Casa Cornell exterior, motion experiments

Everything below is sealed run evidence, bundled with the page. Videos loop; click any to scrub. Sealed manifests: verification/generative-video-splat-20260803-r2/ and verification/camera-guide-20260804-r6-dolly-candidates/.

1 Â· R2 paid generation (Wan VACE depth, $0.51, seed 20260803), rejected, generator exonerated

Artefact-free and faithful to its control, but the control (R5 guide) was rotation-dominant, so COLMAP classified every pair as homography. The failure was inherited, not generated.

Generated clip (81f Â· 720p)

fal-ai/wan-22-vace-fun-a14b/depth Â· request 019fcacc-2504â€¦

Its control: R5 guide (13.25Â° yaw, ~0.03u translation)

210/210 pairs homography, fails the arbiter itself

2 Â· r6 dolly candidates, translation-dominant controls (free, local renders)

Pure lateral dolly, yaw exactly 0, from the same pinned source photo + depth. Only the baseline differs. Orchestrator-run COLMAP gate results under each.

t015 Â· 0.15u rejected

77/210 epipolar pairs Â· mapper: no model

t025 Â· 0.25u selected control

124/210 epipolar Â· mapper (pinned intrinsics): 21/21 images, 7,830 points, track 9.6, first real sparse model in the project lineage

t040 Â· 0.40u rejected (appearance)

156/210 epipolar Â· solves 21/21, but depth-map error visibly widens the facade at 15Ã— the R5 baseline

3 Â· R3, the accepted run (Wan VACE depth + t025 control, $0.51, seed 20260803)

Same prompt, params, and seed as R2, only the control changed. The arbiter fully solves it: 21/21 images, 10,781 points, mean track length 10.2 (and it solves even under the legacy no-intrinsics protocol that R1 and R2 failed). First reconstructable generated clip in the project lineage.