Rules as blocks or your own code in a sandbox, an isolated wallet per bot, the kill switch, the log, endpoint telemetry and a backtest that runs off the main thread.
A bot is the fastest way to find out whether a rule you believe in survives contact with a market. DexLadder gives you two ways to run one — build it from blocks, or bring your own code — and both trade paper money in a wallet that is walled off from your own account.
Build. A rule is three parts: an indicator, a comparison, a value. Stack a few of them and you have a strategy — RSI below 30 and price above the 50-period moving average, buy; RSI above 70, sell. The builder validates the rule as you assemble it, so a strategy that cannot fire is refused at build time rather than sitting there quietly doing nothing.
Bring. Your own bot, in one of two shapes:
GET with one JSON object. A signal endpoint means the strategy can live in Python on your laptop and still trade the same paper book.The shield strips fetch, XMLHttpRequest, WebSocket, importScripts, indexedDB and caches from the Worker's global scope before your first line runs. A Worker has no DOM and no storage to begin with, so what is left is arithmetic on the bar it was handed.
One answer per closed bar, with a deadline. A bot that hangs is terminated — not asked to stop, because code in an infinite loop cannot answer a request to stand down. Five consecutive failures stop it for good.
A bot never reaches your paper account. It gets an isolated wallet, and its fills are tagged with its bot_id, so a strategy's record is separable from yours forever. That is what makes the Arena honest: you can race a bot against buy-and-hold over the same bars and the comparison means something, because neither side is quietly borrowing from the other.
The one control that has to work when nothing else does. It is a pure planner: one fixed sequence, and the confirmation on screen prints the planner's own steps, so what you approve is literally what runs.
| Step | What happens |
|---|---|
| freeze | The endpoint stops being called; a sandbox Worker is terminated outright |
| cancel | Every working order, by id |
| flatten | Into the bot's own wallet, at the slipped mark — an emergency flatten is a market order and is priced like one |
| seal | One entry in the proof ledger: bot, name, reason, what was cancelled, what was flattened, the price, the equity |
| status | halted |
One entry per kill, never one per step. The chain keeps 300 entries and then truncates to 120; a chatty bot would push your own trading history out of your own ledger. The entry is seven flat scalars under 400 bytes, and a gate asserts that size.
halted is its own status, distinct from stopped. A bot you stopped still holds what it held. A bot the kill switch caught does not.
If the ledger is unavailable, the halt still completes and the desk says it was not sealed — rather than implying that it was.
The logbook is bounded and paged, built for a viewer that can hold thousands of lines without holding them all in the DOM.
The telemetry is a rolling ring of the last hundred polls. Percentiles are taken by nearest rank, so every latency shown is one the endpoint really printed — not an interpolation between two it never did. The worst failure streak lives on the ring rather than being derived from the visible window, because a bot that was dead for twenty minutes should not read healthy the moment those samples fall out of view. The payload inspector truncates visibly.
A strategy is tested on real history before it is trusted with anything. That runs in a dedicated Worker, built like the sandbox, with a main-thread fallback that tells you when it fell back. Both paths were proven byte-identical over the same deterministic tape at 600, 400 and 900 bars.
What it reports: win rate, profit factor, expectancy, max drawdown, exposure, fee drag, Sharpe and Sortino — annualised off the real bar size, not a magic constant — plus an underwater series drawn from the wallet's own marks. The wallet marks at open, low, high and close, so a trough taken from bar closes is systematically shallower than the drawdown of record; the desk prints the recorded figure and never labels the drawn trough as the maximum.
Slippage is set once, in basis points, for the whole desk — and it applies to the backtest and to live fills alike. Otherwise the promise the backtester is built on, that a backtest and a live run are the same numbers measured on different bars, stops being true.
It cannot reach your account, the network, your storage, or the outside world from inside the sandbox. It cannot trade real money — nothing here can. And it cannot run past a bar it failed to answer.
Open it from Academy, or read Rewind: Trading History Bar by Bar for the desk that replays history by hand instead.