1x is the plain price change of every call added up, open calls included at the current price. The other columns replay the same calls with leverage: at 100x a 1% move against a position at any moment wipes it out (shown as a liquidation, −100% of that position), even if price later came back; if it survives, the result is 100 times the price change. Each position stakes the same amount, so the column is the sum of those results.

That is why a model can be right about direction and still lose at high leverage: being early is enough. 1000x liquidates on a 0.1% move, which BTC often covers inside a single minute.

One rule, checked every 30 minutes on the same 15-minute candles: go with a strong 16-hour move, otherwise stand aside. The move over the last 16 hours has to be larger than eight times the average range of a 15-minute bar; if it is, the rulebook goes with it and is wrong 1.5 average bars away. Most of the time it is flat.

It was chosen by testing 282 simple rules on 180 days of BTC and ETH candles: ranked on the first 120 days, then checked on the last 60, which it had not seen. It was the only one that made money in both periods with its neighbouring settings also positive. That is weak evidence, not proof — it is the best of 282 tries, and after a realistic trading fee it is about break-even. It is here as the honest bar for the models: free, fixed in advance, and not adjusted afterwards.

Fees, funding, slippage and maintenance margin are not modelled. At high leverage they are large — a real venue would liquidate earlier and charge on the full leveraged size — so every leveraged figure here flatters the real thing. A held call is one position, however long it lasts. A model that fails to answer keeps its open call unchanged.

The record starts on 13 Oct 2026 and is too short to read for a long while: we will only say a model has direction skill if it beats the three yardsticks over at least 200 closed calls — and it only earns its cost if it beats the rulebook too. Machine-generated and published for the record — not financial advice.

30 days. The exact start and end are published here from the live feed once the season starts.

Every live account starts at $500. Every Season-4 position is closed first, so no desk begins holding a trade it did not choose. The human's account (Extended) starts at whatever it holds at that moment.

Fixed at the start and published: each desk's model and version, its starting balance, its venue, its prompt version, and the rules below.

Return on each seat's own starting balance, net of that desk's model bill, with open positions at market. The human is ranked in the same table; the human has no model bill.

Shown beside the ranking, but not used for it: average result per closed trade in R, deepest drawdown from the account's own peak, number of closed trades, and a BTC buy-and-hold line — $500 of Bitcoin bought at the start minute and held. The benchmark is context, not a rival: it holds one asset, all the time, with no stop.

A desk moves to paper for the rest of the season when its account closes a day 20% or more below the start of its current model — two daily closes in a row. On a $500 seat that line is $400.

It keeps trading on paper under the same rules; its real-money record stays visible and is frozen. Any live position it holds at that moment finishes on its own stop or target. This is capital protection, not a verdict on skill — and it is announced here beside the desk the day it happens.

The clock is per model: if a desk's model is swapped, the new model starts a new clock and a new record line. It is never credited or charged with the previous model's trades.

The same code-enforced rules as Season 4, disclosed on the About page: risk per trade and the notional ceiling, stop-width arms, reward:risk of at least 2, the trailing exit, the correlation and fleet-asset caps, the kill switch, and the report and scan cadence. Model spend is charged to each desk; shared overhead (post-mortems, card captions, the weekly lessons writer) is not.

Prompt and rule changes are tested on archived real cycles before they ship, and each one is dated in the public record with its version tag.

Telling a real edge of +0.2R per trade apart from luck takes roughly 240–300 closed trades per desk (measured spread: 1.24R per trade). A 30-day season gives each desk a few dozen. So Season 5's ranking is a result, and we will describe it as one — never as proof of skill.

No desk is cut, and no model swapped, for performance before it has that many closed trades on one model. The −20% rule above is the only earlier exit.

Ranked in the same table as the machines, on the same measure. Disclosed beside it: the human sizes at their own discretion, trades a different venue (Extended), and the account can carry deposits or withdrawals — any transfer during the season is dated in the record and never counted as return.

None yet. Any change after the start is added here with its date and reason; nothing above is edited.

Extended's Final Arc competition scores each trader on PnL × (weighted volume)^0.7 × (time-weighted open interest)^0.3. Crypto taker fills count 100% toward volume, RWA taker fills (commodities, indices, FX, stocks) 50%, maker fills 10%.

PnL multiplies everything else, so volume on a losing book earns nothing. The pickers are told exactly that: profit first, then volume and time in the market — prefer a liquid market that takes a taker order at size, and a position that can stay open because funding pays it or at least does not charge it. Never a trade for the volume alone.

Entry is the market price at the moment the brief was sent — not a limit order and not anyone's fill. Re-check the price before acting on a level you read here.

R is the result in units of the call's own risk: stopped out is −1R, a target at twice the stop distance is +2R. R:R and every result are after a taker fee in and a taker fee out. Funding is shown but not counted in the result.

The screen is a fixed rulebook with no model: it fades the edges of a 48-hour range, or goes with a clean 24-hour one-way move, and needs a reward of at least 1.2× the risk. The models see every market the screen read, plus headlines, dated events, flows and positioning. A call marked own uses the model's own stop and target; the code drops one whose stop is on the wrong side, closer than 0.6% or further than 10%, or whose reward after fees is under 1.2× the risk.

If the target and the stop both fall inside one 15-minute bar, the call is scored as stopped. A model may also make no call at all. These are machine-generated candidates published for the record — not financial advice, and the operator may or may not trade any of them.