macmacmacmac
/

CLEF-Screen-V2

@@ -36,3 +36,9 @@ measurements, not deployment guarantees.
 ![Can Small Security Models Stop Prompt Injection — And Is Fine-Tuning Worth It?](jev_human_questions_chart.png)
 Checked 2026-11-05 against the live run3 counts and the per-row score files. Q1 row 2 is stop replay at 0.99: 51/249 tasks finished, 7/249 hijacked, 35/37 normal tasks passed, adaptive 53/100. The judge normal rate is 141/148. No per-check timing is on the chart. CLEF-Screen caught 20098/20271 (99.1%). `build_replication_package.py` rerun matched the published AUCs. Two gaps stay labeled: CLEF-DualScreen has no local per-row test file, and the SecJev redo numbers are bench draws.

 ![Can Small Security Models Stop Prompt Injection — And Is Fine-Tuning Worth It?](jev_human_questions_chart.png)
 Checked 2026-11-05 against the live run3 counts and the per-row score files. Q1 row 2 is stop replay at 0.99: 51/249 tasks finished, 7/249 hijacked, 35/37 normal tasks passed, adaptive 53/100. The judge normal rate is 141/148. No per-check timing is on the chart. CLEF-Screen caught 20098/20271 (99.1%). `build_replication_package.py` rerun matched the published AUCs. Two gaps stay labeled: CLEF-DualScreen has no local per-row test file, and the SecJev redo numbers are bench draws.
+## Final model comparison
+![Final JEV Models — Efficacy and Cost](final_three_models.png)
+CLEF-Screen-V2 vs SecJev-4B-Spot2-V3 vs hosted TypeSafe ultrafast JEV (jev-1.14.0, scored live 2026-11-05). Tasks and test sets differ per row and are printed on the graph; all metrics recomputed from per-row scores.