3D engine that AI agents drive
Flicker probe results
traps-1280x720.json is the ground-truth output of node bench/flicker/probe.mjs --traps --viewport 1280x720 --assert with
configs reversed=1;reversed=0;reversed=1,offset=0;reversed=0,offset=0 (reversed-Z float depth and 24-bit depth, flush pass on and
off). It needs a GPU browser, so it is a local run, not part of CI; rerun it after a change to the flush pass or the depth path.
--assert exits 1 if any trap shows a swapped pixel with the flush pass on.
melbourne-drape-1280x720.json is the exact ground truth (facets=2, see probe.mjs) of the Melbourne drape GLB at 12 camera
spots (roofs, bridge, river, overview and 8 glass spots), with the flush pass on and off (offset=0), reversed-Z float depth, shadows off.
Summed over the 12 spots: 8 true visibility flips (flipped) with the flush pass on against 109,522 with it off. The 8 are single
pixels on a triangle edge, where the ray cast and the rasteriser disagree about the pixel centre, not coplanar fights. offRay (pixels
showing a face the pixel-centre ray misses) is the rasteriser and the ray cast differing at edges and is not depth fighting. A render
pass (albedo, object ID) cannot stand in for this: its target has a 24-bit depth buffer, which fights far more than the beauty path.
melbourne-drape-offray-1280x720.json is the same exact run with offRay taken apart, at hits=6 and hits=32 (hits kept per ray in the
ray cast) with the flush pass on, and hits=6 with it off. Summed over the 12 spots, flush pass on: offRay is 18,590 at both hit
counts, so no pixel hides in a stack deeper than 6 faces. It is not edge-only: 363 touch an owner change of the cast in their 3x3
neighbourhood (offRayEdge), 18,227 do not (offRayInterior). But the colour shown there is within 16 bytes of the colour expected
for 18,367 of the 18,590 (offRayNearColour 8,261 within 2, offRaySmallGap 10,106 within 3 to 16): consecutive triangle numbers are
consecutive colours, so a rounding step of the frame names an unrelated triangle. Only 215 interior pixels are further off
(offRayUnexplained), so the true flips are at most flipUpperBound = 8 + 215 = 223 of 5.9 million contested pixels. With the
flush pass off the same bound is 129,361. Reproduce with probe.mjs --configs "reversed=1,facets=2,shadows=0,hits=6;reversed=1,facets=2,shadows=0,hits=32;reversed=1,facets=2,shadows=0,offset=0,hits=6".
melbourne-drape-single-sample-1280x720.json is the same 12 spots with 4 samples (samples=4) and with 1 sample (the default of facets=2
since #751), flush pass on, and 1 sample with the pass off (offset=0). The 18,590 offRay pixels of the 4-sample frame were the MSAA
resolve: a pixel shared by two faces resolves to the average of their ids, and ids of neighbouring triangles are neighbouring numbers, so
the average names a third, valid triangle (a pixel that the ray through its centre cannot match). viewer.js colourStepDetails() shows
it: the wrong pixels differ from the expected id in one or two channels, never three, by a few bytes, in runs along thin faces. With
1 sample the same frames hold 8 wrong pixels in 5.9 million contested ones, no swapped pixel and no true flip (flipUpperBound 1); the
8 "true flips" of the 4-sample frame were blends too. With the flush pass off the 1-sample frame still shows 119,404 flipped pixels, so the
probe still sees fighting. Reproduce with probe.mjs --configs "reversed=1,facets=2,shadows=0,samples=4;reversed=1,facets=2,shadows=0;reversed=1,facets=2,shadows=0,offset=0".
effects-melbourne-drape-1280x720.json is effects-probe.mjs on the Melbourne drape at the same 12 spots (per-part colours, ids=1, shadows
off, reversed-Z float, 4 samples, D3D11): AO, line art and DOF through the stacked scene with the flush pass on, against the same scene with the
pass off. spurious counts the pixels an effect changes where the plain frame has no colour edge within 2 px. On a real city AO and DOF
change many such pixels by design (real occlusion, a blur of a shaded gradient), so the absolute counts say nothing; the two sides are
the comparison. Summed over 12 spots, flush on / off: AO 2,066,622 / 2,052,039, DOF 822,357 / 824,561, line art 44,654 / 50,794.
traps-exact-1280x720.json is the exact ground truth (facets=2, every triangle drawn unlit in a colour that is its own number) of every
trap scene, 88 rows: 11 traps x reversed-Z float depth and 24-bit depth x flush pass on and off x 1 sample and the default 4 samples
(MSAA). Reproduce with node bench/flicker/probe.mjs --traps --viewport 1280x720 --prefix exact --configs "reversed=1,facets=2,samples=1;reversed=0,facets=2,samples=1;reversed=1,facets=2,samples=1,offset=0;reversed=0,facets=2,samples=1,offset=0;reversed=1,facets=2,samples=4;reversed=0,facets=2,samples=4;reversed=1,facets=2,samples=4,offset=0;reversed=0,facets=2,samples=4,offset=0".
failingTraps judges an exact run on flipped (a face behind covers one in front that is not ranked below it), not on swapped, which
also holds the faces the flush pass lifts on purpose. Flush pass on: 0 flips on all 44 trap runs, 0 offRay, 0 blended, 0 wrong.
Flush pass off: 0 flips on the stacked traps too, because exact coplanar faces are tied (the scene names no winner, e.g. 335,243 pixels on
plaque-2m, 137,786 on kitbash-duplicates, 122,353 on stack-30-exact); the one trap with a real flip off is plaque-closeup-0.5m (330 flips
at 1 sample on reversed-Z, 207,626 on 24-bit), where the wall must be cut around the plaque.
The 18,365 wrong pixels that traps-1280x720.json shows for plaque-closeup-0.5m with the flush pass on are not flips: that run draws
the lit scene, where wrong also counts shading steps and anti-aliased edge pixels (inferred: the lit frame has 18,365 of them, the unlit exact frame none). The exact frame has 0 wrong pixels.