01 / InputFormat and size
LUN accepts JPEG, PNG, or WebP images up to 8 MB. Validation runs in the browser and is checked again on the server.
- One invoice or receipt per image.
- Front-facing, complete, and unobstructed.
- Focused, well lit, and free of glare.
02 / CaptureHow to photograph it
- Lay the receipt flat and frame it from the front without excessive perspective.
- Include the header, every line, and the final total.
- Use even lighting; avoid shadows, glare, and motion.
- Before sending, check that names and prices are distinguishable.
03 / More useful readingsWhich receipts work best
Extraction works best when each line describes a real product and retains identifying details.
Include, when shown
- Individual lines, not only a grouped total.
- The full product name.
- Brand, variant, SKU, or identifiable code.
- Date, prices, quantities, and printed total.
Categories with strong potential
Grocery, cosmetics, cleaning, and household goods. Textiles or electronics also work when the product, model, or variant remains identifiable.
Low-efficiency categories
Generic restaurants, services, fuel, and transport often hide the product or the evidence needed. Totals without lines and cryptic abbreviations also limit the reading.
If the image only shows ‘service,’ ‘consumption,’ or a code without context, information must not be filled in by intuition.
05 / CriteriaWhat can support a SUS line
SUS requires product-specific evidence, not a general impression. These are examples of strong evidence that may support a conservative reading:
- An explicit, recognized product certification.
- The exact product found in a sourced catalog.
- Quantified recycled content.
- Certified material.
- An unequivocally refillable or reusable product.
- Recognized energy efficiency.
- Identifiable electric mobility or solar product.
- Certified organic product.
Insufficient: NOS is retained
‘Natural,’ ‘green,’ or ‘eco’ without support; merchant reputation; the brand alone; an ambiguous product; or illegible text do not demonstrate the product's condition.
One result: product-by-product classification
The interface lists each detected product with its SUS or NOS label and summarizes the monetary share of SUS among analyzed lines. The summary is not a rating for the whole receipt and states coverage or discrepancies against the printed total.
07 / ExamplesTeaching examples
These fictional examples explain the product labels; they are neither real receipts nor benchmark results.
Example A · mixedDetergent refill → SUS · Rice → NOSThe refill has exact supporting evidence; the rice line has no product-specific evidence and remains NOS.
Example B · exact productElectric bicycle model X → SUS · Service → NOSThe identified bicycle can be matched to concrete evidence. A generic service line cannot.
Example C · no evidenceGreen product → NOSWithout certification, source, or an identifiable description, the vague claim is insufficient and remains NOS.
A conservative signal can become trust infrastructure.
LUN's value is not in labeling more products sustainable, but in transforming scattered evidence into a consistent, explainable, and extensible reading. Every NOS caused by missing evidence protects system credibility; every verifiable SUS strengthens a reusable catalog.
Defensible advantageStructured evidence, not automated greenwashing.
- 1Read
Extract merchant, date, and product lines from the image.
- 2Resolve
Normalize each product and search for identifiable matches.
- 3Classify
Assign SUS only with concrete evidence; ambiguity remains NOS.
- 4Show
List every detected product with its SUS or NOS label and summarize SUS value only over the analyzed lines.
Conservative by defaultReduces false positives and reputational risk in a sensitive category.
Cumulative traceabilityEvery verified product can strengthen future readings without relaxing the criteria.
Controlled scaleThe catalog and rules can grow by market, category, and versioned source.
LUN is currently a SUS/NOS signal demonstrator. It does not replace environmental certification and does not yet provide a complete life-cycle assessment. That explicit limitation is part of its trust proposition.
09 / Real benchmark217 real images, 202 valid readings
The expanded benchmark processed 217 individual receipt images: 200 CORD v2 receipts and 17 traceable Wikimedia Commons images, including six reproducible crops from a licensed collage. The flow completed 202 readings and rejected 15 responses without manually repairing results.
199distinct valid transactions
4images with a SUS line
535detected NOS lines
2distinct SUS transactions
Three SUS images are views of the same Ekoplaza purchase containing ‘Het Blauwe Huis Fenegriek 60 g.’ The fourth is another real Lidl transaction whose line reads ‘Bio Haferdrink.’ Coverage therefore grew from one to two SUS transactions without presenting photographic variants as new purchases.
Main finding: 539 lines were resolved—4 SUS and 535 NOS—with 13 reproducible derivatives to respect the 8 MB limit. Sources: CORD v2 (CC BY 4.0) and Wikimedia Commons (explicit licenses or public domain). The SUS subset is enriched to test coverage, not to estimate the prevalence of sustainable consumption.
10 / PrivacyWhat happens to the image and history
To perform the reading, the image is transmitted in encrypted form to a selected multimodal AI provider. LUN does not store the original image by default; it retains only the structured result needed for internal MVP traceability.
- Upload only your own receipts or those you are authorized to analyze.
- Avoid documents with unnecessary personal data.
- There is no global public history of merchants, products, or amounts.
- The immediate result remains visible on the screen that performed the analysis.
Before enabling accounts or user retention, LUN must add authentication, owner-level authorization, deletion, and an explicit retention policy.
SUSReal · licensed
