Momentic’s QA Bench repository reports Mo at 84.2% precision and 19.3% recall in its November 2, 2026 results table for the 60-minute setting. Those percentages describe different parts of the evaluation. Together, they help a team decide what to inspect before using an AI bug-finder in a release workflow.
Read the scoring rules first
QA Bench compares agent reports with 1,396 labeled bugs across 20 areas of four open-source applications. Its judge groups reports of the same failure and classifies them as correct, incorrect, unmatched or duplicate. Momentic builds Mo and maintains the benchmark; these are vendor-published results, not an independent test performed for this article.
The repository defines precision as correct findings divided by correct plus incorrect findings. Unmatched findings do not affect that ratio. Recall divides correct findings by labeled bugs. An unmatched report may describe a new bug, but its classification does not establish that the report is valid.
Work through a small example
Imagine a fictional test area with 100 labeled bugs. An agent produces 30 correct findings, 10 incorrect findings and five unmatched findings, with no duplicates. Under these definitions, precision is 30 divided by 40, or 75%. Recall is 30 divided by 100, or 30%. The five unmatched findings need separate review; they do not turn the precision denominator into 45.
Keep your pilot results reviewable
For your own staging application, choose one small area and agree on the expected behavior before running a bug-finder. Record each distinct finding, its reproduction steps and the reviewer’s classification. Keep unresolved findings visible instead of treating them as confirmed defects or silently discarding them.
Use the same application version, task scope and time budget when comparing runs. If your own app has no complete list of known defects, report how many findings reviewers confirmed; do not call that a measured recall rate. A tool can produce credible reports while still missing defects. No agents, benchmark fixtures or applications were run for this article.