PROTOCOL.md

Reproduction protocol

The decision set is the Euclidean unit ball and the adversarial constraint is a half-ball. Linear objective gradients have norm at most one, so the exact paper assumptions hold with G=1 and D=2. The feasible comparator for every prefix is solved analytically, never estimated. All arrays are deterministic NumPy float64 values on CPU.

Algorithm 2 / Theorem 4.2

The implementation uses the paper's preprocessing alpha=(2GD)^-1,

lambda_t = 1 / [4 sqrt(t) sqrt(log(t)+1) (log(log(t)+1)+1)],

Phi_t(Q)=exp(lambda_t Q)-1, and

Q(t)=(lambda_(t-1)/lambda_t)Q(t-1)+alpha g_t^+(x_t).

The surrogate gradient is passed to the displayed adaptive OGD/AdaGrad update. Every one of 750,000 prefixes is checked against the exact regret envelope 2GD(sqrt(t)+1) and exact CCV envelope 8GD sqrt(t) sqrt(log(t)+1)(log(log(t)+1)+1)log(4t). The matrix covers 16 configurations, dimensions 2/5/10, and a maximum horizon of 300,000. Shared-prefix replays prove that neither lambda_t nor the action update uses the final horizon.

Time-varying Lyapunov certificate

At every round the reproduction checks the convex drift inequality, potential monotonicity, and Q(t)/alpha >= CCV_t. A separate 1,374-point grid checks the exact transport identity that compensates for the pointwise decrease of Phi_t. An additive-queue mutation removes the lambda_(t-1)/lambda_t factor and produces a strict potential decrease, so the audit is sensitive to the paper's core mechanism.

Algorithm 3 / Theorem 5.3

The dynamic update replaces lambda_t with the paper's lambda_t/sqrt(1+P_t) and uses only the path length observed through the current round. Thirty configurations cover five motion rates, three seeds, a feasible-cost suite, and a constraint-pressure suite. Each of 600,000 prefixes is checked against both exact Theorem 5.3 envelopes. Truncated executions reproduce full-run prefixes exactly, proving that neither future path length nor final horizon is consulted.

Algorithms 4–5 / Theorem 6.1

The optimistic meta-algorithm uses delayed multiplicative queues and Euclidean optimistic mirror descent. Six prediction-noise levels are swept independently for objective and constraint predictions over 192,000 rounds. Perfect objective predictions yield exactly zero E_t(f~); perfect constraint predictions yield exactly zero E_t(g~). A 512-point certificate recovers the square-root error exponent as 0.5, and four deterministic error schedules verify the lambda-weighted summability term used by the proof. Corrupting objective predictions worsens regret; corrupting constraint predictions worsens CCV.

Replay

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
CUDA_VISIBLE_DEVICES=-1 PYTHONDONTWRITEBYTECODE=1 \
  .venv/bin/python reproduction/run_experiments.py
.venv/bin/python reproduction/run_tests.py
.venv/bin/python reproduction/verify_repeatability.py

Release scripts build the poster, execute a genuine local trackio.log_artifact, render the six-page judge-facing logbook, and then audit ingestion, privacy, source hashes, and publication metadata. No remote write occurs during these steps.