# Measured results and evidence

## Standalone release verification

One cold workflow run on the project's Apple Silicon Mac, using Python 3.11.16
and the versions in `requirements.txt`. No download or fresh package installation
was performed during this run: existing official weights and the existing
dedicated interpreter were reused. All BF16/Q8 conversion and both conditioning
caches were newly generated in an isolated directory.

Input/output: `examples/input.png` / `examples/result.mp4`. The exact prompt,
profile and artifact hashes are in `benchmarks/cold/result.json`.

| Stage | Seconds, including process exit and normal-pressure wait |
|---|---:|
| T5 FP32→BF16 conversion | 22.833 |
| T5 BF16→Q8 conversion | 13.185 |
| Text encoding | 7.113 |
| Image encoding | 5.076 |
| DiT, four steps | 12.686 |
| VAE decoder | 7.113 |
| FFmpeg encode | 2.539 |
| Complete measured workflow | **83.276** |

The total includes weight-file hashing, stage setup, all stages and final media
validation. It excludes installation, downloads, backend copying/patching and
the initial pressure preflight. The sum of stage times is smaller because it
excludes work between stages. Preparation from scratch is not the same as
uncached OS disk I/O: operating-system caches were not flushed.

MLX allocation peaks: text encoding 5.299 GB, DiT 4.459 GB, decoder 12.570 GB.
These are decimal GB and do not represent total unified memory or RSS.
This run did not record a full system-memory or swap time series, so no zero-swap
claim is made. Per-worker numerical logs and allocator samples are in
`benchmarks/cold/`. Paths were replaced with `<RUN>` / `<WORKSPACE>`; numeric
measurements were preserved. Raw runtime evidence remains local.

Technical checks passed: finite tensors, expected shapes, lossless tensor reload,
49 decoded PNGs, 49 MP4 frames, H.264/yuv420p, 384×224, 24 fps,
2.041667 seconds, no audio, complete FFmpeg decode without errors.
Contact-sheet inspection at frames 0, 24 and 48 shows a preserved orange circle
and floor, with minimal visible motion. This is a technical fixture, not evidence
of strong semantic motion or a quality-approved YouTube shot.

## Historical cached experiment

The original staged experiment recorded **25.763535 seconds** for cache validation,
DiT, decoder, encoding and its original stabilization gates.
Text and image conditioning were **skipped using previously validated caches**.
It used a different private input and a different orchestration harness.
Its performance report and sanitized numerical worker logs are preserved in
`benchmarks/historical/`. Its MP4 and input are not distributed.

Historical diffusion alone took 2.941842 seconds; decoder computation alone took
2.449214 seconds. Neither number is end-to-end latency. Historical telemetry
reported one transient WARNING pressure sample; this is not proof of universally
comfortable operation on 16 GB.

Do not average the two runs or present a speedup: the inputs and measured work
differ. No larger resolution, longer duration, FastMetal performance or
production-quality claim is validated here.