Prototype (h): sample collection to try a compute job

Branch proto/sample-collection. One line: eight public-domain page images from the National Diet Library, on which a visitor with no wallet can run an analysis from start to finish — choose pages, choose an analysis, run it, and see that only result files come back while the images stay with the custodian.

• URL: /sample-collection (ja: /ja/sample-collection). Not linked from the menu.
• Collection: 歳旦発句牒 (芭蕉 編, 1682), NDL info:ndljp/pid/2534020, marked PDM. All 8 canvases of the IIIF manifest, 1600 px wide copies, in public/sample-collection/saitan/. Bibliographic values are copied verbatim from the manifest into collection.json (not translated). Page notes ("front cover", "text") are ours and are badged as such.
• Two analyses, each badged with how it runs:
  • Fixity check — runs here, now. SHA-256 and pixel size per file, computed in the browser; CSV download. Each checksum is compared with the checksums recorded by the OCR run ("same file as the OCR run read?"). Verified equal to shasum -a 256 of the files.
  • Text recognition (NDL 古典籍 OCR-Lite) — replays a recorded run. Run once on 2026-09-26 on the author's Mac, in the same container and under the same rules as an Ocean Node compute job (no network, 2 CPUs, 2 GB, uid 1000, write access to /data/outputs only). Record in public/sample-collection/saitan-ocr-run/ (run.json = container digest, command, input checksums; logs.txt verbatim; outputs/ = txt / TEI / NDL XML / JSON per page). The log is shown unchanged; the chosen pages' lines are highlighted.
• Results view: files in /data/outputs vs "what did not come back" (the images). OCR viewer: pointing at a line highlights its box on the page.
• Honest parts kept on purpose: page 8 (back cover, no text) failed inside the OCR program (ZeroDivisionError on zero lines) — shown as "page 8 failed, and the job still finished". A "read this as a draft" note gives two real errors from this run (歳旦 → 歳且; the ruler read as "1.3.3").
• Section 5 lists what changes on the real network (catalogue, wallet, node, minutes) and what stays the same, and says the test-network trial is not open yet.

Known limits: the OCR cannot run in the browser (models ~79 MiB, ~70 s per 8 pages on 2 CPUs), so it is a replay, and the page says so. The phone header overflows by 8 px on every page of the site (existing menu, not this page).

Next step if chosen: publish the same 8 images and the OCR algorithm as a compute asset on Sepolia via the mdx-clio node, so the "real network" section can link to it (needs Vicki's OK for a public trial).

Screenshots: page-ja.png, fixity-en.png, ocr-result-ja.png, ocr-failed-page-en.png, phone-ocr-ja.png.

Source: docs/prototypes/sample-collection/

Screenshots

fixity-en.png

ocr-failed-page-en.png

ocr-result-ja.png

page-ja.png

phone-ocr-ja.png