What the model was told it could not change, written down before any frame was generated: 2 stories, 4 porch columns, 3 balcony bays, a five-sided brick bay tower on the right and a single-story glazed sun-room on the left. The building is deliberately not left-right symmetric, which is what tells a frame apart from its own mirror image. Every later frame is checked against this list.

Screening formula: 35% feature continuity, 25% camera motion, 25% source retention, 15% cut safety, over SIFT correspondences with RANSAC fundamental geometry. It never ran a structure-from-motion solve, so it could not see whether the generated 3D was right. Persistent 2D texture matches features even when the geometry underneath is wrong, which is how Veo scored first and then failed the solve.

Step 1: pick a model

Step 2: load it

How far the camera can turn before the model breaks down.

This limit took four attempts. The first was a guess, and its −55° preset showed pure smear. The second came from a sharpness metric, which said three of the four models hold out to ±80°; at +80° there is no building in the frame at all. A collapsing splat does not go soft. The gaussians stretch into streaks and the foliage breaks into confetti, and both of those raise sharpness, so the metric scored the collapse as healthy. The third used solved camera positions, which are accurate but answer a different question: where a solved frame stands, not how far the solve carries. They are symmetric when this property is not.

The fourth was set by looking. 52 frames were rendered through this viewer with the panel hidden, four models across thirteen azimuths at 10° steps, and each was marked clear, degraded or gone. The marks are recorded per frame, so any one of them can be argued with by pointing at the frame.

Wider capture stops buying wider viewing arc.

Line the four models up by how much camera arc each solve was given and the return collapses monotonically: 5.06× readable degrees per camera degree, then 1.75×, then 1.01×, then 0.83×. A 10× spread in capture width bought 20° of extra readable view, and the two widest captures read no wider than the arcs their own cameras stand on. Past that knee, adding generated frames further round the house buys very little. The next gain has to come from the pixels already captured rather than from more camera positions.

Marking invented pixels as unknown, and why the trainer cannot be told.

Feeding generated frames into a reconstruction costs −1.43 dB against a baseline built from solved frames alone, measured on held-out views. The cause is known too. A warped frame is two things stitched together: pixels resampled from a solved frame, and pixels a model invented where nothing was visible. The warp records which is which, per pixel. So the obvious repair is to give the trainer the resampled half and mark the rest unknown.

That repair assumes unknown is something the trainer can be told. The only candidate is a transparent input pixel, and the manual is ambiguous about what one means. So it was tested rather than assumed: a rectangle punched to transparent in half the training views, the other half left intact, everything else identical.

This is strictly worse than feeding the invented pixel through. A wrong color is one bad observation competing with good ones; a wrong emptiness deletes geometry that other views got right. Masking a warped frame down to its resampled pixels would carve the disoccluded regions out of the house. Two short training runs and one rectangle cost about four minutes and stopped a larger experiment that would have returned a negative result for the wrong reason. Sealed as refuted in alpha-semantics-20260811-r1; the same idea still has a coarser form left, selecting whole frames by how much of them was invented.