Reproducibility
Every 2026-onward paper lands with seeds, wheel SHAs, code hashes, and the exact experiment configs so a reviewer (or a future us) can rerun from scratch and get the same numbers back.
All runs are driven by Maneuver.Map’s
/experiments endpoint. Each experiment writes an experiment.json
with a provenance block that captures git state, installed-package
hash, and hardware fingerprint. See the
provenance module source .
Per-paper runbooks
| Paper | Canonical experiment ID | Runbook |
|---|---|---|
| ICCD (CRDT + DTN, ) | exp_iccd_main_2026q2 | /research/reproducibility/iccd |
| DMB + TF-ACO (phase transitions, ) | exp_dmb_main_2026q2 | /research/reproducibility/dmb_tfaco |
| HMA (lunar construction) | exp_hma_main_2026q2 | /research/reproducibility/hma |
Every result table in §5 of each paper derives from one canonical
experiment.json. Plugging the exp_id into notebooks/load_experiment.py
loads the run as a pandas/xarray object and the matching template
(phase_diagram.py, criticality.py, etc.) reproduces the figure.
The reproducibility contract
For each paper we commit to:
- Seeds. The seed tree is published: experiment-level, per-
generation, per-candidate, per-repeat. Rerunning with the same
leaves yields byte-identical
final_swarm_metrics. - Wheel SHAs. The exact Gossamer and Leviathan wheels that produced the numbers, pinned in the appendix.
- Configs. JSON bodies that were POSTed to
/experiments. Copy and paste into a local Maneuver.Map to re-run. - Provenance.
git describe,pip freezehash, CPU count, and CUDA visibility at run time. - Artifacts. Expected
experiment.json, final figures, and summary CSVs produced by the run, archived at a stable URL.
When a future paper refines a baseline, the old runbook keeps working; we version the configs rather than overwriting them.