1Three modes are distinguished (Section~\ref{sec:limitations}). All commands are run from the repository root with the Python virtual environment in \texttt{.venv}; every command below was executed during this study.23\paragraph{Infrastructure tests (no provider access).} \texttt{make test} runs the state-machine, event, gate, evaluator and bundle-hygiene tests. \texttt{.venv/bin/python evaluators/run\_validations.py} re-runs every oracle's seeded validation set (requires the hidden oracles under \texttt{\textasciitilde/wp29-oracles/}).4\paragraph{Artefact replay (no provider access).} \texttt{make analyze} rebuilds the analysis dataset from \texttt{runs/state.sqlite}, recomputes every estimate, table and figure, and \texttt{make paper} recompiles the manuscript; \texttt{release/validate\_release.py} verifies the hashes of the released files and re-runs the headline analysis in a clean copy.5\paragraph{Fresh live replication (requires an Anthropic API key and executor hosts).} Install the gate and key file on the nodes (\texttt{make deploy}), probe isolation (\texttt{make probe}), then register and run an assignment: \texttt{python -m pgar.cli assign --split confirmatory --tasks <ids> --repeats 3 --seed 301} and \texttt{python -m pgar.cli run --split confirmatory --workers 8}. Model stochasticity is not seedable; results of a fresh replication are expected to differ run by run and should be compared at the task-cluster level with the intervals reported here.6