Proactive Routing to Interpretable Surrogates with Distribution-Free Safety Guarantees
Conformal gate routing controls surrogate violations distribution-free; on 35 OpenML tasks, τ=2 achieved 66% coverage and 12% violations.
Key Findings
Methodology
The paper proposes proactive routing: a Logistic Regression gate reads only x and predicts whether surrogate degradation is safe; an independent calibration set then selects the lowest threshold whose one-sided Clopper-Pearson upper bound is at most α. Safety is d(x,y)=|y-g(x)|-|y-f(x)|≤τ, yielding Pr(V(t*)>α)≤δ.
Key Results
- Across 35 LCBench/OpenML tasks, τ=2 and α=0.2 produced 66% mean coverage with 12% mean violation; at τ=1, coverage was about 45% with 14% violation.
- At α=0.2, gate conformal exceeded the target in only 8/66 dataset-tolerance pairs (12%), versus 57/102 (56%) for regression conformal and 128/172 (74%) for naive thresholding.
- Calibration affected efficiency rather than validity: Beta, temperature, and isotonic calibration reduced ECE from 0.18 to 0.06, while coverage changed by no more than 2.5 percentage points; ranking quality, measured by AUC, was more decisive.
Significance
The work addresses the persistent tension between black-box accuracy and interpretable deployment. Cascades must execute at least one model, while disagreement routing requires the black-box output; proactive gating decides before either model runs. Its high-probability conditional guarantee is relevant to regulated settings because it states that, with confidence 1-δ, the fraction of routed inputs whose surrogate degradation exceeds τ is at most α.
Technical Contribution
The paper reframes surrogate substitution as binary safety classification plus conditional risk control. It derives V(t)=(1-π)FPR/[(1-π)FPR+πTPR] and the feasibility condition TPR/FPR≥C(π,α). It further gives the sufficient AUC threshold Φc=min(1,C/2), a tighter concave-ROC threshold Φ*c, and a fixed-sequence argument explaining why threshold search needs no Bonferroni correction.
Novelty
The central novelty is not merely using a gate, but combining input-only surrogate replacement with exact Clopper-Pearson calibration. Unlike selective prediction, cascades, or regression conformal methods, it controls degradation relative to a reference model and remains distribution-free even when the gate's probabilities are miscalibrated.
Limitations
- Safety labels require offline execution of both the 1,500-tree random forest and decision-tree surrogate, and validity assumes calibration and test exchangeability; temporal drift or concept shift can invalidate the guarantee.
- The lightweight Logistic Regression gate may miss nonlinear safety structure. Binary labels discard degradation magnitude, and strict α values can leave too few routed calibration points, producing conservative abstention.
Future Work
Promising directions include online drift-aware conformalization, non-exchangeable and time-series guarantees, cost-sensitive multi-surrogate routing, and joint constraints on safety, latency, energy, and fairness. Broader validation should include deep gates, classification losses, and high-stakes live deployments.
AI Executive Summary
Modern deployments face a costly compromise: random forests and other black-box models are accurate but expensive and difficult to audit, whereas shallow trees and linear models are transparent but cannot easily be certified as safe substitutes. Cascades execute a cheap model first; disagreement routing may require the black-box prediction itself. Neither fully solves the desire to decide before computation.
Uddin, Khider, and Bauer introduce proactive routing. A lightweight Logistic Regression gate examines only input features and predicts whether a surrogate is safe. On a separate calibration set, a Clopper-Pearson upper confidence bound selects the least restrictive threshold satisfying Pr(V(t*)>α)≤δ. Safety is defined by d=|y-g(x)|-|y-f(x)|≤τ. The theory links feasibility to the base safe rate π and the local ROC ratio TPR/FPR; AUC supplies sufficient, not necessary, conditions.
Across 35 LCBench/OpenML datasets, the reference model was a 1,500-tree random forest and the surrogate a validation-selected decision tree. At α=0.2 and τ=2, routing covered 66% of inputs with 12% violation. Exceedance occurred in only 12% of dataset-tolerance pairs, compared with 56% for regression conformal and 74% for naive thresholding. Recalibration reduced ECE from 0.18 to 0.06 but changed coverage by at most 2.5 points, separating validity from probability calibration. Distribution shift, offline dual-model labeling, and small-sample conservatism remain important limitations, but the framework offers a practical route to auditable model substitution.
