control
control_static
Final loss 0.000 · travel 0.000 Blender units · lens change 0.0 mm.
Closed-loop camera experiment · Classroom · v0.1
Can VTL steer a release shot?
Early evidence, yes, from one bounded test. A deterministic search varied the Blender camera while the room, lighting, and starting composition stayed fixed. The best searched path beat the best ordinary moving control at both preview and final cadence, on the loss it was selected by. It has not yet been compared with an equal-budget random search, the test that would show the measurements guiding the search rather than only scoring it.
search_07final winner
24.8%lower loss than moving control
0.886Blender units traveled
What we built
The loop generated 24 deterministic camera paths plus four controls from the same production camera. It rendered nine-frame Cycles previews, measured structural G1, the VTL kernel, ISR island turnover, and the experimental Layer 6 drain, then selected three parameter-diverse finalists. Those finalists, the strongest moving control, and the static floor were rendered as 49-frame, 24 fps clips and measured again.
The protocol was frozen before rendering. Candidate measurements did not redefine the kernel scales or objective. The source Classroom file remained unchanged; exact camera transforms, lenses, targets, render hashes, and source hashes are recorded in the manifests.
Watch the result
VTL-selected
search_07
Final loss 0.544 · travel 0.886 Blender units · lens change 3.4 mm.
VTL-selected
search_16
Final loss 0.644 · travel 1.026 Blender units · lens change 14.3 mm.
VTL-selected
search_18
Final loss 0.691 · travel 1.168 Blender units · lens change 16.7 mm.
control
control_dolly
Final loss 0.724 · travel 1.100 Blender units · lens change 0.0 mm.
Final-cadence comparison
The preview win survived the denser 24 fps evaluation. Lower is more release-like under this experiment’s objective. The static shot is shown as a continuity floor, not as an eligible winner.
| Path | Loss | Travel | Kernel Δ | ISR turnover | L6 drain Δ | G1 range |
|---|---|---|---|---|---|---|
| control_staticcontrol | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| search_07search | 0.544 | 0.886 | 0.082 | 0.010 | 0.013 | 1.448 |
| search_16search | 0.644 | 1.026 | 0.090 | 0.053 | 0.009 | 1.227 |
| search_18search | 0.691 | 1.168 | 0.126 | 0.065 | 0.005 | 0.922 |
| control_dollycontrol | 0.724 | 1.100 | 0.097 | 0.051 | 0.026 | 1.443 |
What the system discovered
The winning move is controlled, not inert. It travels 0.886 units while preserving the room’s relational cast: the chalkboard remains the primary structural anchor, the desk field advances without repeatedly replacing its islands, and the light pattern changes gradually. Its advantage is distributed across the VTL spine rather than produced by one zeroed metric.
The straight dolly is a meaningful control. It travels 1.100 units, so the search did not win merely by refusing to move. The dolly expands foreground desks and window contrast more uniformly, increasing structural G1 variation and leaving Layer 6 too uncertain to contribute a measured drain term.
The diverse finalists expose the objective’s shape. They are not runner-up copies. Some narrow the lens and transfer dominance toward the board; others introduce stronger lateral or vertical change. They show where “release” becomes emphasis, reveal, or instability. This makes the system useful as a navigable field of alternatives, not merely an automatic best-shot button.
Layer 6 remains diagnostic. Its null gate is honored and uncertainty is penalized. The result therefore does not claim that the drain is a stable steering signal. The most credible evidence remains continuous kernel change, ISR membership continuity, and structural G1 behavior.
What this proves, and what it does not
This is evidence that VTL measurements can close a loop around actual scene inputs: generate, render, read, select, and reproduce. It is not yet evidence of general cinematic taste. The scene count is one, the path family is bounded, the weights encode a release hypothesis, the comparison is against ordinary camera moves rather than an equal-budget random search, and the metric winner still requires human viewing. The next serious test is repeated optimization across different scenes and deliberately opposed objectives such as suspense and clarification, followed by blinded human comparison.
Reproducibility
The preview and final rankings, exact final camera states and the frozen protocol are retained in the experiment archive.