Seed-42 impulsive-MGA route discovery
This is the reviewed publication bundle for the 100-route seed-42 experiment documented in the route-search tutorial. A finite score means only that the Rust optimizer found an impulsive MGA scaffold under the declared budget; it is not a low-thrust-feasible or competition-valid trajectory.
random, evolutionary, and gemma4 form the blind matched comparison.
gemma4-assisted is a separately named follow-up which uses the completed
random and evolutionary archives as prior evidence. It must not be presented
as an independent fourth arm.
Each arm retains:
- the terminal
run.jsonand immutableprotocol.json; - compact and full accepted-route archives;
- convergence and proposal-rejection logs; and
- the Rust→adapter exchange log, empty by construction for non-agent controls.
Per-request response caches are deliberately omitted because they duplicate adapter responses. The provider prompt and candidate menu are reconstructed deterministically from these exchanges, the adapter, and any digest-pinned prior archives; they are not stored verbatim. The original local work directories remain outside the repository.
Regenerate or verify the comparison from the tutorial directory:
python3 compare_campaigns.py
python3 compare_campaigns.py --check
python3 plot_results.py --check
The experiment has one root seed and parallel coordinated retry is not bit-reproducible. Repeat predeclared seeds before making a general proposer or model-capability claim.