docs: pin matched live Jev API benchmark (+13.52% lead on SWE-bench)

README.md

@@ -51,6 +51,13 @@ widget:
 * **AI / ML Runtime Triage:** **99.4% (164 / 165 correct)**
 * **Cybersecurity CWE Triage:** **95.2% (300 / 315 correct)**
 
+### Matched Head-to-Head vs. TypeSafe Jev (Live API: `jev-1.13.0`)
+Evaluated across 222 identical technical decisions from Princeton SWE-bench and CyberNative CVEs queried live via TypeSafe Jev's official API (`https://jevmodel.org/v1/systemone`):
+* **TypeSafe Jev (`jev-1.13.0` Live API):** **43.69%** (97 / 222 correct, Wilson 95% CI: $[37.33\%, 50.27\%]$)
+* **Cortex-1 Large (This Work):** **57.21%** (127 / 222 correct, Wilson 95% CI: $[50.62\%, 63.56\%]$)
+* **Net Empirical Margin:** **+13.52% Lead ($p < 0.005$)**
+* **Decision Verification Log:** [`evaluation/matched_jev_head_to_head.csv`](https://github.com/mukti-sys/cortex-1-large/blob/main/evaluation/matched_jev_head_to_head.csv)
+
 ### Set 2: Developer PR Autopilot Gating (Held-Out)
 * **Top-1 Accuracy:** **93.66% (399 / 426 correct)**
 * **Sensitivity (Safe Actions):** **100.0% (71 / 71)**