01The goal
“I want to give it some money and see what it does with it in a month — focusing on gaining as much as possible, while keeping tax implications in mind.”
“Take in all information — including what the information is and the time it was received — and politician purchases as well.”
“I’m hoping to learn from this simulated experience. Make better choices, not faster choices.”
What the month will prove
That the machine works: data arrives, agents reason, limits hold, orders fill, the books balance, and nothing runs away overnight.What it can’t prove
That the strategy makes money. A month is a few dozen decisions at most — at that size, luck and skill look identical. Judging the strategy means replaying years of history, which comes later.02Seven rules the design follows
- The brain never touches the money.Agents only write proposals. A separate piece of ordinary, tested code — the gate — decides whether a proposal may become an order. Agents can be wrong or fooled; the gate still holds.
- Code does numbers, models do judgement.AI models are poor at arithmetic and good at weighing arguments. Anything that doesn’t need judgement isn’t an agent.
- Doing nothing is the default.A trade has to win an argument. Most nights should end with no trade.
- Buying is hard, selling is easy.Every check can block a buy. Almost nothing can block a sell.
- Two timestamps on everything.When it happened, and when we could have known. Agents only ever see what was knowable at that moment — otherwise any test quietly cheats.
- Everything is written down.Every memo, objection, vote, rejection and fill. If a trade can’t be explained six months later, the system isn’t finished.
- It explains itself in plain language.Learning is a goal, so the nightly journal is written for a beginner.
03The account and what it may trade
- Alpaca practice (“paper”) account. Free. The practice and real accounts use the same connection, so nothing gets rebuilt if this ever goes live.
- Cash only, and only the cash that’s in it. No borrowing, no options, no betting on stocks falling.
- $610 starting balance — exactly what would be funded for real, so the month’s behaviour carries over.
- Fractional shares, so $150 can buy a slice of an expensive stock.
- It trades over days, not minutes. After a sale, that cash is usable the next business day.
- Tax records from day one — purchase price, holding period, gain or loss per lot, and wash-sale flags.
About 40 symbols, on purpose
Broad market funds (SPY, QQQ, IWM, DIA), the 11 sector funds, the ~25 largest US companies — plus NANC and GOP, two funds that already copy congressional trades and serve as the benchmark. Reliable prices, deep news coverage, and nothing that can collapse overnight. The list grows only on evidence.
04The agent team
Runs every night before the market opens.
plain codeAI agenthuman
1
ScreenComputes the numbers for all 40 symbols: price moves, volume, new filings, upcoming earnings.
code2
ScoutPicks the 3–5 names worth a full debate. Everything already owned gets a “did anything change?” check.
agent3
Four analysts, each working aloneNone sees another’s work, and each only gets its own slice of the data.
agents- Fundamentals — the business itself
- Technical — how the price is behaving
- News & events — news, calendar, Fed tone
- Filings — congressional + insider trades
4
AuditorChecks every fact an analyst cited against the stored data. The defence against an AI making things up.
code + agent5
Red teamReads all four memos and writes the strongest possible case against the trade.
agent6
RebuttalThe analysts answer the red team. One round only.
agents7
Risk officerLooks at the whole account: how big, overlap with what’s held, free cash — and sets the exit plan.
agent8
Portfolio managerThe only one who can propose a trade — and must answer every red-team objection in writing.
agent9
The gateHard limits in tested code. No argument gets past it.
code10
SteveApproves on his phone. Three modes get tested: approve everything, approve above an amount, watch only.
human11
ExecutionSends the order and records it.
code12
JournalExplains the night in plain language. Weekly, a post-mortem agent scores what each agent said against what actually happened.
agentWho can block a trade
| Who | Power |
|---|---|
| Auditor | Hard veto. One cited fact that doesn’t check out kills the trade. |
| Analysts | A buy needs at least 2 of 4 in favour, and more for than against. |
| Red team | No veto, deliberately — an agent paid to object that can also block would stop everything. Instead every objection must be answered in writing; an unanswered one means no trade. Its track record is kept, and being right earns it weight. |
| Risk officer | Hard veto on buys, and can shrink any trade. Cannot block a sell. |
| The gate | Absolute. |
| Steve | Final say. |
The honest weakness: every agent is the same underlying AI model wearing a different hat, and copies of one model tend to agree. The defences: analysts are isolated, each sees different data, the fact-check is code, the red team is told to attack rather than assess — and the disagreement rate is measured. If nobody ever disagrees, it’s a rubber stamp and gets redesigned.
Shadow rules
The ledger also records what alternative rules would have done each night — a stricter 3-of-4 vote, a fixed price target, different limits — without acting on them. At month’s end they’re compared on identical days. Free learning.
The hardware
An old desktop: a 2015 Intel i7, 31 GB of memory, and a 6 GB graphics card. Far too small for fast AI — but this job runs overnight, so slow is fine. That’s the trick that makes it free. It uses the largest open model that fits, because the brief is better choices, not faster ones.
05Every position has an exit plan before it’s bought
The moment to decide when to get out is before you own it, while you’re still calm.
- Stop — the price where we admit we were wrong and sell.
- Trailing stop — for winners, the sell price follows the stock up, and triggers only when it falls back from its high. Winners can keep running; the exit is still a rule.
