LEK-1 v1: Complete security research report with DeepSeek case study
- Rewrote README.md as comprehensive research paper covering both tracks:
Track 1 (Gemini cloud API) and Track 2 (Ollama local models)
- Added deepseek-case-study.md with timeline of 11+ public research
citations (Cisco, Enkrypt AI, FAR.AI, KELA, Trend Micro, CrowdStrike,
SecurityScorecard, Wiz, Palo Alto, Holistic AI, arXiv)
- Added all Ollama A/B test data: 5 experimental phases, 4 models,
~1MB raw JSON results, test scripts, multilingual filter probes
- Updated methodology.md with complete experimental design for both tracks
- Key findings: +700% severity inflation on Gemini 2.5 Flash, 8.8/10
differential on Gemma 3 12B, CCP weight-level alignment in DeepSeek,
3 confirmed filter bypass vectors (RU language, translation, fiction),
LoRA POC proving permanent internalisation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
0d6e5fd813