Best-arch × maps · 32×32 / 10 agents · four maps × 3 seeds (12 runs)

Best-arch vs optimal — how close to the god-view ceiling, map by map?

What it tests: the single best architecture we have — explorer/relay role split, a learned-Lagrangian SOFT connectivity constraint (no hard mask), a frontier-attention LPAC explorer, and SLAM wall-sensing on — run on all four maps at 32×32 / 10 agents and scored as cover_r = 0 REAL visited coverage. Each map has a different god-view-oracle ceiling (open 72% · rooms 37% · mixed 31% · crowded 40%), so every tile is read as coverage-vs-optimal: not the raw percent, but the fraction of what is even reachable in 100 steps on that map. The question is whether the policy is uniformly good across very different terrain or whether corridors and clutter break it.

Key finding Measured live from the finished r9090 seeds below. The best recipe holds ≈ 98–100% real connectivity across every terrain; raw coverage runs ≈ 18–44% (open highest, the obstacle maps lower because their reachable ceiling is lower). The bet: if the fraction-of-ceiling turns out uniform across maps, an architectural improvement should transfer across terrain rather than help one map and hurt another. Three seeds per map.
Per map — mean of the finished r9090 seeds · cover_r = 0 real coverage · connectivity = mean giant component over the rollout
MapOracle optimalReal coverageCoverage ÷ optimalReal connectivityseeds
computing from the rendered rollouts…

Scrub any rollout below to watch the explorer/relay split work the map. The table above is computed live from the rendered rollouts (not hand-typed), so it self-corrects whenever the tiles are re-rendered.

t = 0