---
title: Reproduction - Conditional Coverage Diagnostics for Conformal Prediction
emoji: 📐
colorFrom: green
colorTo: gray
sdk: static
pinned: false
short_description: All 6 claims tested, arXiv 2512.11779
tags:
  - trackio
  - open-reproductions
  - icml2026-repro
  - paper-vaApZm6MKM
---

# Reproduction bundle

Reproduction of **"Conditional Coverage Diagnostics for Conformal Prediction"**
(arXiv:2512.11779v1, OpenReview `vaApZm6MKM`).

The ERT family is implemented from Table 1 -- L1, L2 (Brier) and KL, each as
`R(1-alpha) - R(h)` for a cross-fitted classifier `h` of the coverage
indicator -- together with the asymmetric split of Section 3.3, the k-fold
estimator of Algorithm 1, and CovGap for comparison.

For the synthetic experiments the data-generating process is chosen so the
**true** conditional coverage p(x) is available in closed form, so every
estimate is scored against ground truth rather than against another estimate.
The oracle interval gives p(x) = 1-alpha exactly and is the negative control
that each metric must return ~0 on.

Claims 2 and 5 are audits of the authors' own released outputs
(`ElSacho/Conditional_Coverage_Estimation`), which are vendored under
`authors/`: 2,800 v1 rows, 5,600 v2 rows and 80 classification rows.

```bash
python3 run_claims.py        # all six claims
python3 build_pages.py       # regenerate pages/
```