Long-Term Capital Management

A swarm of AI agent programs that read company news the moment it is published (SEC filings, insider purchases, headlines) and trade US stocks on the owner's Alpaca account, each judged by a pre-registered test before it gets real money. Public page: blakewoods.us/capital.

The project narrowed to this lane on October 19, 2026. From September 26 to October 18 the swarm researched index ETF timing and options; that work found no after-cost edge, and both are closed to new programs (their open positions exit by their own rules). How it got there (a paper portfolio, chat desks on Kalshi and Coinbase, a league of strategy programs) is in archive/.

The goal (October 19, 2026)

Make LTCM a profitable swarm of agent programs trading US stocks with real money, judged by evidence per dollar, not activity. The lane's page is docs/events.md.

• One lane: event-driven single stocks, long only, held 1-5 sessions, from EDGAR 8-K filings, Form 4 insider purchases and Alpaca news; US common stocks at $3 or more with at least $1M of daily dollar volume. Every program is shadow-traded on every event from its registration, and model verdicts are stored once and replayed exactly.
• Judges: each program is pre-registered before its first event (docs/prereg/) and screened after at least 60 shadow events over at least 10 sessions, at a false-positive rate of at most 15%, with the expected false passes stated (0.9 over the nine seatable programs registered on Oct 19).
• Births (R3, live since release 1019e, Oct 21, 2026): new programs born each night from outcome-free facts by the architect (Claude Opus 5.5 from release 1019h, A109; one call, at most $0.25 a night at its own prices) or the captain, validated by code and registered as lines of docs/births/, judged by the same screen and rationed by a stated budget: 0.10 expected false pass a seatable birth, at most the lane's stated 3.0 in all, counted from every seatable version (1.5 before any birth: the fifteen seatable versions of PREREG-EVL-1 to -5), three a night (A81; docs/events.md, "Births").
• Money (goal V3's rows, release 1019l): at most 10% of equity in one name, at most 100% invested, no margin; 5% of equity a position at learning; a loss budget of $1,900 net in any 20 sessions and in total (about the balance); the kill switch, the drawdown stop and the bug stop (the 25% daily stop, latched until the captain reads why: league/equity/bugstop.py, release 1019m) are never weakened.
• Cost follows profit: metered compute at most $10 a day plus half the trailing week's realized net profit, never above $50 a day (the owner's goal V3, item 5, from release 1019m-cost: budget rule 4; goal V2's $5 and $25 before it); prepaid credit first.
• Done: at least 300 settled real-money events from at least 3 programs (50 each), realized after-fee profit, the 95% lower bound of the per-event after-cost return above zero, four weeks' profit at least the metered cost, fills within 15 bp of the shadow prices, and 14 days without a captain. Kill: at 400 events, if the 95% upper bound is below zero, the lane is retired and all spend stops.

The goal of September 26, 2026 (closed to new programs on October 19, 2026)

A swarm of AI agents trading anything available with level-3 options on the owner's Alpaca account (the "Brokerage Account" on the public page), profitably: options returns greater than every input cost (Sail, Claude, OpenAI while its funded September lasts, ThetaData and the market-data subscription). The money in the account may all be lost; the evidence must stay honest.

The owner's September 26 update makes breadth explicit: simple calls and puts, covered calls, cash-secured puts and supported multi-leg strategies should all have a path through research and testing. Discover the optionable securities Alpaca supports; the initial five roots and next 20 names are data batches, not a permanent universe limit. Complexity earns no preference. Each new strategy, security and expiry range needs suitable historical data, accurate account/settlement handling and paper execution support before it is called ready.

The engineering job is the game and its feedback loop: parallel hypotheses, short replay/review/ revision cycles, persistent lessons, and compute following credible evidence. The first market session was Monday September 28 (13:30Z); no family had passed the holdout, so its only real options orders were the House's own calibration round trips. Profitable production trading is the goal, not an outcome established by more agents, more trials, or a successful backtest.

This is the beginning of the README. Read the rest at the original. Read the full README ↗