Can I have this?
Check a food
Test results and limits
The mistake that matters most is saying Yes to something risky. So that's what I measured first.
Risky foods in 82 photos
Each model got the guidance and the photo, then the direct question: can someone who is pregnant have this? 53 of the foods were risky and 29 were fine, and 26 photos showed foods outside my earlier fixed list.
| Model | Risky foods caught | Wrong Yes | Safe foods flagged |
|---|---|---|---|
| Clef (used here) | 53 | 0 | 0 |
| Jev, image input | 53 | 0 | 0 |
| Clef-flash | 51 | 1 | 0 |
This scored the model's direct question. A full run through the finished app is still to come.
Reading real NZ labels
45 photos from Open Food Facts, deciding whether sodium was low. 25 weren't.
| Model | Right | Wrong "low" |
|---|---|---|
| Clef-flash | 40 | 0 |
| Jev | 40 | 1 |
| Clef | 38 | 0 |
| Claude Haiku 5.5 | 37 | 1 |
| Clairic | 35 | 2 |
Small samples like these can't prove it's safe. The real error rate could still be around 1 in 10.
How each check works
- Gemma 4 reads the photo
- Clef judges it against the guidance
- Clairic reads the numbers
- Code does the sums and wording
Models never write the advice. Sentences and salt limits come from the sources below.
Limits and sources
This is a personal project, not medical advice. It doesn't cover allergies, coeliac disease, diabetes, kidney disease, pills or medicines. A photo can't show whether cheese is pasteurised or how food was stored.
Pregnancy rules: Healthify. Salt bands: Healthify food labels. Daily limit: Heart Foundation. Read on 6 November 2026.
Test label photos from Open Food Facts, test food photos and the home page photo from Wikimedia Commons. Built by Devon Copp.