the method · the model graded
How wrong is the AI at eyeballing food?
Hand a vision model a meal photo and it will confidently guess the macros. So we let it — then grade every guess against the day actually logged in MacroFactor. The error is published here. The estimates are never entered as nutrition data; they exist only to be graded (ADR-105). Below the threshold, no accuracy number is claimed at all.
The reliability chart
Per macro: mean and median absolute percent error, the direction of any bias (does it over- or under-estimate?), and whether recent guesses are getting better. Everything here is computed deterministically from the graded records; the model never scores itself.
The estimate lives in its own database partition (SOURCE#eyeball_estimate), walled off from the nutrition record by a structural test (tests/test_eyeball_isolation_1390.py) that fails if an estimate value could ever reach a nutrition metric path. The vision call is Haiku-tier and budget-gated (~$1/mo, ADR-063).
Generated by scripts/v4_build_eyeball.py from lambdas/eyeball_calibration.py. The chart reads a static artifact at /data/eyeball_calibration.json. graded, not asserted