Optimal Anytime Algorithms for Online Convex Optimization with Adversarial Constraints — exact 6/6 CPU candidate

Source-locked reproduction of OpenReview 7FG30Wld9Q / arXiv 2510.22579. The conservative forecast is 6/6; this is not an official verdict.

• Algorithm 2 is independently transcribed in NumPy and checked at every prefix over dimensions 2, 5, and 10, including a horizon-300,000 run.
• Across 750,000 static rounds there are zero violations of either exact Theorem 4.2 envelope, zero Lyapunov-drift violations, and zero queue-upper-bound violations.
• Horizon-obliviousness is an exact shared-prefix property: truncating a run does not change any action, queue, potential, regret, or CCV value on the common prefix.
• The time-varying transport identity Phi_t((lambda_(t-1)/lambda_t)Q)=Phi_(t-1)(Q) passes 1,374 direct checks; removing the multiplier makes the potential decrease and is caught.
• Algorithm 3 is checked over 600,000 rounds, 30 configurations, two constraint-pressure modes, and realized comparator path lengths from 0 to more than 1,000, with zero regret/CCV envelope violations.
• Algorithms 4–5 are exercised over 192,000 rounds. Objective- and constraint-prediction errors are swept independently; perfect predictors have exactly zero corresponding error and the high-error mutations worsen regret or CCV.
• Eight destructive controls, 48/48 science gates, 70/70 independent tests, and a bit-exact full replay pass.
• Independent NumPy paper-equation implementation; CPU only; no GPU/MPS.

Target: neonforestmist/optimal-anytime-oco-repro. Exact tags: icml2026-repro, paper-7FG30Wld9Q. Resolved artifact target: https://huggingface.co/buckets/neonforestmist/optimal-anytime-oco-repro-artifacts.

Remote publication remains fail-closed until the coordinator assigns a rolling Space slot and reconfirms live tag/target absence.