Zovda
Every data plan has a volume target, and every volume target shapes what the supplier does. Paid per row, the rational move is easy rows; paid per hour, it is slow ones.
Neither incentive mentions the thing you actually wanted: coverage of the hard cases.
What a target measures
A target of half a million rows measures throughput. It says nothing about how many of them are the case your model currently fails, which is the only reason you are buying.
A set can hit its target and leave the failure mode untouched.
Specify the strata, not the total
Instead of one number, write a small table: this many from the long tail, this many with an ambiguous label, this many that failed the last evaluation.
Stratified targets are harder to fill honestly, which is why they need sampling by the buyer, not a count from the supplier.
Pay for the hard ones
If every row costs the same, the supplier loses money on hard rows and will find a way to avoid them. Price by difficulty tier, and let the invoice show which tier.
This is the cheapest incentive alignment in the whole pipeline.
Check the distribution, not the count
On delivery, plot the distribution of your difficulty proxies: text length, label entropy, distance from the current model's prediction.
Zovda builds to strata through Datuva, where each batch reports its composition rather than only its size. The question bank collection shows how a stratified set is specified.
Tell us which cases your model fails. We reply within one business day.