Twenty-six letters have told you to write things down before you need them: a survival line, a flinch line, a kill fact, a trim ladder, a sale order. All of it rests on one claim: that a rule written in calm weather beats a decision made in a storm. It's a fair claim. It is also, so far, a claim I've asked you to take on faith.

Arc Four stops asking. Over the next weeks I'm going to show you something being built under those rules, in public, with the receipts attached, and let you judge whether the method holds up when money is on the line. The thing being built is a small trading desk. The trader is an AI. The rules are written down. And the first month, which closed last week, lost money.

I could have waited for a good month to start this arc. The better letter is the honest one, and the honest one starts with the audit I was handed on October 29. Here it is, with nothing taken out.

Since September 28, a small book on a regulated U.S. exchange, about $1,400 at the start, has been run by an AI acting as portfolio manager. Every rule it follows is in a manual that only changes on Sundays, in writing. Every trade is journaled. Every Sunday the journals are reconciled against the exchange statement to the cent. There are three parts: a slow, rule-based bitcoin position that is only long when price is above its 20-week average and is sized to a fixed volatility target; a small "challenge" account that was allowed to trade on its own judgment with strict risk caps; and a paper-only stock book that will go live in late November if its test month stays clean. No leverage worth the name. No withdrawals possible from the trading key. A kill switch at −20% that puts everything on paper for a month.

The point was never the money. $1,400 cannot make a living and this letter has said so every week since The Floor. The point was to find out, with real fees and real fills, whether a process built from these letters survives contact with the market. Call it the proof of concept for Arc Two.

Two of the three parts did exactly what they were supposed to. The bitcoin position is down because bitcoin fell from $86,000 to $77,000 during the month; a trend rule is down when the trend is down, and it will go to cash the first weekly close under its line (around $72,900) and buy back the first weekly close above it. That's not a failure. That's the rule being the rule. The stock book sat on its hands for two weeks because no ETF met its entry condition, which is also the rule being the rule.

The failure is concentrated in one place, and it's the place where the rules were loosest. The challenge account was allowed to pick its own trades, and it picked ten, and nine of them lost.

Here is the finding from the audit, and it's embarrassing in the useful way: the losses had almost nothing to do with being wrong about the market.

The account was about $800. The risk cap per trade was 1.5% of that, about $12. On an exchange where the smallest bitcoin contract is worth about $800, a $12 risk budget means the stop-loss has to sit within 1.5% of the entry. Bitcoin moves 2 to 3 percent on an ordinary day. So nine times out of ten the trade was knocked out by normal daily noise before the idea had a chance to be right or wrong. The stop wasn't set by the market. It was set by the size of the account. And the software, when a trade didn't fit, printed a message that said tighten the stop.

A trader with $100,000 running the identical ideas with honest 4% stops would have had a different month. Not necessarily a winning one, but a fair test. The small account didn't get a fair test; it got a coin flip rigged by its own arithmetic. We replayed all ten trades afterwards against the coins' own recent behavior, using only data that existed at each entry: every one of them had break-even odds before it was placed. Nobody checked.

The fix went in the same day. The stop now has to be at least one full day's average range from the entry, or the trade doesn't exist. Which means, at $800, the account can only trade four coins where that honest stop fits inside the risk budget. Bitcoin isn't one of them. That's a humbling sentence to write about a bitcoin publication's own desk, and it's correct.

Here is what I want you to take from this, and it's a tool like the others: fifteen minutes, numbers you already have.

Once a month, on a fixed day, write one page about your own position with four questions on it. Not how it did. Whether it was run.

One. Did I follow the written rules? Yes or no, per rule. If there is no written rule for something you did, that's a no.

Two. What did I change, and when? Every change to size, stop, allocation or plan, with a date. If a change happened on a bad day, flag it. Changes made in storms are the ones that cost.

Three. What did the test say, and what happened? Whatever you expected going in, from a backtest, a base rate, or just a written guess, next to the real result. The gap between them is the only real information you have about whether your method works. Most people never write down the expectation, so they can never measure the gap, so they never learn.

Four. What did it cost? Fees, spread, slippage, the actual numbers from the statement. Our desk assumed 0.6% per round trip and measured 1.8%. That one line killed a whole category of fast trading before it could lose real money, and it was found by spending $1.63 on two tiny test trades. Most people never measure it at all.

That's the whole page. The desk's version of it is what you just read, and it produced the one finding that mattered. Yours will too, usually something you half knew and had not been forced to look at.

First, the sample is tiny. Ten trades is noise; a one-win month can be followed by a six-win month with no change in skill either way. The finding about the stops isn't drawn from the win rate; it's drawn from the arithmetic, which is why I trust it. But anyone selling you conclusions from ten trades, including me, should be read with that in mind.

Second, maybe the lesson is that an AI shouldn't be trading at all. I don't think so, and the slow rule-based parts of the desk argue against it, but the discretionary account's first month is a legitimate data point for the skeptic. The desk's answer is to run twenty more trades under the fixed stop and publish those too. If they're also bad, the discretionary idea gets retired, in writing, and this arc will say so.

Third, publishing a loss might read as false modesty. It isn't meant that way. The purpose of the audit page is that it gets written whatever the number is. A publication that only shows you its audit in the good months is showing you marketing.

Fourth, none of this is a recommendation. The desk trades derivatives, with small leverage, in a sandbox sized so that total loss changes nothing about my life. The canon of this letter for you hasn't moved: one to ten percent, spot, self-custody, no leverage, entered on a schedule. Arc Four will show a system being built; it will not tell you to build this one.

The Blind don't keep one. The account did what it did, and the story they tell about why changes depending on who's asking.

The Scared keep one, read it in a bad month, and quit, which converts a measurable loss into a permanent one.

The Prepared write the page on the same day every month, in good months and bad, and treat a bad page the way an engineer treats a failed test: as the most useful thing that happened all month. Then they change one thing, in writing, and run it again. This letter is the desk's first page. There will be one every month in this arc, and you can hold me to the number.

Arc Four, letter one. The arc is called Build because that's what it's about: not a thesis to believe but a thing to make, piece by piece, where every piece has a written rule and a measured result.

If this one was useful, three things:

One, pick your audit day. The first Sunday of the month works; it's when I do mine. Put it in the calendar now, before you have a reason not to.

Two, write the first page this week, even if your position is just a slice you bought on a schedule. Four questions. You'll find a rule you never actually wrote down. Write it.

Three, reply with your measured cost, if you have ever measured it. Just the number. I suspect most replies will be "I haven't," and that will be a letter of its own.

Next Sunday: The Floor Plan, how the desk's three parts fit together, why slow money and fast money are kept in separate rooms with a wall between them, and the one-page design you can copy for a position of any size, including one with no fast money in it at all.

— Bill2Billion

P.S. — Not financial advice, and the desk's results are a record, not a recommendation: it is a small, deliberately expendable experiment, and the canon for readers stays one to ten percent, no leverage, no timing, ever. Every number here was taken from the desk's own reconciled journals and the October 29 audit; the live market numbers were re-verified this weekend: bitcoin near $77,000, three-month bills at 4.00 percent, core PCE at 3.1 percent for September, the federal funds range held at 3.75 to 4.00 percent on October 28. The audit page on my own position was written this morning. It had a no on it.