arXiv:2602.03168v1 ·
OpenReview EKmiOtjZjI
UP-OCP is implemented from Algorithm 1 / Eq. (11); the KT baseline from
Eqs. (41)–(43); bounds from Theorems 4.3 and 4.4. The paper releases no code,
so the AXP score stream is rebuilt from the conformal-time-series
pipeline and validated against that benchmark's released AMZN forecasts
(corr 0.99999 over 2,919 days).
| Claim | Subject | Verdict | Headline |
|---|---|---|---|
| 1 | Parameter-free | reproduced | worst |miscov−α| = 0.0145 over 15 settings |
| 2 | Theorem 4.3 regret rates | reproduced | 945 checks, 0 violations; both rate regimes located |
| 3 | Theorem 4.4 coverage | reproduced | 135 checks, 0 violations; worst ratio 0.765 |
| 4 | AXP Table 1 | reproduced | 0.9329 / 14.94 / streak 4 vs paper 0.932 / 14.8 / 4 |
| 5 | Pareto frontier | reproduced | UP-OCP dominates KT at 50/50 rates |
| 6 | Synthetic sinusoid | falsified | tracking holds; tuned P-Control is 1.66× better |