Five AI models read the market every 30 minutes, and you see every vote. A rules-based grid places the orders. The two never touch — and this page shows you both, all the way down.
Every 30 minutes, for each of the eight coins, five models each receive a role-specific slice of the data and answer the same question: is this market trending up, trending down, or ranging? Each vote lands on your bot's card next to the weighted read. Nothing in this section places an order.
The weights are fixed and published — 0.40 / 0.15 / 0.20 / 0.15 / 0.10, in the order above — and every model's accuracy is tracked and shown on every trade receipt. We do not claim the weighting makes the read smarter. In our own scoring it has not yet beaten a naive always-RANGE baseline, and we say so rather than hide it.
A real read, captured from the same data your bot's card is drawn from — BTC, 2026-08-25 at 20:30 UTC. Four seats called it UP; the Risk seat called it a RANGE. The weighted read is UP. You see who said what.
Every order your bot places comes from a short list of published rules. Here they are, in the order they run.
The engine takes the last seven days of hourly prices for the coin. The floor is the 5th percentile of the lows; the ceiling is the 95th percentile of the highs. A volatile coin gets a wider band automatically. You set nothing here.
The band is divided into rungs, evenly spaced in percentage terms, and your allocation is spread across them. The rung count is derived from the allocation: the minimum is $250, a slice under $500 runs fewer, larger rungs (about $50 each), and $500 or more runs the full 10. Every rung must clear our $10 minimum order (the exchange's own minimum is $1); bigger rungs keep a smaller slice from running short of cash per rung.
Each rung below the current price holds a buy order. The bot does not chase the market: orders rest on the exchange as limit orders and fill only when price trades through them. Resting orders are charged the maker fee.
When a buy fills, a sell is placed one rung above it. When that sell fills, the buy is re-armed on the rung below. Buy the dip, sell one rung up, re-arm — that loop is the whole strategy, and every turn of it is a receipt.
When the seven days are up, a new band is set from the new seven days. Coin bought in the old band is not dumped at market: it is carried into the new band and re-listed as sell orders at the rungs above the current price. A falling week means holding coin until price climbs back to those rungs.
One rule sits above the grid. It is price math, not an AI decision — and we describe exactly what it costs.
If the coin falls 15% below its 30-day high, the bot cancels its resting orders, sells everything it holds at market, and waits in cash. It realizes a loss to cap a larger one.
The grid resumes when the coin recovers to within 7.5% of that 30-day high. Between those two lines it does nothing — and your card says so.
A gate that sells near the bottom is worse than no gate in a market that recovers. In our paper pilot, the gated arm earned less than the ungated arm over a window that recovered. We say that wherever we mention the gate.
A grid bot spends real stretches of time doing nothing on purpose. Every idle state on your card says what the bot is doing, why — with the live numbers — and what ends it.
The coin ran above the band top. Buy rungs are armed below; nothing fills until price comes back down or a new band forms at the next re-band. The card shows how far above the band you are and when the next band forms.
Price is under the band floor. If rungs are armed, buys fill on the way down and carry. If none could be armed, the card says so and names the price that re-arms them.
After a falling week, the cash left over may not cover a full ladder of $10 orders. The bot pauses the block — nothing is sold — and the card shows the exact math: cash, rungs, dollars per rung, and when it is re-checked.
The drawdown gate fired. The card shows how far the coin sits below its 30-day high and the recovery level that resumes the grid.
Every fill writes a receipt at the moment it happens — nothing is reconstructed after the fact. The mechanical reason comes first: which side, which rung, what filled. The five-model read sits beside it as context, so you can see what the models thought of the market at that instant — not because the read placed the order. It didn't.
It's the difference between "trust our algorithm" and "here is what the algorithm did, line by line."
Start in the Sandbox — paper trading on real market prints, with real fees modeled and no real money.
// A note on transparency. The five models are shown by role, not by vendor; the mechanism is shown in full — real votes, real weights, real math. The market read above is a real read (BTC, 2026-08-25 20:30 UTC), captured from the same data your bot's card is drawn from. This page explains how the product works; trading performance is reported separately, and always with its fill model, window, capital base, and fees stated.