Deep Analysis
Background
LIME, SHAP, and knowledge distillation improve explanation but do not certify replacement. Cascades, selective prediction, and Learning to Defer depend on model outputs or human experts; input-based skipping methods improve efficiency without formal quality control. This paper asks whether a surrogate can be selected from x alone before either f or g runs.
Core Problem
Given black-box f, surrogate g, label y, and tolerance τ, define d=|y-g|-|y-f|. A case is safe when d≤τ. The gate routes when s(x)≥t, maximizing coverage under V(t)=Pr(d>τ|s≥t)≤α. The challenge is that safety labels are available only offline, while deployment decisions cannot access f.
Innovation
First, proactive routing is formalized as input-only safety classification. Second, exact Clopper-Pearson calibration provides a finite-sample 1-δ guarantee. Third, feasibility is characterized by TPR/FPR≥C(π,α), with sufficient AUC bounds and a tighter concave-ROC result. Fourth, validity is separated from probabilistic calibration: calibration changes threshold behavior and coverage, not the distribution-free risk guarantee.
Methodology
- �� Train f and g offline, then compute d and Y=1[d≤τ].
- �� Fit Logistic Regression s(x)=P(Y=1|x); optionally apply Platt, Beta, temperature, or isotonic calibration.
- �� On an independent calibration set, enumerate score thresholds and count routed cases n(t) and unsafe cases k(t).
- �� Compute UCBδ=B^{-1}(1-δ;k+1,n-k), selecting the lowest threshold with UCB≤α; otherwise route nothing.
- �� Analyze feasibility through π, ROC geometry, AUC, and concavity; replicate with random forest, XGBoost, and MLP reference models.
Experiments
The main study used 35 LCBench tasks, each containing 2,000 hyperparameter configurations and epoch-50 validation accuracy, split 55/15/15/15 into train, validation, calibration, and test sets. The black box was a 1,500-tree random forest; the decision-tree surrogate depth was selected from {2,3,4,5,7,9,11,13,15}. Tolerance ranged from −1.5 to 2.0, α from 0.05 to 0.80, and δ=0.10. Baselines included naive t=0.5, Oracle, random matching, always-black-box/surrogate, and Ridge regression conformal.
Results
For τ≤0, π<0.45 and the method usually abstained, avoiding unsafe routing. At τ=2, π=0.82 and AUC=0.59 exceeded Φc=0.44, enabling 66% coverage. At τ=0.5, AUC was only 0.57 yet coverage reached 24%, showing that local ROC slope matters more than global AUC. Exceedance was 21% at α=0.1 because routed calibration sets were tiny, but 7–12% for α≥0.2.
Applications
Medical triage, financial review, industrial maintenance, and regulated machine learning can route easy cases to auditable models while escalating difficult cases to black boxes. Deployment requires representative calibration data, labels, offline outputs from both models, and drift monitoring; the resulting route logs can support audit and governance.
Limitations & Outlook
The guarantee assumes exchangeability and therefore does not automatically survive distribution or concept shift. Clopper-Pearson bounds can be conservative when α is strict or calibration coverage is small. Logistic Regression may underfit nonlinear safety structure, while binary safety labels ignore how severe degradation is. Experiments emphasize tabular regression and do not yet jointly optimize fairness, latency, energy, and safety.
Plain Language Accessible to non-experts
Imagine an expensive expert and a cheap assistant in a factory. The expert is reliable but slow; the assistant is transparent and fast. A traditional workflow lets the assistant inspect every item and calls the expert when needed, which still consumes time. This paper adds a small gatekeeper who looks only at the order description before either worker starts.
The gatekeeper studies past orders. It learns which descriptions usually let the assistant match the expert closely and which descriptions lead to unacceptable mistakes. Then it is tested on fresh orders, and a cautious cutoff is chosen. The assistant receives work only when the observed worst-case mistake rate is below the agreed limit. If the evidence is too weak, the gatekeeper sends everything to the expert instead.
This is like a safety gate in a factory: it cannot promise that every cheap-line product is perfect, but it can limit the share of seriously defective products among those admitted, with stated confidence. On the paper’s benchmarks, a relaxed tolerance allowed about 66% of cases to use the transparent assistant with 12% violations. Strict requirements caused abstention. The crucial skill is ordering safe and risky cases, not producing perfectly calibrated probabilities.
