Reproduction

Does each Ajax agent learn like the implementation it reproduces? Each comparison trains Ajax and a reference implementation (Stable-Baselines3, Brax, or the authors' code) on the same JAX environment and budget, and AdaStop decides how many seeds the answer takes.

• Protocol: how comparisons are set up, run and judged, and what each run records.
• Specifications: every comparison's settings, written before its runs.
• Results: the verdicts, speeds and learning curves.

The harness is benchmarks/reference, run on demand, not in CI. The conformance checks, which do run in CI, compare one update rather than whole runs.

The runs are published as a Hugging Face dataset, which ajax.reproduction.load reads (the reproduction extra).