How the loop works

Does any way of picking trades beat random entries, after costs? Most trading results cannot answer that: a market that rises makes every long look clever. So the loop is built to answer it and nothing else.

The desk is one paper account, started on 2 Nov 2026, 03:10 UTC with $3,750 of paper money. Claude models scan the crypto futures market all day, shortlist ideas and decide the entries; code sizes them, sets the exits and enforces the risk rules. Orders fill against live public prices from Bybit, with fees, slippage and funding charged. No real money is involved anywhere.

Every idea the desk takes is filed under one method before it is placed:

• Momentum. Trades with the move: breakouts, breakdowns, pullbacks inside a trend and retests of a broken level.
• Fade. Trades against a move that looks exhausted, or from the edge of a range back into it.
• Rule. Fixed geometry written in code: the models do not choose its entry, stop or target prices.

The methods share one market scanner, one decider and one market, so they are not fully independent. That is a known limit, not a hidden one.

For each real order, a separate service places two random paper trades on a hidden account: one on the same side, one on the opposite side, with the same size, the same stop distance and the same time. The twins never count in the desk's result; they exist only to be compared with.

A method's score is its excess over random: its net R minus the average of its two twins. The long and the short twin cancel whatever the market did on its own, so a rising market does not make a method look good. An order that never filled scores 0, so the score is per idea placed, not only per trade that happened.

Ideas of one method on one side within two hours are counted as one cluster: a burst of similar ideas is one observation, not ten. The range of each score is computed over clusters.

Each method has a weight between 0.5 and 1.5, starting at 1.0. It sets how much the desk pays for one idea of that method, never how many ideas it takes. Weights are recomputed once a day, move only after a method has 40 clusters and a clear lead or lag, and change by at most 0.05 a day. The decider never sees the weights or the scores, so it cannot chase them.

How an order is filled can matter as much as the idea. So each idea's fill style is drawn by a hash of the idea, not chosen: as asked half of the time (the order type the desk chose (a resting limit, a stop entry or market)), market a quarter (filled at once at the market price, whatever the desk asked for) and chase a quarter (a limit order that follows the best price until it fills). The split is published, but it does not decide any weight.

A second paper account of $1,250 buys BTC and ETH in small amounts on a fixed schedule, with no model involved. It measures one thing: whether the market itself went up while the desk traded. It is never part of the desk's result.

The run is reviewed once, on 14 Dec 2026, six weeks after the start. A method succeeds only if all four hold:

1. the 98.3 % range of its excess over random is above zero;
2. the excess itself is at least +0.10R per idea;
3. its own net R is at least zero, so it made money, not just lost less than chance;
4. the excess has the same sign in the first and the second half of the run.

A method needs about 230 clusters by then; one with fewer is "inconclusive". If no method succeeds, the answer is no method beats random, the desk stops paying a model for entry decisions and the work goes back to research. That is a result too, and the most likely one.

The run stops early only if the random twins miss more than one order in ten, if the twins turn out biased, or if the desk loses 30 % of its money.

Bugs are fixed when found. A change to how one method works restarts that method's count (its version goes up) and leaves the others alone. Changes that touch all methods alike are batched, at most once a week, and recorded with their date.

• Fills are simulated: a resting order fills when the price touches it, with fixed slippage. Real fills can be worse.
• Six weeks is one or two market moods. A method that works in one may fail in the next.
• The methods share a scanner, a decider and the market's direction, so their results are related.

Published: the desk's account value, its daily result, every closed trade (coin, side, method, entry type, result in R, how long it was held and how it ended), every open trade's coin, side, method, age and result so far in R, a short feed of the desk's decisions (what it looked at, passed on, entered and closed), the methods' scores and weights, the fill split and the market check. As JSON for machines: /data/loop.json.

Not published: the entry prices, stops, targets and sizes of open trades, the models' prompts and reasoning, the desk's internal rules and thresholds, and the random twins' account as a trader. Open trades are shown so the experiment can be followed live; this is not advice and not a signal service, and by the time you read a line the price has moved.