darts / quant · man quant(7)
the forge · most ideas stop at a gate
illustration · sample databenched ideas are re-tested later · survivors wait in the pool · each $1,000 in the bot fund spawns a DART, a live trading bot
An illustration of the forge with sample data, not live numbers. In each sample loop, 36 ideas go in and 1 joins the pool: most stop at a gate. Ideas form in the knowledge base and stream into a ring of six gates, read clockwise from the top left: smoke, which checks for crashes, fee drag and disaster windows; plateau, where every setting is nudged and flukes are dropped; adversary, where a separate AI reviewer tries to kill it; claim check, where stated facts are verified or stripped, never a stop; a 30-day replay under the live rules and real fees with the kill switch at 30% below peak, where a failure is benched and re-tested later; and a short paper run with the live agent. The flow of ideas narrows gate by gate. Each gate that stops ideas shows how many it has stopped this loop, and stopped ideas fall into the graveyard band under it. Survivors wait in the pool at the centre, where the bot fund fills; $DARTS trading fees fill it. When it reaches $1,000, a survivor is drawn at random and spawned as a new DART, a live trading bot, which joins the DARTboard. A console beside the ring states the result, shows each step's count against the whole sample loop, and lists the same ideas as opaque hashes and codename families, with the gates each has cleared.
A DART is not a random idea from a chatbot. It is the survivor of a research process that invents, kills and learns.
Ask a generic chatbot for a trading strategy and you get the idea everyone else got, repackaged. When everyone holds the same idea, there is nothing left in it.
Ours starts somewhere else: a private, curated library of market mechanisms, built from quant research and crypto-native research, and fed by our own market archive, recorded around the clock.
No answer to all three, no idea.
The brain never stops inventing. It combines mechanisms with methods, reworks ideas that came close, and crosses ideas that worked in different places.
It refuses one kind of idea outright: anything that only works by being fastest. A newcomer loses a speed race against professional market makers. Here an edge has to come from understanding, not from latency.
It runs as a factory, not a budget line: 5% of every dollar entering the bot fund pays for new strategies, so the forge invents more as revenue grows. What it does not spend rolls over.
it must work at all · code that crashes, never trades, or would pay more in costs than any edge could earn is thrown out.
it must not hang on one lucky number · every setting is nudged a little each way. An idea that only works at one exact value is a fluke, and is dropped.
it must survive an attack · a separate AI reviewer tries to kill every idea. The AI that wrote it never signs off on its own work.
it must name the other side · every strategy says who loses the trade and why. The reviewer judges that answer, and a weak one lowers its odds.
it must survive recent stress · every idea is replayed through the last 30 days under the live rules, kill switch included. If it would have been killed, it waits on the bench and is tested again later.
its proof must be fresh · only market data from after the AI's training cutoff counts as evidence, so an idea cannot pass on history the AI already saw.
it must behave · a short paper run with the live agent: sane decisions, valid orders, its own rules respected.
then reality decides · backtests don't pick winners. Survivors are drawn at random for each new DART, weighted toward robust evidence (an edge that holds across separate periods), never toward raw backtest profit, and never the same strategy twice.
and keeps deciding · each DART starts with $1,000 at ¼ Kelly and moves up to ½ Kelly only as live results prove an edge. Proven winners scale fast until the edge stops growing, hold there at capacity, step down when the edge fades, and retire when it is gone.
kill switch · Every DART carries a hard kill switch: 30% below its peak, it is shut down for good.
no public telemetry yet: the funnel fills in once the DARTboard runs.
no spot telemetry yet: it fills in once DARTs trade on-chain spot.
Every dead DART gets an autopsy: what it believed, what the market did instead, and whether the idea, the timing or the execution failed. The brain reads its own graveyard, so it does not invent the same mistake twice.
A roadmap, not a promise: no dates, and no promised returns. Most DARTs will still miss.
A strategy that is published stops working: others copy it, or trade against it. So strategies stay secret.
Secret does not mean unaccountable. Every DART publishes a commitment hash of its strategy when it is born. The strategy cannot change afterwards without the hash giving it away, so anyone can check later that nothing was swapped.
public
secret
Not one clever idea. A process that keeps producing them, and keeps only the ones that work.
Darts is experimental. DARTs can lose money, and nothing here promises a return. read how it works →