Decide nothing here. Scan, grade, next. Anything doubtful → Online, sorted in Pass B. A judgement call puts the loop back to five minutes.
Time a run of ten and let it work out the average. Don't stopwatch single books.
Leave both blank if BiblioScan found nothing — that records as not assessed, which is not the same as worthless.
Cross-check fifteen books, not thirty-five — the top 10 by net value and the bottom 5. These rows are picked for you as soon as Pass A has enough books in it.
Watch for: deep offer depth at around €1 on books BiblioScan valued highly. If that is the pattern, the money is not in barcoded stock at all — and that finding outranks everything else in this run.
Don't chase a price here. No keyed price signal reaches a book without an ISBN — zero of 400 sampled pre-1982 French records carried one. Identify cheaply, then let the market do the pricing.
The model's job is to read the marks, never to name the edition. Ask it which edition this is and it invents a confident, period-appropriate answer.
Copies tab-separated. Open a Google Sheet, click cell A1, paste. No file, no import dialog.
JSON includes the photographs, so it is the one worth keeping — those images are ground truth for every future model test.
Everything lives in this browser only. Nothing is uploaded anywhere except the photographs you explicitly send for extraction.
Optional. Without a key the pre-ISBN section falls back to copy-prompt / paste-result, which works everywhere.
Everything here — camera, barcode scanning, automatic extraction — needs the page served from an https address. Deploy it once and it is a URL you open on any phone in the shop:
bunx wrangler pages deploy docs/testing/site \ --project-name petites-fugues-atelier
Then open the .pages.dev URL on the phone and use Add to Home Screen — it launches full-screen with no browser chrome. Your data and key stay on each device; nothing syncs between them.
Below either line → shelf. At or above both → online. No data → online, because unknown is not worthless. These are a starting hypothesis; this batch is what calibrates them.
Used for both automatic and manual extraction. Worded to force transcription rather than recall.