
A model takes about a minute to decide. One measured cycle ran 61 seconds across three reasoning rounds before an order could exist. Option quotes do not last that long. Proxima Quant is built around that gap. AI LOGIC. Every Alpaca capability is wrapped as a Python function inside a sandbox with no credentials and no network. The model writes a program and runs it. One live program: 130 lines, 10 Alpaca calls. It enumerates liquidity-gated structures across expiries, builds three independent return distributions (lognormal, an empirical bootstrap, and a fat-tailed Student-t), scores every candidate under all three, discards anything surviving only one, and stress-tests the resulting portfolio. It ended in decision.no_trade — 39 of 59 cycles did. A watcher runs continuously at ten-second cadence with no model involved. Orders need two separate programs: one stages an intent, a later one confirms it after the host re-prices from fresh quotes. RISK GATES. Twelve host gates run as pure functions generated code cannot relax, from paper-endpoint assertion and account identity through spread, economics, concentration and correlated portfolio scenario stress. Quantity is the smallest of five headrooms, the binding one named back to the model. A fresh-price check requires the weakest measure to clear the real round-trip spread cost by 1.5x. Exits are host-owned, including a 15:15 ET expiry-day liquidation. ALPACA. Execute mode routes every order through the official CLI, pinned and checksum-verified at v0.0.14. Multi-leg mleg limit orders with per-leg position intents. Four concurrent streams; caps measured at 30 equity, 200 option symbols. Greeks computed locally, validated within 0.5%. RELIABILITY. Requests are fsynced as PRE_SUBMIT before Alpaca is contacted; timeouts and duplicate-ID 422s become UNKNOWN and reconcile by client_order_id. One session: 59 cycles, 17 orders, 59 reconciliations, zero unresolved. 509 tests, including live-broker fault injection.
4 Sep 2026