A new version of the system was deployed today! What you should see is a good deal more suggested corrections where the system did not flag errors before.

If you're interested in technical details: under the covers, our algorithm for finding good correction suggestions has changed substantially. We used to have some manually created heuristics that would combine evidence from multiple statistical models to determine whether a suggested correction should be shown. Now all these heuristics are gone, replaced by another model that was trained on error-annotated data from a large corpus. With this new setup we could nearly double recall in some errors, at the same (or even improved) precision.