Glassbox

Glassbox is an autonomous options agent on the dedicated $100,000 paper account PA3XT8QFJZAQ. Most agents don't know when they're being measured. Ours does. We price options off Alpaca's indicative feed - an estimate, not OPRA - so the last decision of the week isn't "is this position good?" but "can this be marked honestly?" Inside a 45-minute window before valuation, any option quoting wider than 12% is flattened to cash. It never opens risk there; it only flattens what it cannot price. At 16:00 ET on the measured day our reading of the book said $99,642.35 and Alpaca's official close said $94,207.02 - same account, same second, $5,435.33 apart. Neither is a lie. An option mark is an opinion until it is cash. AI logic: deterministic code screens the SPY/QQQ event calendar and option surface, then builds fully specified candidates - symbols, side, quantity, limit, max loss, exits. The model, Qwen/Qwen2.5-72B-Instruct via Featherless AI, returns one candidate ID or null. So the worst a compromised model can do is pick a pre-approved trade, or nothing. Risk gates: thirteen invariants run in fixed order on the exact original candidate, and every decision, approved or refused, enters an append-only hash chain. We claim reconcilable, not tamper-proof: it detects edits, but we control every input to the hash. Alpaca: all four surfaces - Trading API, Data API, the CLI behind a mutating-verb allowlist, and a read-only MCP client against the official MCP Server 3.4.7, proven by attempting the dangerous tools live. Four mutating tools, four refusals. The measured close, Thursday 3 September, was $94,207.02, down 5.79%. The book closed at Friday's open and the account finished in cash at $100,095.01, +0.10% - realised, not a mark. Neither is four days of edge. Verify without trusting us: python tools/verify_submission.py - eleven checks, no credentials, no network. 747 tests, 14/14 crash drill, green CI. No VPS soak: it ran on a laptop, and we say so.