WARDROBE LOOP · SEPTEMBER 5, 2026
Scene-aware bounded Camotion strength is preferred over fixed .08 for Phase 1.
This chapter records the final bounded Camotion experiment for Phase 1. It is not Camotion 01.13. The 01.8 operator was not changed. 01.12 was not reopened.
Notation: A / B = pristine canonicals. A′ / B′ = Camotion-conditioned shooting frames. A→B = generated traversal shot.
Use the minimum bounded conditioning needed to establish useful motion state.
Stronger Camotion does not automatically mean stronger perceived camera travel. The Cinematographer inspects each actual canonical keyframe and chooses from LIGHT .02 / MEDIUM .04 / STRONG .08. Strength is per canonical, not necessarily one value per shot. Camotion itself remains deterministic: agent reasons, CV observes where useful, Camotion renders, video model films.
Controlled three-pair design
PROTOCOL
Shots: A→B seed 80, B→C seed 81, D→E seed 82. Not C→D. Not E→A. Control = Integration Test 01 01.8 frames at .08/.08. Adaptive = same operator, strength from {.02,.04,.08}. Same Seedance prompt, settings, and seed within each pair. Only intended difference: conditioned endpoint images. Selections frozen in planning.json before any Seedance call. Prompts are the authoritative Integration Test 01 prompts. COMMON_VIDEO_INTENT was not prepended. Seed does not guarantee deterministic reproduction.
FROZEN ADAPTIVE SELECTIONS
Gemini 3.1 Pro inspected the actual stills plus existing CameraMotionPlans. Canonical B is reused as A→B end and B→C start. Nobody independently selected .08; that was preserved, not forced.
A LIGHT .02 · B LIGHT .02 · C MEDIUM .04 · D LIGHT .02 · E LIGHT .02
Therefore A→B adaptive = .02/.02, B→C adaptive = .02/.04, D→E adaptive = .02/.02.
A→B · SEED 80
01 · fixed .08 / .08
CONTROL
prediction skcrer6721rmy0d0edkb60ay5c · retry 0
02 · adaptive .02 / .02
HUMAN 2× WIN
prediction h9v7vh433xrmr0d0ednbc3y58g · retry 0
B→C · SEED 81
03 · fixed .08 / .08
CONTROL
prediction gwfxjk5nv5rmt0d0edq8x6kct0 · retry 0
04 · adaptive .02 / .04
HUMAN 2× WIN
prediction 9rtx0pa8j5rmy0d0eds8k4awyr · retry 0 · B′ .02, C′ .04
D→E · SEED 82
05 · fixed .08 / .08
CONTROL
prediction mmbkwk7qwsrmt0d0edvb230b60 · retry 0
06 · adaptive .02 / .02
HUMAN 2× WIN
prediction g7mfjxhtbnrmw0d0edx8kb2yew · retry 0
Adaptive 3–0, twice.
Primary reviewer: adaptive wins A→B, B→C, and D→E. Independent second review: the same 3–0. Shared qualitative reading: adaptive preserved same-or-better perceived traversal, geometry remained strong, conditioned endpoint blur/smear was less noticeable, and stronger .08 conditioning did not show a compensating locomotion advantage in these three pairs. No numerical quality score.
Complementary to 01.12 and the seed-70 2×2
PRIOR RESEARCH
01.12 still stands: on Ghost Library stills, substantial baked exposure creates objectionable structured-content duplication, and there was no useful pristine-source operating window. Camotion remains motion-state conditioning, not image enhancement. The seed-70 A→B 2×2 still stands: conditioned endpoints plus authoritative Cinematographer locomotion prompting won at one controlled seed. This experiment adds that the useful conditioning does not need to be fixed at .08 for every scene.
Limitations
BOUNDED
Three shot pairs. One seed per pair. Wardrobe Loop only. Gemini 3.1 Pro qualitative selector, not trained on empirical strength/video mappings. No adaptive selection used .08, so this does not establish that .08 is unnecessary in general. Not a universal optimal strength. Not cross-seed proof. Not a claim that .02 is globally best.
Camotion Phase 1 is frozen.
Frozen: 01.8 Route-Preserved Exposure operator, current geometry, destination protection, depth behavior, sample count, route preservation, bounded strength vocabulary {.02,.04,.08}, and Cinematographer ownership of per-canonical strength. Do not start .06, new kernels, sample-count sweeps, compositor experiments, or Ghost Library retunes unless a concrete later movie failure provides reason. The later shootability / intermediate-volume experiment is complete; see chapter 09. TunnelVision Research Phase 1 is complete. See the experiment report.