How accurate is Voco?

Every number on this page is the average of three full runs of the whole test, back to back, on the exact parser that ships, release 5.2. No single run failed a bar. A single run is a screenshot, not a measurement.

How to read this

• Signed error is direction. Read it first. Misses that cancel out cost you nothing over a week. Misses that always run low silently break a cut.
• Calorie error is size. A 7% miss on a 700-calorie meal is about 50 calories.
• Hard-answer vs range-scored. Hard-answer cases check against a published number. Range-scored cases have no published truth, so they score against a band and look better by construction. We lead with the hard-answer figure.
• ✓ Verified is separate from all of it. It only asks: when we stamped the checkmark, was it right?

Where we land, food by food

The test: 200 written food logs, phrased the way people talk, scored against sources we cite per case. USDA, current manufacturer labels, and chains’ own published nutrition. No consumer aggregators.

Chain food is our worst number: 15.08%. Order a burrito bowl without weighing it and expect the estimate to be off by roughly a sixth. It is also the category the verified match layer exists to intercept: when a named chain item matches a curated record, the estimate is replaced by the published numbers.

The gap between weighed and unweighed is the argument for owning a food scale. Saying “about five ounces” instead of weighing costs roughly eight points of accuracy.

One more defence against the quiet undercount: the cooking oil you did not mention still gets counted 87.9% of the time (same 200 cases, 2026-08-09). That misses about one in eight, and we say so.

The all-200 figure, 8.04%, is lower than the 9.67% headline only because 41 of the 200 cases are range-scored. We publish both and lead with the higher one.

The 1.84% we will not brag about

Vague inputs like “a plate of stir fry” score 1.84% (30 cases, 2026-08-09), the best number on this page, and it is not a precision result. Nobody publishes what was on your plate, so those cases are scored against a range built from our own written portion assumptions, and the app runs on the same assumptions, so most answers land inside the band by construction. It measures consistency, not truth. Compare us on 9.67%, not on 1.84%.

When we stamp ✓ Verified

✓ means the item’s numbers came from a curated record, a manufacturer label, USDA, or the chain’s own published nutrition, and replaced the estimate outright. You also have to have stated a measure. Most items do not earn it, by design.

The traps are the near-misses that have actually bitten us: a sports drink versus its zero-sugar twin, a bar versus its mini, a food that is not in the database at all but looks like one that is. Zero false badges across 295 traps, in each of three runs. With 295 traps, the strongest honest claim is a true false-badge rate below about 1.01%, at 95% confidence. We will not print a bare “100%”.

Behind the badge: 1,687 curated records, and a matcher that declines to guess about 85% of the time on purpose. A missed badge costs a moment. A wrong one costs the reason to trust all the others.

A sealed set of 56 further traps has never been run. It runs exactly once, at App Store launch, and gets published whatever it says. That is the only way to know we have not been tuning against our own answer key.

What this page does not measure

• We wrote our own test. No third party has verified these numbers. Every case cites its source, the pass bars were set before the parser was chosen, and the misses are published. Until someone re-runs it, treat these as our numbers, honestly reported.
• The test types the meal, it does not speak it. Speech is tested separately: 0 material errors on 17 reference clips, three runs (2026-08-01). Seventeen clips, one speaker, one microphone. Small, and stated as small.
• It measures the parser alone. The verified match layer is switched out for these runs, so real-app accuracy on covered foods should be better. That is an expectation, not a measurement.
• 200 cases is small. One case can move a category by points. That is why we run three times and average.
• The badge tests use hand-labeled inputs, and it has gone wrong once. In one early development run the matcher took a trap it should have refused. It has not recurred. A repeat is a hard blocker, not a number to average away.
• Everything here is offline except the ~30% coverage figure.

Version history

One row per release, each from a fresh run. No row is ever deleted.

2026-05-30

Parser locked after a six-way comparison. 8.84% · +2.30%

2026-07-25

Per-item confidence added; deterministic settings pinned. The measured +4.45% broke the original ±3% bar, so the bar was widened. 8.4% · +4.45%

2026-07-31

Test set rebuilt from source; roughly 40% of the old answers were wrong and were corrected. The ruler changed, not the app. 7.7% · +1.66%

2026-08-01

A revised parser tested and not adopted. 8.9%, unshipped

2026-08-02

Shipping parser re-measured. 7.32% · +1.59%

2026-08-09

This page. Release 5.2: example foods our parsing instructions shared with the test set were removed, and everything re-measured. The trap set grew from 29 to 295 cases. 9.67% hard-answer · 8.04% all · +2.11%

Want the long version? Every case carries its source ID and retrieval date, and we intend to open-source the harness and the answer key so anyone can re-run this. Until then, ask us anything: hello@voco.fit.