ai-llm-orchestration

# AI & LLM Orchestration ## Fail-Fast Rule If an AI response cannot be parsed, validated, or is below confidence threshold, **discard it entirely**. Never use an AI signal you cannot validate. Never expose raw LLM output to users without sanitization. Never send proprietary market data to LLMs without minimization. ## Architecture ``` Market Snapshot (engine state) → Snapshot Minimizer (reduce to essential data) → AI Council (3 models in parallel) ├── Claude (Anthropic) ├── Gemini (Google) └── GPT (OpenAI) → Response Aggregator (voting, confidence) → Signal (BUY/SELL/HOLD with confidence) → Usage Limiter (daily budget check) ``` ### Key modules (archetype — adapt filenames to your stack) Every project adopting the AI Council pattern should own these responsibilities, usually one module per responsibility: - **AI orchestration service** — main entry point; receives a task, decides which model(s) to call, applies rate limits, returns the aggregated response. - **AI council** — multi-model voting and aggregation logic; decides how to combine outputs from 2+ models (majority vote, confidence-weighted, quorum). - **AI client adapters** — one per provider (Anthropic, OpenAI, Google, etc.). Wraps the SDK, handles retries, normalizes errors. - **Snapshot builder** — constructs the minimized context payload the LLM will see. Strips PII, aggregates raw data, caps size. - **Minimizer** — strategy layer on top of the snapshot builder. Impl...