# Original JPEG feasibility — 2026-09-14

The three photos in `data/2025-09-01-HarrisHillCCW-highres` are 2048 × 1365/1366 pixels. They are individual photos of James from a different track day, not expert/subject pairs. They remain unpaired in both the manifest and reviewer.

Before viewing detector results for these photos, Codex recorded thirteen provisional source-only garment landmarks with 25-pixel uncertainty. They are not independent human anatomical ground truth; older 2026 detector outputs had already been seen. The white-helmet rider on the silver motorcycle was selected, excluding the other rider in H02. No hidden opposite-side anatomy was annotated.

`artifacts/highres-feasibility/index.html` contains six raw inspection PNGs and three outcomes. Both 2D and 3D returned a pose for H01, H02, and H03, without runtime errors. Source hashes were verified unchanged, and a copy of the input manifest is preserved in the output directory.

Some visible locations look closer to their source landmarks, but substantial errors remain. For example, H01's projected left elbow is about 12 pixels from the provisional source mark while its left ankle is about 174 pixels away, near the seat rather than the boot. H02's projected left elbow differs by about 94 pixels; its 2D left elbow does not pass the display-confidence threshold. These illustrative distances describe provisional annotations, not a validated accuracy benchmark.

This is not a controlled resolution experiment: the photographs, track day, viewpoints, crop sizes, and poses differ from the 2026 screenshots. More detailed originals support clearer manual review; this run does not establish how much resolution alone improved detection. Keep unobservable joints unassessable and use the editor for visible corrections. No expert comparison or coaching conclusion is implied by the presence of a skeleton.