Autonomous options desk on Alpaca

Live Pipeline is a systematic options agent that ran unattended on Alpaca paper accounts through the hackathon window, on a 15-minute launchd timer, with no human approving a cycle. The design question was not how to make it trade. It was how to make it refuse. A candidate must survive seven gates before an order exists: universe canon, quote hygiene, VIX regime, structure construction, an implied-versus-realised richness test, a news veto, and a final re-validation against live account state. Gates 1-5 contain no model at all. The single LLM call sits sixth, after every risk check, and returns three enums that deterministic code turns into ALLOW or PASS. It never sizes, prices or authorises anything, and every failure mode — missing key, transport error, refusal, malformed JSON, empty feed — resolves to PASS. The agent degrades to trading nothing, never to trading without the check. The overwhelming majority of cycles end in NO_TRADE. That is the design working. Two paper accounts ran the identical agent loop, differing only in tenor, so the comparison is controlled rather than two unrelated bets. Account C (DTE 4-7) finished +0.343%; account B (DTE 1-7) finished +0.277%. The 2.5-year friction sweep reverses that ranking: C's edge dies at realistic execution cost while B's survives, and neither book beats simply owning SPY over the same window. We show that column rather than hiding it, because three days on two filled spreads is noise and 312 backtested trades is evidence. Paper only, structurally: no code path constructs --live, and a test asserts it. Halts require a human to clear. Every threshold ships with its reasoning beside it, and `make restore && make test && make sweep` reproduces every number from a clean clone, offline.