🤓 Latest Submissions
Killswitch Capital
The model proposes; the gates dispose. A language model reading option chains, Greeks and headlines is good at noticing that several weak pieces of evidence point the same way. It is not good at applying a risk rule the four-hundredth time exactly as it did the first. So Claude gets the noticing; plain Python gets the discipline. Claude Opus 5 sees one brief per symbol — 82 sessions of price history, contracts pre-filtered to the delta band and expiry window, live Greeks, implied volatility and news — and returns a direction, a confidence and a rationale, schema-validated. That view must then survive fifteen deterministic gates: kill switch, market clock, position slots, sector concentration, cooldown, delta band, spread width, expiry window, directional balance, premium richness, theta decay burden, risk budget, buying power. Two invariants, asserted in code and proven by tests: a gate may reject or shrink but never enlarge or invent an order, and exits are never gated — a rule that can block an exit can trap a loss. Every failure path returns confidence 0.0, so a system whose reasoning broke does not trade. The Alpaca split is the architecture. Every read — chains, quotes, Greeks, bars, positions, account, clock, news — goes through the Alpaca MCP server. Every write goes through the Alpaca CLI as a subprocess. The MCP server also exposes order-placing tools; this agent never names them, so "the model cannot trade" is a capability never handed over rather than a rule we hope it follows. Unattended on a Ubuntu server every fifteen minutes since 1 September. Result: $102,361 against $100,000, up 2.36%. Realized +$3,280 over three closed round trips — NVDA +59.3%, IBIT +52.2%, GLD −27.2%. The journal records every symbol considered on every pass, with the model's reasoning and the full gate trace. An agent that trades twice a week is indistinguishable from a broken one until you can read the ninety decisions in between. 312 tests. Write-up: docs/WRITEUP.md
4 Sep 2026
