An invoice-extraction tool fails schema validation about 8% of the time — usually because line items don't sum to the stated total, or because a required field is missing. Your current handling: when validation fails, re-run the same prompt against the same document. That recovers about 60% of failed cases on the second try.
What's the highest-leverage improvement?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Retry-with-error-feedback dramatically outperforms blind retry. When the second prompt includes the failed output and the specific reason it failed, the model can address a concrete error rather than re-attempting blind. "Your previous extraction returned total: 247 but the line items sum to 232" routinely recovers 90%+ of failures versus the ~60% you're getting now.
Stronger models help at the margin, but the fundamental issue here isn't capability — it's that the second attempt has no information the first didn't. A weaker model with good feedback often beats a stronger model running blind.
Three blind extractions in parallel tend to fail in similar ways, especially on structurally hard documents. Running three at 92% success each doesn't aggregate to "near 100%" — it produces the same kinds of errors three times.
Human review is appropriate as a last resort, not as the first response to a failure mode you haven't yet tried to fix programmatically. Burning a reviewer on every failed extraction when retry-with-feedback would catch most of them is a waste of reviewer capacity.