wtf is franklin_conviction?

Franklin Conviction Engine is a tool that assembles a simulated investment committee of AI agents to analyze a cryptocurrency token and produce a final buy/sell recommendation with suggested position sizing. Instead of a single AI answering a question, it runs seven specialized agents in a structured pipeline designed to surface disagreement and force synthesis. When you enter a token, four analyst agents run in parallel: a Bull analyst (arguing the case for buying), a Bear analyst (arguing against), a Macro analyst (looking at broader market conditions), and a Narrative analyst (examining sentiment and story). A disagreement gate then checks the gap between the Bull and Bear scores, if that gap exceeds 30 points, a premium Arbitrator agent is brought in to adjudicate. Otherwise the Arbitrator is skipped, keeping costs down. A Synthesis agent then merges all the inputs, and finally a Risk Officer issues a conviction score, entry zone, stop-loss level, position size, and time horizon. Live market prices come from BlockRun's PriceClient, which is backed by Pyth, a real-time price feed. Agents stream their outputs back to the browser as they complete, using Server-Sent Events. The tool has two modes: FREE mode, which routes all calls through zero-cost NVIDIA models and requires no API key, and SMART mode, which uses Franklin wallet routing profiles including the premium Arbitrator when needed. Each session uses a wallet key you supply yourself, nothing is stored server-side. Built with Express and TypeScript on the backend and React with Vite on the frontend. MIT licensed. The README includes a disclaimer that outputs are research scaffolding, not financial advice.