The paper then studies independent candidate selection, planning within fixed semantic boundaries, and unequal allocation of the testing budget. Studies cover intent classification, controlled shifts, and four additional application domains. Constrained planning gives larger gains on HWU, while budget allocation gives smaller gains in simulations. The main contribution is a deployment framework and an empirical diagnosis, rather than a universally better planner.

Critical strengths
  1. The paper highlights a useful distinction between a valid certificate and the ability to obtain one with limited data. Treating certification probability as a planning objective connects safety requirements, reporting granularity, and service coverage. This gives practitioners a concrete way to assess whether their certification requirements are feasible.

  2. The paper identifies two useful ways to improve certified coverage. Constrained planning gives substantial gains on HWU while preserving declared scenario boundaries. Unequal testing-budget allocation also improves coverage with noisy planning data, and accounting for estimation uncertainty improves it further. Both operate within the paper’s validity guarantee.

  3. The experiments distinguish population opportunity from what finite planning data can recover. Comparisons with random-order designs and matched candidate pools show that candidate diversity and independent selection matter more than the proposed reliability ordering. These findings are useful even where the structured planner does not win, because they identify estimation and selection as practical bottlenecks.

Critical weaknesses
  1. The scenario-preserving HWU results are a meaningful strength. However, the direct planner changes both the number of reporting units assigned to each scenario and the grouping within each scenario, while the baselines use simpler allocation rules. As the authors acknowledge, these effects are not separated. This leaves the source of the gains unclear: better allocation across scenarios, better grouping within scenarios, or both.

  2. Both constrained partitioning and testing-budget allocation show useful gains, but their practical evidence has different limits. The HWU gains appear across architectures and trained models, yet come from a post hoc replay of inspected traces. Budget allocation is tested with noisy planning data, including a posterior-predictive version, but these gains come from controlled synthetic studies. Independent real-data evaluation would strengthen the case for both mechanisms.

  3. The paper establishes valid certification under its chosen procedure and shows gains from unequal testing-budget allocation. It also acknowledges that sharper familywise procedures could improve availability. What remains uncertain is how the coverage frontier and the benefits of both planning mechanisms would change under another valid, less conservative procedure. This limits how broadly the reported effect sizes can be interpreted.

Questions for authors
  1. Could you compare the HWU grouping methods while fixing the number of reporting units per scenario? A second comparison that fixes the grouping rule and varies only the allocation would help explain which part drives the gains.

  2. Could you fix the planning procedures in advance and evaluate constrained partitioning and testing-budget allocation, separately and together, on untouched real data? Using the same total label budget would help show whether the two gains carry over and whether they complement each other.

  3. Beyond the unequal allocations already studied, could you test a sharper valid familywise procedure under the same label budget and reporting constraints? How much of the constrained-planning and budget-allocation gains remains relative to matched baselines?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Regarding no engagement

Dear Area Chair,

Thank you for your consistent efforts to elicit responses from reviewers to our replies. Despite your reminder and our response to most reviews more than a week ago, we have not received much engagement from a few reviewers. This is discouraging and does not align with the spirit of the rebuttal period. It is particularly concerning for authors when months of work are assessed with a low rating and no active subsequent engagement with our response.

We kindly ask that you take this into consideration when evaluating our work. The lack of follow-up has left us uncertain about the concerns underlying the rating and without a meaningful opportunity to address them.

Thank you again for your efforts to encourage a constructive discussion between authors and reviewers, and for your gentle reminders.

Sincerely, The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.