当社のライブボットは5.31ドル稼いだ。取引所のダッシュボードは15.43ドルと表示。

エンジニアリングフィールドノート。 On 27 July 2026 a deterministic ladder took over a MEXC futures wallet holding 103.44 USDT and began trading the SOL_USDT perpetual with real money. Twenty-three days later the exchange dashboard reports an account equity of 113.33 USDT and a 30-day P&L of +15.43 USDT. Neither of those numbers belongs to the bot. This note is the reconciliation.

What the machine actually does

The strategy is a long-only ladder, and it has no model in it at all. It opens a first rung of $5 margin at 3x — about $15 of notional, or 0.2 SOL. It takes profit when price is 0.8% above the average entry of the whole ladder. If price instead falls 0.5% below the ladder's running low, it adds another rung. That is the entire decision surface.

In 23 days it logged 28,365 decisions: 27,455 holds (96.8%), 863 add attempts, 26 opens and 21 take-profits. There is no language model anywhere on this path. That is an architectural rule in this codebase, not a preference — the only function permitted to reach the exchange is a single order-submission call, and a test asserts it at the syntax-tree level.

The exchange's number is not our number

MEXC's own P&L analysis page, filtered to SOLUSDT Perpetual, shows +15.4291 USDT over 30 days and an account equity of 113.2702 USDT. Both are true and neither is the bot's result. Pulling the account's closed-position history and transfer records explains the whole gap:

The rest of the exchange's headline is history that predates the bot: manual trades on 21 July (+$6.75) and 26 July (+$3.44) sit inside the same 30-day window. Stretch the view to 90 days and the same account shows only +10.52, because manual trading on 7 June lost $4.91. None of that is the ladder, in either direction.

This is why the ledger keeps its own attributed curve instead of reading the account balance. The account balance answers "how much money is there". It cannot answer "what did the machine do".

A 100% win rate is a design artifact

Twenty-one trades, twenty-one winners. That statistic is worth nothing, and the reason is in the rules: the ladder only closes at +0.8%. A rung that moves the wrong way is never sold — it is averaged. So a losing trade cannot appear in the record until the strategy is forced to capitulate, and in 23 days it never was.

The averages say the same thing more quietly. Mean adverse excursion across the 21 trades: −3.01% of the margin posted; worst −9.94%. That is the ledger's unit — a 3x position moves against its margin three times as fast as the price does, so −9.94% of margin is a −3.31% move in SOL. At full extension on 1 August the ladder held 21 contracts — 2.1 SOL, about $153 of notional on a $103 wallet — with $51.08, or 49.4% of the account, posted as margin and $4.60 of open loss. The attributed equity curve fell from 104.51 to 99.57, a 4.73% peak-to-trough drawdown. Read that against the 5.14% the whole experiment earned: the account risked nearly as much as it made.

The day the ladder ran out of money

On 2 August at 18:52 UTC the operator moved 95.41 USDT out of the futures wallet — while the ladder was extended. From 3 August 01:27 to 5 August 01:04 the strategy asked for a new rung and the exchange answered Balance insufficient. It did so 811 times.

Nothing was borrowed, no order was forced through, no limit was widened to compensate, and the failures were written to the gate log rather than swallowed. That is the behaviour we want. It is also a clean demonstration of the ladder's structural weakness: the moment it most needs to add is the moment it is most likely to be unable to, and a ladder that cannot average down is simply a stuck long. Capital availability is part of this strategy, not an operational detail beside it.

What it cost

All 94 fills were taker at 0.08%. Fees came to $1.30 against $6.61 of gross profit — 19.7% of everything the strategy earned went to the exchange. Funding cost another $0.05. At a 0.8% take-profit target, two taker crossings eat a fifth of the move before anything is left over; this is the single largest structural drag on the design.

Our internal projection also carries a modelled rebate line of +$0.65. We are not counting it: no rebate has been paid into this account, the ledger column for a confirmed rebate is empty, and modelled income is not income.

Against the only benchmark that matters

SOL was 76.41 when the first rung opened and about 77.3 today — +1.2%. The wallet did +5.14%. So the ladder beat holding the coin by roughly four points over 23 days, on an average deployment of only 17% of the account.

Now the part that decides how much that is worth. We measured the maximum drawdown of SOL inside our 23-day window and compared it with every other 23-day window since February 2021 — 1,977 of them:

A ladder earns from oscillation and suffers in a sustained trend. We tested it in the calmest 1% of SOL's history — which is precisely the regime it is built for. That does not make the +5.14% fake; it makes it unrepresentative, and it means the rate cannot be carried forward as if the market always looks like this.

So what would $5,000 become?

This is the question that gets asked about every working system, and it deserves an arithmetic answer plus the reasons that answer is not a forecast. We wrote the calculator into the repository so the number always arrives with its caveats attached:

python -m ai_os.trading.mexc_projection --capital 5000 --years 1 2 3

It reads the live ledger, compounds the measured daily rate of 0.2175%, resamples the 21 observed trades for a confidence band, and then applies an empirical liquidation tail taken from five years of SOL daily candles. At 3x an isolated position is wiped by a 33% adverse move; that has happened in 9 of 66 non-overlapping 30-day blocks (13.6%) since 2021. At 2x the threshold is 50%, which occurred in 4 of 66 blocks (6.1%).

Lower leverage is not a smaller version of the same bet. At 2x the rung buys two thirds of the notional, so the naive number is far smaller — and the liquidation threshold moves from a 33% drop to a 50% one, which is roughly half as frequent. Over three years the cautious setting ends ahead of the aggressive one in the tail-adjusted columns, which is the entire argument against the aggressive one.

What our own gate says

The ledger's verdict function does not read any of the above as encouraging. It returns insufficient_sample: 21 trades < 100, and until that threshold is met no result from this experiment is allowed to be quoted as an edge — including by us, in this article. The initiative is registered as a trial with a stated kill criterion: if the verdict has not turned after 100 closed trades, the family is culled and the write-up becomes an autopsy rather than a case study.

What would actually change our mind is narrow and boring: at least 100 closed trades, at least one sustained downtrend survived without capitulation, a random-entry benchmark of the same trade density that the ladder beats, and a fee structure where 19.7% of gross profit does not leave the building. Until then, the honest summary of 23 days of real money is that a simple deterministic ladder made $5.31, did not blow up, and told us more about our own accounting than about the market.

All figures are quoted from the live ledger and cross-checked against the venue's own closed-position and transfer records on 19 August 2026. This is a published engineering result, not a forecast, not a signal, not an offer, and not financial advice. Leveraged perpetual futures can lose more than the margin posted.