Reproducing the HMA paper

Target paper: Macro-Micro Synergy: Planetary Engineering and Resource Harvesting.

Versions pinned for the canonical results

Fill from the canonical experiment.json provenance block once the revised runs land.

Environment

python3.11 -m venv .venv && source .venv/bin/activate
pip install -r maneuver-map/backend/requirements.txt
# HMA is CPU-bound; inprocess engine avoids HTTP overhead on the
# large micro-agent count.
export ENGINE_MODE=inprocess

Canonical trial — HMA lunar base pad

Submit via POST /experiments:

{
  "name": "hma_lunar_pad",
  "steps": 20000,
  "num_agents": 2000000,
  "dt": 1.0,
  "generations": 1,
  "output_frequency": 50,
  "environment_bound": 2000.0,
  "algorithm": "hma_market",
  "algo_params": {
    "agents": {
      "micro":   { "count": 1976000 },
      "macro":   { "haulers": 120, "printers": 24 }
    },
    "depots": { "count": 6 },
    "hma":    { "auction_interval_sec": 4.0 },
    "energy": { "soc_wh": 100.0, "radio_j_per_kb": 0.3 }
  },
  "seed_base": 42
}

Baselines

The paper reports against:

• FCFS — first-come-first-serve at depots; submit the same body with algorithm: "fcfs_baseline".
• Central planner — true MILP replanner budgeted to 2 seconds of wall-clock per cycle. Implementation lives in gossamer.algorithms.coordination.central_planner (post-P2 refactor).

Energy accounting caveat

The paper’s ε values are mobility-only (kinematic energy from hauler motion + micro excavation effort), not thermal energy for sintering. Sintering thermodynamics are ~1–5 MJ/kg for lunar regolith; add that separately when comparing to thermodynamic baselines. The rewritten paper states this explicitly in §4.5.

Expected wall-clock

Per-trial cost at 2M agents / 20k steps: ~90 minutes on the 32-core reference node. Three seeds per condition for the full table: ~7 hours per method.