CREDO

Conformalized Risk Estimation for Decision Optimization

A compact interactive demo for auditing whether prescribed decisions remain near-optimal under uncertainty.

Try Demo Paper Code Citation

Wenbin Zhou · Agni Orfanoudaki · Shixiang Zhu

Interactive CREDO Demo

Audit a candidate decision in simplified optimization problems

Compare saved tabs for simplified fixed-context linear, quadratic, Knapsack (2D–2D), and Knapsack (4D–2D) demos, with Linear and Quadratic switching between 2D and true 3D views.

Current z: (0.00, 0.00)
Switches between linear, quadratic, and finite knapsack examples. Use Dimension / view to choose Linear 2D or true Linear 3D.
The number of extreme points defining the editable feasible decision region \(Z\).
3
The distribution type \(P_{Y\mid X=x}\) that the generative model is fitted over to generate outcome samples. In this simplified demo, \(x\) is held fixed.
Controls how dispersed the simulated outcome distribution is.
0.35
The number \(K\) of generated outcomes used to estimate decision risk.
60
The tolerance \(\epsilon\) used to determine whether candidate decision \(z\) is near-optimal for an outcome \(y\).
0.08
Selects the radius construction used in the risk estimator: Monte Carlo, \(p\)-value, or \(e\)-value.
Dimension / view

2D is the default. Linear 3D is a true LP; Quadratic 3D is a true fixed-Q tetrahedral QP.

Optimization problem

\[ \begin{aligned} \underset{z}{\operatorname{minimize}}\quad & y^\top z\\ \text{subject to}\quad & z \in Z = \operatorname{conv}\{v_1, v_2, v_3\} \end{aligned} \] Editable boundary vertices define the feasible polytope \(Z\); here \(m=3\).

Decision space

Drag \(z\) or the boundary vertices.

Feasible region \(Z\) Selected \(z\) Editable boundary vertex

Outcome space

Samples vs. \(\pi^{-1}_{\epsilon}(z)\).

Ground-truth simulated outcome density Inverse feasible region \(\pi^{-1}_{\epsilon}(z)\) 3D inverse volume / slices Near-optimal Not near-optimal

Risk estimate

Estimated non-near-optimal probability.

0.00 \(\hat{\alpha}(z)\)
Approx. true risk under simulated distribution 0.00

Educational 2D approximation; not a reproduction of the paper's full guarantees.

Simplified educational demo. Conformal-style balls and finite-decision margins are visual intuition, not full theoretical reproduction.

Code

Code repository

The official CREDO research code is maintained in the authors' repository and includes experiment notebooks, implementation files, and visualization code accompanying the paper.

View official code repository

Citation

BibTeX

@article{zhou2025credo,
  title  = {Conformalized Decision Risk Assessment},
  author = {Zhou, Wenbin and Orfanoudaki, Agni and Zhu, Shixiang},
  journal = {arXiv preprint arXiv:2505.13243},
  year   = {2025}
}