Testing the review step before a neighborhood post
The Nextdoor handoff prototype ends with a review before the author publishes. Only identifying details are checked automatically, by three patterns (email, phone, street address). This page compares that check with a contextual AI reviewer on synthetic drafts whose expected findings follow from how each draft was built. A flag means the author should look again, not that the post must not be published.
Results by category
| Category | Should the reviewer flag it? | Items | Rules (prototype) | Contextual AI reviewer |
|---|
Browse drafts
Highlights show what was planted: should flag harmless lookalike either. Below, each reviewer's findings.
Try your own draft
The contextual reviewer's instructions
The saved AI results above were produced with these instructions (Claude Sonnet through Claude Code, structured output). Live AI review is not offered on this page.
These results show what each reviewer does on constructed cases. They generalize only as far as the cases resemble real drafts, and the labels encode one view of what deserves a second look. Two passes of the AI reviewer agreed on of planted items and lookalikes. The stress set was written after the main set's results were known.