The gate rendered its lookdev half at erosion_render's OWN defaults — strength 0.22, erosion_scale 0.15 — while the Godot half ran the map's 0.0057 / 0.0014. Two different terrains, reported as MATCH max |godot - lookdev| = 904.639 m FAIL which reads as the engine and the lookdev disagreeing about the terrain. They did not. The gate was comparing parameters, not implementations. Fixed by PARSING the constants out of map_dc_diffusion.gd — floor, ceiling, terrain height, gully band, depth, K, cell scale, octaves, gain — and mirroring the map's own two derivations. A gate that keeps a private copy of the numbers it is checking drifts from them silently, which is precisely the failure being fixed, so it now cannot hold one. With both halves current and the depth formula corrected: MATCH max |godot - lookdev| = 0.002 m rms 0.001 m PASS CONTROL erosion off: max diff 16.8 m REJECTED Two millimetres over a 1.8 km window between DCTerrain::sample_density and a standalone Vulkan compute shader that shares no path with it beyond erosion_filter.inc and the wrapper. And the control now DISCRIMINATES, which it could not before: at the old 0.227 m amplitude, erosion-on and erosion-off were the same field inside the 2 m tolerance, so the gate accepted anything. That blindness was the third independent sighting of the depth bug, after the renderer at map settings and amplitude_law.py. ⚠ Re-run BOTH halves. Editing the map and re-running only the python half compares a fresh lookdev against a stale /tmp/dc_surface.txt; that misread cost a cycle here and shows up as a MATCH failure of about the depth.