What it is
Shows how to fix a detector's blind spot by writing one sentence.
What it does: Asks Jev several English questions about a text and weighs the answers into an attack probability.
Why it belongs in the showcase: Honest comparison against a single question and four other detectors.
Caveats: Ties, not beats, the single-question baseline.
Evidence
Public MIT repo, created 2026-10-21, with a live demo. On a 2,193-text confirmation set it reaches AUC 0.979, and 0.983 with one added sentence; the author's paired bootstrap finds that version tied with a single Jev question (0.983). Adding one sentence as a neuron, with no retraining, raised password-extraction catches from 40.8% to 70.4% at a cost of 13 new false alarms in 2,731 benign texts. One Jev call per text: median 79 ms, 2.0e-5 USD.
Caveats
- Does not beat one good Jev question on unseen data; the gain is editability and explanations
- Questions were found with Claude in a search loop
- Training and test sets are public prompt-injection datasets
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.