- Time limit — the idea has worked by this date, or it gets closed.
Code checks these daily. Hit one and the sale is automatic — no debate. Agents may tighten a plan but never loosen it; “it’ll come back” is exactly the mistake this exists to prevent.
The trade-off, understood going in: it trades on rules instead of emotions, and the price of that is sometimes selling a winner early.
06The limits
The brief: “whatever keeps the account in good standing.”
The broker’s rules — not ours to choose
- Never buy with cash that hasn’t settled from a sale
- Never order more than the settled cash available
- Cash only — no borrowing, no options, no short selling
Our own safety settings — starting values
| Setting | Start at | At $610 | What it prevents |
|---|---|---|---|
| Most in any one stock | 25% | ~$150 | one bad pick sinking the account |
| Cash always kept back | 10% | ~$60 | being stuck with no room to act |
| Sell-at-a-loss point | 4–12% below cost | ~$6–18 | a small loss becoming a big one |
| Idea deadline | 15 trading days | — | dead ideas tying up money |
| New buys per day | 2 | — | a burst of enthusiasm |
| Total orders per day | 6 | — | a software bug firing hundreds of orders |
| Daily pause | down 4% in a day | ~$25 | no more buys until tomorrow |
| Weekly pause | down 8% in a week | ~$50 | no more buys; human notified |
| Full stop | down 15% from start | at ~$520 | everything halts until restarted by hand |
Pauses block buying only — selling always works. These are sensible beginner-safe starting points, not precise science; the month shows which are too tight or too loose.
07The data — all free, official sources first
| What the agents need | Where it comes from |
|---|---|
| Prices, dividends, splits | Alpaca |
| News per stock, back to 2015 | Alpaca (Benzinga’s feed) |
| Second news feed, earnings calendar | Finnhub, free tier |
| Company financials, filings, insider trades | The SEC’s own data service |
| Congressional trades | House and Senate official records |
| Fed speeches and meeting dates | federalreserve.gov |
| Rates, inflation, jobs | FRED (St. Louis Fed) |
- The timing catch: members of Congress get up to 45 days to report a trade; company insiders get 2 business days. Only the publication date is something you can act on — the trade itself is long gone.
- Collectors switch on early, weeks before the month starts. Our own “when did we learn this” timestamps are the honest record, and can’t be bought or rebuilt later.
- No paid data source guarantees returns. If one did, its owner would trade on it rather than rent it out. Paid feeds buy speed and convenience — and speed is worthless to something that decides overnight.
08The build — 3 to 5 weeks, then the month
| Step | What gets built | Done when |
|---|---|---|
| 0 | Free signups, keys stored safely | keys work |
| 1 | Project set up on the server, in the nightly backup | skeleton backed up |
| 2 | Broker connection + ledger. No AI. | a hand-placed trade matches the broker exactly |
| 3 | The gate, with a test for every kind of bad order | every bad order is rejected |
| 4 | Data collectors + two-timestamp store (left running from here) | the store never shows the future |
| 5 | One agent, end to end | an AI-caused trade traceable from prompt to fill |
| 6 | The full team; model timing and quality test | runs unattended overnight; the red team blocks some trades |
| 7 | The month | 30 days, no runaway, a readable reason for every trade |
Steps 2–4 contain no AI at all — the boring, reliable part gets built and proven first. Afterwards: an honest review, then either a replay over years of history to measure real edge, a better model, or the real $610.
09Risks, kept in view
- Fooling ourselves with a small sample. The biggest one. A good month proves nothing; neither does a bad one.
- Poisoned text. News is written by strangers, and an article could try to instruct an agent. Everything agents read is treated as information, never instructions — and the gate is the backstop.
- Made-up facts. Why the auditor has a hard veto.
- Agreement theater. Why disagreement is measured.
- Quiet decay. A feed changes or a model updates and quality slips unnoticed — so it gets health checks.
- It can lose money. The honest purpose of version one is to find out, at zero financial risk, whether this has any edge at all. It may not.
10How we'll judge it
Set on September 21, 2026 — before any results exist, so the bar can't be moved afterwards.
Why not just a percentage target? In an ordinary month the whole market moves 4–5% for no reason anyone can name, and an account holding three or four stocks swings more. A month is a few dozen decisions at most — like judging a coin after ten flips. A target like “+3% means it works” would mostly measure luck.
- Safety — pass or fail. A month can prove this.No trade ever got around the safety gate. Our books matched the broker's every day. Every exit plan fired when it should have. The 15% full stop was never hit, and nothing ran away overnight.
- The number — against doing nothing clever.Compared with simply holding the whole market (SPY), the two funds that copy Congress (NANC, GOP), and cash. Goal: finish ahead of SPY. Tripwire: trailing SPY by more than 8 points means stop and look. In between is “can't tell yet” — not a win, not a loss. Short-term gains are taxed more heavily, so beating SPY needs a margin before it truly beats buying SPY and holding it.
- Decision quality — a month can measure this too.How often the fact-checker caught a bad claim. How often the red team blocked a trade. Whether the agents ever disagree. Whether average wins are bigger than average losses. Whether it got out when a trade's “what would prove me wrong” came true.
- The real test — after the month.Replay the same agents over several years of history: thousands of decisions instead of dozens. That is what separates skill from luck.