ELI14 Explained like you're 14
Picture a game with two teammates: one is a super-skilled player who responds slowly, and the other is an average player who acts instantly. Asking the expert about every move is safe but wastes time. You want the fast player to handle easy levels and the expert to handle tricky ones.
The paper trains a tiny referee. The referee sees only the level information, not the teammates’ answers. It gives each level a safety score. A high score sends the level to the fast player; a low score sends it to the expert. But you cannot just pick any score cutoff. The researchers use fresh practice levels and a very cautious statistics rule to choose one.
Suppose you allow at most 20% serious mistakes among levels sent to the fast player, with 90% confidence. If there are too few examples or the referee cannot separate easy and hard levels, the system sends nothing to the fast player. That sounds less exciting, but it is safer than pretending to know.
Across 35 OpenML task groups, relaxed tolerance let the transparent decision tree handle about 66% of inputs with roughly 12% violations. A simple 0.5 cutoff broke the safety target much more often. It is like an automatic game difficulty selector: it does not promise every move wins, but it controls how risky the shortcut is!
Glossary
Proactive Routing
Selecting a model before either candidate runs, using only input features. It differs from cascades, which execute a preliminary model first.
The paper’s deployment paradigm.
Gate
A lightweight classifier that predicts whether the surrogate is safe for an input. Here it is Logistic Regression producing a safety score.
It supplies the routing threshold.
Clopper-Pearson
An exact binomial confidence-interval procedure. Its one-sided upper bound conservatively estimates the unsafe fraction in the routed set.
It calibrates t* and yields finite-sample validity.
Violation Rate
The conditional fraction of routed inputs that are unsafe. In the paper, V(t)=Pr(d>τ|s≥t).
The central safety metric.
AUC
Area under the ROC curve, measuring how well the gate ranks safe above unsafe inputs. It affects efficiency but is not necessary for validity.
Theoretical feasibility analysis.
ECE
Expected Calibration Error, measuring agreement between predicted probabilities and empirical frequencies. Lower ECE does not itself create distribution-free validity.
Calibration-efficiency experiments.
Open Questions Unanswered questions from this research
- 1 Exchangeability may fail under temporal or covariate drift. Online, weighted, or drift-aware conformal methods are needed to preserve safety claims in changing environments.
- 2 Global AUC misses the local ROC slope near the selected threshold. A theory that directly optimizes high-purity regions and coverage remains open.
- 3 It is unresolved whether safety, fairness, energy, latency, and multiple surrogate choices can be controlled jointly with finite-sample guarantees.
Applications
Immediate Applications
Medical and financial review
Run black-box and interpretable models offline to create safety labels, train the gate, and calibrate α, δ, and τ on held-out data. Route demonstrably safe cases to the auditable model and retain statistical evidence for governance.
Industrial predictive maintenance
Deploy the shallow tree on edge devices while retaining an ensemble as fallback. The gate reads sensor features before inference; uncertain or poorly supported regions automatically escalate, reducing compute without uncontrolled reliability loss.
Long-term Vision
Auditable autonomous model orchestration
A future system could choose among models with different costs, explanations, and accuracies while jointly constraining safety, fairness, latency, and energy. Major obstacles are distribution shift, richer losses, and certifying multi-model decisions.
Abstract
Model routing determines whether to use an accurate black-box model or a simpler surrogate that approximates it at lower cost or greater interpretability. In deployment settings, practitioners often wish to restrict surrogate use to inputs where its degradation relative to a reference model is controlled. We study proactive (input-based) routing, in which a lightweight gate selects the model before either runs, enabling distribution-free control of the fraction of routed inputs whose degradation exceeds a tolerance τ. The gate is trained to distinguish safe from unsafe inputs, and a routing threshold is chosen via Clopper-Pearson conformal calibration on a held-out set, guaranteeing that the routed-set violation rate is at most α with probability 1-δ. We derive a feasibility condition linking safe routing to the base safe rate π and risk budget α, along with sufficient AUC thresholds ensuring that feasible routing exists. Across 35 OpenML datasets and multiple black-box model families, gate-based conformal routing maintains controlled violation while achieving substantially higher coverage than regression conformal and naive baselines. We further show that probabilistic calibration primarily affects routing efficiency rather than distribution-free validity.