# Bottle study 02 — five contrasting wines

Live: https://wine-bottle-poc.dbooij.com

## Revised selection
The first study used five Ridge wines, which was too homogeneous. This revision replaces every active specimen. These are deliberately selected for contrast, not a statistically random sample.

| Year | Wine | Region | Bottle | Size |
|---|---|---|---|---|
| 2015 | Joh. Jos. Prüm Wehlener Sonnenuhr Riesling Kabinett | Mosel, Germany | Tall flute; white/gold neck and scenic label | 750 ml |
| 2001 | López de Heredia Viña Tondonia Reserva | Rioja, Spain | Classic shoulder; wire net and mature label | 750 ml |
| 2010 | Travaglini Gattinara | Piedmont, Italy | Asymmetric body, separate vintage stamp | 750 ml |
| 2012 | Dom Pérignon Vintage | Champagne, France | Broad body, long foil and shield label | 750 ml |
| 2018 | Juliusspital Silvaner trocken | Franken, Germany | Flattened Bocksbeutel | 3 L Doppelmagnum |

“Kabinett” is the interpretation of the user's “cabinet”, given the Mosel context. Each specimen has a different producer, label and silhouette. There are four countries and two German regions. Vintages span 2001–2018. Sizes are explicit: the producer's Bocksbeutel reference is 3 L, not 750 ml. All models use a common display height, not a common physical scale.

## Manual process performed
1. Selected contrasting wine styles, countries, vintages and bottle formats.
2. Located exact vintage/cuvée pages and inspected all five seed photographs. Prüm's label visibly says 2015, Wehlener Sonnenuhr and Kabinett (not Spätlese); Tondonia says 2001 Reserva (not Gran Reserva); Travaglini's separate vintage stamp says 2010; Dom Pérignon's shield says Vintage 2012; Juliusspital's label and producer sheet say 2018.
3. Preserved original source images unchanged. The Juliusspital reference is the original embedded JPEG (PDF image object 19) extracted losslessly from the producer's vintage-specific PDF, which is also retained. The Travaglini seed retains the auction house watermark. `wines.json` records sources, SHA-256, image notes and crop pixels.
4. Followed public CellarTracker tasting-note links where available. Direct CellarTracker retrieval remains blocked. Tondonia maps to iWine=874025 and Dom Pérignon to iWine=3979364 via cited public notes. Exact CellarTracker IDs for the other three remain unresolved and are marked as such. The viewer offers a search link for unresolved entries; it does not present that as a verified record.
5. Manually bounded each bottle in its photo, traced smooth profile anchors guided by the visible outline, and retained changes in the centerline. Initial automatic dark-edge tracing produced reflection artifacts and was replaced after visual inspection. Reconstructed elliptical cross-sections. The flattened Bocksbeutel uses an inferred body depth/width ratio of 0.48; Travaglini uses 0.72 and a changing centerline. Both blend to a near-circular neck. These depth assumptions are not measured from the image.
6. Chose label crops manually. Applied the actual photographic labels, preserving year and typography. Travaglini gets a second patch for its vintage stamp. No image generation is used. The geometry and texture choices are documented in `reconstruct.py`, a one-off reconstruction helper, not an ingestion pipeline.
7. Exported five standalone glTF files with embedded geometry and textures and updated the existing temporary site. Three.js 0.170.0 is vendored, with its MIT license.

## Limitations
Only the front image is observed. Depth, rear surfaces, base, closure detail and material properties are approximations. Photographic shading and perspective remain in label crops. The Rioja wire net, embossed glass, foil wrinkles, neck artwork and unobserved back labels are not reconstructed. This establishes a more diverse manual visual experiment, not automatic wine identification, calibrated scanning or recognition accuracy. Attribution does not establish production reuse rights for third-party source photos.

## Files and reproduction
- `wines.json`: current sources, identity status, model assumptions and crop coordinates.
- `sample-export.csv`: synthetic example export with size and region.
- `assets/`: original seeds, original producer PDF and derived texture crops.
- `models/`: five self-contained glTF 2.0 models.
- `archive/ridge-*`: first-study notes and manifests; original binaries remain in Git history at cd4d0ff.
- `reconstruct.py`: requires Pillow. On this machine run `/workspace/shared/.venv/bin/python reconstruct.py`.

Source: `infra/analysis/wine-bottle-poc`. `/workspace/analysis/wine-bottle-poc` is a convenience symlink on this machine.

## Deployment and removal
From this folder run `python3 /workspace/repos/infra/infra.py deploy`. The existing subdomain and route are reused. `/health` returns `ok`.

To pause: `python3 /workspace/repos/infra/infra.py pause --reason 'temporary experiment finished'`. For full removal, remove only this project's routes from `infra/caddy/projects.Caddyfile` and deploy Caddy, remove its project registry entry, then remove this project's containers and `/opt/wine-bottle-poc` directory on projects. Wildcard DNS needs no separate record deletion. Retain the Git source if the experiment should survive.
