A Game World That Cannot Lie About Water

SATELLITE TERRAIN — A GAME WORLD THAT CANNOT LIE ABOUT WATER

NASA 30 m elevation → rainfall simulation → 641 verified rivers, branching where the physics allows · 2026-07

1 · Background & problem

A strategy game set on real terrain, built in two steps: 1. render the land from public NASA satellite elevation data, 2. earn the water from a rainfall simulation and a branching river network — nothing drawn by hand.

Problem: procedurally generated worlds routinely cheat about water — rivers that flow uphill, lakes hovering above their valleys — because the generator places whatever the designer asks for, true or not.

2 · Gaps & countermeasures

gap	countermeasure
water is usually placed by hand	rainfall is simulated at 1 m resolution; rivers are selected from the flow results, never drawn
no limit on how much water is added	a hard cap derived from the region's real-world water coverage (~10% of land here)
rivers can carve through the terrain	cut depth is limited to 1.0 m per cell and audited after every build
a river can fail to reach the sea	nine float-probes are simulated down every river; any river that strands them is rejected
real river networks branch (deltas, distributaries)	a branch is accepted only if it passes four physical tests (direction of flow, no cycles, flow share, downhill) — from the published GRIT method

BEFORE · the world was a 387 MB hand-built asset · its rivers were plausible-looking geometry that could not be verified against any physics

AFTER · the world rebuilds from ~3 MB of public elevation data plus code · 641 rivers pass every audit · every river provably drains to the sea · branches enter the network only through four physical tests

3 · Results — the evidence

In-game world map with simulated rivers — **The result in-game:** the world map, captured directly from a playtest. The bright red rivers are a deliberate placeholder color used during development for visibility.

River banks before smoothing — visibly jagged — **Before smoothing:** the raw computed riverbank is visibly jagged.

River banks after smoothing at the chosen production setting — **After smoothing:** the production setting was chosen by visual review, overriding the purely numerical optimum.

River network map with accepted branching links — **The branching network:** a candidate river split joins the network only after passing four physical tests — water flows into it, it creates no cycle, it carries a sufficient share of the flow, and it descends.

A river's mouth meeting the sea as a straight-sided plunge — **A river meets the sea:** the final chute holds straight-sided to the shoreline and plunges below the sea plane — geometry produced by the simulation, not modeled by hand.

HYDROAUDIT, 3-LANE FLEET 634 sources x 9 ducks = 5706 launched: 5691 sea / 0 pool / 6 GAP / 3 RISE / 6 TILT / 0 STUCK

Verbatim line from the drainage audit: 5,706 float-probes were simulated down 634 rivers; 5,691 reached the sea and none became stuck — the proof that every river actually drains.

4A · Resolved: the water is physically plausible and verified

641 rivers shipped with zero audit failures, using 0.40% of land against an 8.0% ceiling · every river provably drains · river branching runs through the same audited machinery

4B · Unresolved: visual polish & larger regions

Rivers still use the placeholder color; the visual-design pass comes next · the pipeline's next target is the Mediterranean, where large river deltas will exercise the network-branching method (based on the published GRIT framework)