
SecondOrder is an autonomous fund built solo in seven days on a premise: the market prices the headline, but it prices what the headline depends on much more slowly. Instead of buying the obvious names behind near-predictable futures like AI compute, quantum computing, and munitions restocking, it walks the dependency graph outward, who supplies the builders, who supplies them, and trades the exposure with defined-risk options structures. An agent hierarchy works the graph like a real desk: a scout proposes theses (all held until human Board approval, enforced in code), junior analysts walk one hop outward with fetched evidence, senior analysts try to kill each candidate, a sector lead proposes structures, a CIO sizes the slate. The risk desk is not an agent: it is deterministic Python with 307 tests, measuring max loss from the actual payoff curve, and nothing reaches the broker except through it. Exits run on a pure-Python guard every 30 minutes with no LLM in the safety path. Cycles fired on a scheduler all week via Alpaca's MCP server, whether anyone was home or not. What it found: of 55 second-hop suppliers identified, only 15 had listed equity and one cleared the liquidity screen, the deeper the walk, the less of the chain is public. Tradeability is option volume, not size: NVIDIA quotes 3.2% wide, Lockheed Martin 24.9%. The venue only fills marketable orders, proven with a one-lot experiment that filled 63 milliseconds after pricing at the natural, after sitting hours at the mid. The fund also refuted the evidence behind its own biggest position the night before deploying it, cut the size, and revalidated, every step in the journal. Every number on the dashboard is re-derived from broker data by one command; nothing is hand-typed. It finished the judged week green, with entry friction paid, measured, and earned back. Public repo, MIT, live dashboard, full decision journal.
4 Sep 2026