In the first beta cohort, just over half of logged items landed in the right colour category. The errors were not random: soiled food packaging went green far too often, and batteries went almost anywhere except red.
The instinct is to write stricter instructions. What actually moved the number was the 'why this matters' explainer attached to each item — one sentence on what contamination does to a whole bag, one on what a battery does to groundwater.
Accuracy is now at 78% and still climbing. It is the metric we watch hardest, because a reward paid on a badly sorted bag is worse than no reward at all — it buys us data we cannot trust.