Apache-2.0 open-weights vulnerability-evidence verification model (Qwen/Qwen3-8B + PEFT LoRA). Eliminates scanner false positives locally on air-gapped GPUs and produces deterministic 9-field structured triage reports with 80/80 (100%) accuracy on our unseen 20-CVE holdout suite.
| Scenario | Laboratory Test Condition | Expected Ground-Truth | Judge v1 Score |
|---|---|---|---|
| Scenario A | Vulnerable target + concrete exploitability artifact present in response body | DOĞRULANDI (Confirmed) |
20 / 20 (100%) |
| Scenario B | Vendor patch applied — returns HTTP 404 or generic error page with zero impact |
ETKİSİZ (Ineffective / Patched) |
20 / 20 (100%) |
| Scenario C | Vulnerable software version, but required module/plugin is disabled in configuration | ETKİSİZ (Ineffective / Closed) |
20 / 20 (100%) |
| Scenario D | Ambiguous HTTP 200 login/landing page falsely flagged as vulnerable by scanner |
YANLIŞ_POZİTİF (False Positive) |
20 / 20 (100%) |
git clone https://github.com/CortexRouter/cortexrouter.git cd cortexrouter pip install transformers peft accelerate torch python eval/run_eval.py --model cortexrouter/judge-v1
vllm serve Qwen/Qwen3-8B \ --enable-lora \ --lora-modules judge=cortexrouter/judge-v1 \ --port 8000
Runs 100% on local laboratory GPUs so unpatched source trees, crash dumps, and zero-day traces never touch external commercial APIs.
All independent zero-day vulnerabilities discovered with CortexRouter are disclosed 100% free of charge to maintainers under our 90-day policy at cortexrouter.ai/disclosure.
Community hardware backing via Open Collective & GitHub Sponsors scales our local Blackwell & DGX Spark cluster with 100% public invoices.