# Camera-guide example: run notes

Tested September 29, 2026. AI-assisted production for Voyager; John Holliman is the approved author. The example uses original fictional geometry, not a real branded product.

## What ran

Blender 4.0.2, Workbench, 960×540, 24 fps, six seconds. The initial desktop creation used the September 29 content-pilot development build. The actual pinned revision used coherent development build fffef4cec with Codex/GPT-6-Astra. Initial setup was performed off-camera. This is not a claim about every released desktop version.

Four paid requests ran through Voyager's managed model CLI: one minimax/h3-max/text-to-video baseline, two original minimax/h3-max/3d-to-video finishes, and one replacement 35-degree finish after fixing the source guide. Each used 768P. No output was discarded to imply a model winner. The first guided request allowed up to one automatic appearance reference; the second supplied a frame from that first finish. They are not identical-input benchmark trials.

The Moda connection did not return itemized charges. On the test date, fal's 3D-to-video documentation listed $0.08/second at 768P plus reference usage. This is provider pricing, not a quote for a Moda account. https://fal.ai/models/minimax/h3-max/3d-to-video/llms.txt

## Source and output

Independent scene inspection found identical mesh vertices and world transforms in v1/v2. Camera endpoint changed 35°→60°; both scenes hold frames 120–144. The agent disabled the oversized floor's Workbench shadow in v2. After publication review, the same rendering fix was applied to the 35° presentation guide, without changing geometry or camera animation. Subsequent temporal review found intermittent quadrant-like background shadows baked into the original first generated finish (for example at 0.54, 0.88, 2.88 and 5.54 seconds). That finish was regenerated from the clean guide with the same appearance-reference frame used by the second finish. The original flashing output remains in original-submitted/. The new presentation output uses the complete replacement take without frame patches, blur or freezes. The 60-degree generated finish is unchanged. The article demo and comparison now use that clean guide. The exact first uploaded guide predates this cleanup and a later local hold-stabilization export; it is retained under original-submitted/ in the download, alongside its original scene and script. The top-level scenes/scripts and offline viewer use the clean presentation versions.

Generated shoulders differ from the gray source; the text-only attempt invented lettering despite the no-text prompt. Generated camera motion was not measured as an exact 35° or 60° path.

## Review

Initial creative review used sampled frames, which missed intermittent generated-background flashes. The replacement finish was then checked across all 144 source frames, with adjacent-frame quadrant luminance measurements. Maximum quadrant mean change fell from 10.53 to 0.31 on a 0–255 scale; the unchanged second finish measured 0.47. These measurements target the reported flash defect and do not certify every aspect of visual quality. Continuous motion has not been creatively reviewed by a video-capable model. Technical decode and browser playback checks are separate evidence. The demo is silent, omits waits, and uses real saved-note and completed-response still holds. No app interface or conversation was reconstructed.

Blender and FFmpeg are separate local dependencies. The HTML download plays included media offline; new renders need those dependencies, and hosted generations require paid access and internet.