Step-by-step answer
Core 90–120 min8 stepsBuild an ingest eval harness
An eval harness for lease ingest on a synthetic, pseudonymised gold set of 20 leases. Page-level unit tests catch the parser's defects; field outcomes give accuracy, precision, recall, invented-value rate and coverage with Wilson intervals, overall, per field and per stratum; and a paired bootstrap over leases says which per-field differences between two prompts are real. The explicit not-stated prompt cuts invented values from 42% to 0%, at the cost of 20% coverage.
Back to the exercise