Type a section name and press Enter. Use “help” for the list.
Cliste experiment
What did learning add beyond the same visual conversion?
Registered 11 Oct 2026 · Snapshot 12 Oct 2026
The question
Separate conversion cost from the benefit of learned temporal construction.
Why this experiment?
Missing-pixel prediction did not establish coherent construction. Video endpoints provide another concrete question: can a mechanism construct an intermediate scene while using the shared VSA interface?
How it works
Keep the already trained predictors fixed and add an untrained feature blend through exactly the same recovered-feature conversion. This separates the contribution of learning from the contribution of the shared interface.
Follow this part of the investigation: Which mechanism should construct new visual content? →Implementation and comparison details
Add endpoint feature interpolation through the already fitted adapter and unchanged VSA recomposition; reuse the preceding learned outputs without fitting weights.
What happened
Against the matched feature blend, synthesis lowers whole-image error by 0.29% but changes changing-pixel error by +0.14%. Transport changes changing-pixel error by +3.91%. Neither reaches the required improvement on each video. Synthesis makes small gains on many frames; two train cases outweigh them. All 32 cases remain visible, and neither model consistently resolves doubled moving contours.
Results and visual artifacts
Open a figure for its input, comparison, and result. Click an image to enlarge it.
Each case shows both supplied endpoints, the withheld middle for assessment, direct VSA interpolation, the matched feature interpolation control, and both unchanged learned predictions. All cases also appear in indexed overview sheets. The no-fit control still uses the fitted visual adapter; it removes learned temporal construction, not all learning. This public record includes 1 of 36 saved figures: summary comparisons and fixed examples, including every figure referenced by the guide. Reported aggregates include every case; individual records and complete outputs remain in the research archive.
Video example 10: matched comparison
Display derivative: WebP quality 90; maximum edge 2400 px. Original scientific pixels remain in the source run; do not measure this preview.
Limits and next question
The same 32 previously exposed test triplets from two videos. This is a diagnostic follow-up, not new independent validation, an optimized model, or a causal isolation of every source of error.
Next question. Train and evaluate both constructors on original encoder endpoint features with their own untrained blend, keeping all other training conditions and full VSA outputs matched.
Inspect the evidence
Download the methods and measurements for this experiment. Large training and saved-state files are not included.