# PRS Web Supplementary Materials

Open `index.html` for the local supplementary page. All assets are local;
there is no analytics, external font, or network dependency. This directory
is a local package, not an already hosted public website.

## Contents

- `downloads/poster.pdf` and `poster.html`: research poster.
- `downloads/supplementary-core-data.zip`: anonymous numerical
  evidence export, including the matched-search data, all-seed comparisons,
  target experiments and the 21 recorded lead-time diagnostic images.
- `downloads/auxiliary-observations.csv`: all 21 existing single-reviewer
  observations, anonymized Chinese transcripts, English summaries,
  corrections, and missing fields. The data archive additionally includes
  the `lead_time/human_review/index.html` viewer and per-image transcripts.
  These records are auxiliary and do not enter Q/T/C.
- `downloads/Supplementary_Video_S1_Town10HD.mp4`: the previously completed
  supplementary video, copied unchanged rather than regenerated.
- `downloads/Supplementary_Video_S1_web.mp4`: H.264 browser-compatible
  transcode for the embedded player; frame count, timing, and disclosures
  are retained. The original MPEG-4 Part 2 download is preserved separately.
- `downloads/Supplementary_Video_S1_CAPTION.txt`: original video caption.
- `downloads/lead-time-rgb-cases-01-02.zip` and
  `downloads/lead-time-rgb-cases-03-04.zip`: original RGB inputs for the
  auxiliary review viewer, split to satisfy static-hosting file limits.
- `assets/`: framework and result images rendered from the paper's
  figure PDFs, plus the original video preview image.

## Video interpretation

Video S1 is a visualization-only Town10HD rerun: 1920 x 1080, 20 FPS,
11.5 seconds. Both scheduler markers are overlaid on the same new recording.
Ten corrupted sensor frames in the event segment use previous-valid-frame
hold, explicitly disclosed on screen. The video is not a bit-exact replay
of the quantitative trajectory and is not used in the reported metrics.

## Evidence scope

This package presents the website, poster, video, and supporting evidence.
Earlier drafts, internal audit archives, machine paths, and private artifact
identifiers are not included. No experiment was run for this export.

The research evaluates temporal request placement under identical invocation
counts. It does not equate call counts with tokens, FLOPs, energy, or wall-clock
compute. Training-seed variation and complete matched-search contrasts remain
reported in the paper and data. The data archive is not a complete training
environment or checkpoint release.

The Overleaf source is delivered separately. Follow the submission portal's
accepted file formats and size limits when choosing between this full package
and the smaller, data-only supplement. Author approval and disclosure checks
remain the authors' responsibility.