The figures show one Beijing development example from PanoCity, by Yijing Guo et al. (2026), licensed CC BY 4.0. The source revision, five image URLs, checksums, processing and measured values are recorded in evidence.json.
The target is frame 2 of a five-frame window. A synthetic 35° nadir disc is removed from the target; each donor has its own 25° blocked nadir. GPT-6 Astra uses masked images, capture order and tool-assisted ground registration. Clean target pixels, depth and poses are withheld. Its donor-coordinate submission is committed before scoring. Blinding is procedural on a shared filesystem, not an OS-enforced boundary.
The figure preserves raw donor colors and abstentions. It is a selected expert comparator, separate from the automated PAVER LoFTR system. It does not establish an upper bound or performance on independent routes. The synthetic viewpoints are elevated; this is not measured recovery under a real car.
- Solid donor links identify the two sources used in the committed output. Dashed links are retrieved candidates, unused in that output. Node positions encode capture order, not a surveyed map or recovered camera trajectory.
- Fractions integrate the solid angle of the original panorama mask. They are not rectangular crop-pixel percentages.
- Geometry passes when the chosen source has depth-based visibility support and source-ray error is at most 0.5°. This is a tolerance, not exact alignment.
- The joint check also requires mean absolute RGB-channel error at most 10 on a 0–255 scale. 62.70% passes both; 36.49% passes geometry but fails RGB; 0.35% fails geometry/support; 0.45% is unfilled.
- The broader photometric integrity checks remain failed. These are conditional scores using provided synthetic geometry, not calibrated real-world accuracy.
The dark donor accounts for 86.79% of aligned RGB failures and averages 17.29 channel levels below the reference. That describes the observed difference; the physical cause—exposure, lighting or rendering—has not been isolated.
python tools/render_readme_figures.pyThis uses the four committed crops and evidence JSON to regenerate the PNGs, diagram and GIF. It does not rerun reconstruction or reproduce the complete benchmark. Crop hashes and rendered-file hashes are recorded. No font downloads or network access are needed; the renderer uses Pillow's bundled font.
The GIF contains three discrete assembly states: masked, donor 3, then donors 3 and 1. It is a playback of already committed pixels, not an optimizer trace. Palette quantization affects only the GIF; PNGs retain RGB. All panels share the same 90° downward projection, with no color correction or enhancement.
This visual supplement was added after the original research snapshot release. It does not rewrite that release's ledger or imply the full PanoCity pilot is included in its archives. The original scientific results remain unchanged.