Choose an item on the left, or add a resume or job. The review screen shows the parser's answers next to the source text, so you only correct what's wrong.
Accuracy check
Parses every gold item again, compares each answer with its key, and checks whether repeated runs agree. Reviewed keys are graded; draft keys only count toward consistency unless you tick “Grade draft keys too”.
Earlier runs
Pairs: does Match++ land in the right band?
Pick a gold job and a gold resume and say which band a recruiter would expect. The accuracy check scores each pair once per run and reports whether the band is right, within one band, and how much the score moves between runs.
Review notes: questions and problems from the test bench
Each note is saved with the job, resume and Match++ it is about. A model reviews the open notes and recommends a fix; a person decides: Accepted, Fixed, Won’t fix or Closed. Cases worth keeping go into the gold set; the rest are cleaned up later.
Keeping the log tidy
The log is append-only: qa_data/notes/notes.jsonl (kept out of git). Closed notes untouched for 90 days move to notes_archive.jsonl, and uploaded files and runs older than 90 days that no note and no gold item needs are deleted: python -m maayu_parser.notes cleanup lists them; add --apply to do it.