Spinning the strand
Loading
Spinning the strand
Loading
Issue 01 · Vision · Conformal prediction · Grounded retrieval
Photograph a dish. Get the category, an honestly calibrated confidence, the full set of candidates the model cannot rule out, and nutrition traced to the USDA record it came from.
Food-101 test split
97.16%
top-1 accuracy
SigLIP-SO400M and EVA-02-L, averaged. Validation was carved out of the training split so all 25,250 test images stayed sealed until the end.
99.56%
conformal coverage
measured · average 1.54 candidates
101
dish categories
each with a USDA-grounded profile
32
nutrients per dish
macros, minerals, vitamins
693
retrieval documents
577 written + 116 graph facts
98.6%
RAG answers correct
76-case gold set, 100% grounded
01
Three commitments, each of which cost something to keep.
SigLIP-SO400M and EVA-02-L each score the image. A parameter-free probability average beat a 3.97M-parameter learned fusion head — significantly, by exact McNemar test.
See the ablation →Raw softmax is not a probability. Temperature fitted on held-out data cut expected calibration error eightfold, so 80% now means right about 80% of the time.
See the reliability curve →Questions outside the knowledge base are declined in under 130ms without calling a language model, and no answer states a number that is not in a cited source.
See the grounding gate →02
Sixty dishes had no single USDA record, so each was composed from weighted ingredient records. Those weights became a graph — and the graph answers questions no document contains.
Every edge carries the FoodData Central record it came from, so an answer built on this graph points at a source rather than at a plausible sentence.
101 categories, 101 grounded nutrition profiles, and a model that tells you when it is not sure.
Start