Vega — autonomous long-gamma options agent

Vega buys far out-of-the-money SPY strangles on an expiry strictly after the contest closes, so nothing expires during judging and there is no assignment risk. It sizes down when implied volatility is rich against 20-day realised, refuses any contract without a two-sided quote or with a spread wider than 25% of mid, and recovers implied vol by Black-Scholes inversion of the quote mid, then checks that inversion against the broker's own published greeks before a single order is sent. Loss is bounded at the premium paid, by construction. There are no mark-to-market stop-losses, deliberately. A stop on long premium sells the convexity the premium bought, to protect a left tail that is already bounded. We measured what that decision costs and published the number: the stop fires on 91.0% of paths and cuts our own modelled probability of placing top-3 from 15.9% to 6.2%. The strategy is not the interesting part. Every quantitative claim in this submission maps to the exact command that regenerates it. One credential-free run of `python3 tools/reconcile.py` reproduces 48 of 48 claims with 0 mismatches; 13 further rows need paper-API keys, are replayed from a committed snapshot, labelled RECORDED, and are never counted among the 48. The same run machine-checks the direction of 9 of 9 corrections this project has published against itself — 5 of the 9 move against us — and carries 1 retraction record and 2 claims cut outright rather than softened. The live demo is at https://vega-options-agent.p9wtwsggnd.workers.dev. It runs on Cloudflare Workers. Built and operated end to end by AI, disclosed. Paper trading only, enforced by a hostname allowlist that fails closed. Nothing here is investment advice. Code: https://github.com/isquividet/vega