# WEVAL ARENA — Brain v2 | 12 avril 2026 | Wave 114

## IDENTITÉ
WEVAL Arena = moteur IA universel de WEVAL Consulting
176 options | 17 catégories | 394 agents | 385 skills | 328 écrans | 4 serveurs | 0€

## PROVIDERS ACTIFS (21/26)
SOUVERAIN: Master(310 intents), Public(chatbot), SearXNG, Qdrant(16K), Ollama(qwen3+gemma4)
API FREE ILLIMITÉ: NIM(Llama+DS-R1 40rpm), Mistral(EU), Alibaba(Qwen Turbo)
API FREE LIMITÉ: OpenRouter(GPT-OSS+Nemotron 50/j), CF(R1+Llama 10K/j)
WEBCHAT FREE: HuggingChat(121 Playwright), Copilot(GPT-4o), Meta AI, DeepSeek Web
CLÉS EXPIRÉES: Groq, Cerebras, GitHub PAT, HF Inference → Blade auto-renew queued

## SKILLS (17/18 EXEC)
PDF Excel PPTX Image TTS Video CodeAgent Debug CI/CD Docker Webhook Browser PR Lint SQL Scrape Cron MultiAgent

## ENDPOINTS
multi-provider.php(33KB 70+routes) | arena-engine.php(consensus+autonomous+stream)
arena-health.php(🟢/🔴 live) | arena-multiagent.php(parallel exec)
arena-autowire.php(discover+renew+scale) | arena-budget.php(cost tracking)
master-api.php(310 intents) | deepseek-proxy.php(DS Web+Gemini cascade)
filegen.php(PDF/Excel/PPTX) | code-agent.php | anthropic.php(sovereign GW)

## INFRA
S204: 158 pages, 108 APIs, 77 Python, 57 crons, 17 Docker, 59 ports
S95: 170 Arsenal screens, 6.65M contacts, PMTA+KumoMTA+Postfix
S151: OVH tracking | Blade: Windows desktop agent, Chrome, PowerShell

## DOCTRINE
Souverain-first | Zéro régression | 2 Strikes | GOLD backup
Validation Yacine | Playwright test | Tout via Master | Push synced

## ANTI-REGRESSION RULES (47 pièges)
- SED/HEREDOC = INTERDIT pour fichiers complexes → Python str_replace
- CX Proxy max 4500 bytes → split base64 chunks
- HTML-Guardian écrase fixes toutes 10 min → update GOLD après fix
- PMTA 0.0.0.0 ne marche pas → IP privée obligatoire
- Authentik SUPPRIMÉ → Auth PHP souverain
- Crontab JAMAIS echo | crontab - → (crontab -l; echo) | crontab -
- php -l OBLIGATOIRE après chaque modif PHP
- 2 Strikes Rule → STOP, root cause, fix structurel
- FAQ complète: /faq-anti-regression.html
- KB condensée: /api/arena-antireg-kb.md


## 9 COUCHES INTELLIGENCE IA SOUVERAINE
1. **Brain Engines** (111 fonctions): Nucleus v3 (23fn), Cognitive Brain (20fn), Opus46 (20fn), Opus46 Advanced (20fn), Expansion (15 domaines), GPU Rotation (13fn)
2. **Agents** (11): Master 310 intents, Autonomous SSE, Director 39obs, L99, WeDroid, Chatbot 134KB, Blade 50+, MiroFish, DeerFlow, Multi-Agent, Paperclip
3. **Providers** (9+): Ollama→Groq→Cerebras→NIM→Mistral→Alibaba→OR→CF→Gemini
4. **KB** (32 tables): 4500+ entrées, 40+ catégories, auto-enrichment */6h
5. **RAG** (Qdrant): 4 collections, 15953 vectors, hybrid dense+BM25+Graph RAG
6. **Tools** (15): Code, Math, Image, Mermaid, PDF, PPTX, Vision, TTS, Search, SSH, Planner
7. **Prompts**: Dynamic system prompt per intent, 8 reasoning techniques (CoT, ToT, dialectical, causal, counterfactual), sanitizer anti-leak
8. **WEVAL Mind**: Cycle Perception→Diagnostic→Planning→Execution→Learning, 21 scripts auto
9. **Nomenclature**: HAMID→WEVAL MIND (documenté, pas appliqué en prod)

## FICHIERS CRITIQUES
- brain-nucleus-clean.php: 1334L, 23fn, cœur souverain
- cognitive-brain.php: 20fn, routing+guard
- cognitive-opus46.php: 20fn, raisonnement avancé
- cognitive-opus46-advanced.php: 20fn, self-correction
- weval-chatbot-api.php: 134KB, 751fn MONOLITHE
- wevia-wave114-intents.php: 710L, 310 intents
- wevia-multi-provider.php: 33KB, 70+ routes