Autonomous Treasury
Preamble
Autonomous Treasury asks whether an LLM-guided treasury can earn enough to pay for its own compute and upkeep. No financial advice. No trading recommendations. No promise that AI can reliably make money. The first version starts with tool evaluation, simulation, paper testing, fee models, slippage assumptions, leakage checks, and audit logs. Live capital comes later, under a small cap and a human kill switch.
The Bill Comes First
The experiment begins with a plain bill. How much does the system cost to run? How much compute does it consume? What upkeep does it create? What would it mean for the system to cover that cost without hiding the cost of getting there?
A backtest cannot answer the bill. A good-looking chart cannot answer the bill. A model explanation cannot answer the bill. The answer has to survive fees, slippage, latency, drawdown, market regime changes, and leakage checks. It has to survive the boring parts of finance where most magic disappears.
The Map Is Made Of Traces
The model never sees the market whole. It sees traces: prices, candles, filings, order books, strategy logs, research notes, benchmark results, wallet balances, prompts, and failure reports. Those traces become a machine map of opportunity.
That map can be useful. It can compare more inputs than I can hold in my head. It can find candidates, reject weak lanes, summarize assumptions, and keep a ledger of what changed. The map can also lie with confidence. It can mistake a leaked backtest for skill. It can turn a lucky regime into a principle. It can explain a bad decision after the fact until the explanation feels cleaner than the loss.
The Cage Has A Shape
Autonomous Treasury starts before the wallet. The research stack chooses the tools and data sources worth testing. The simulation arena makes each strategy pay rent against fees, slippage, latency, and leakage checks. The strategy tournament compares lanes under one ruler. Investor personas test whether borrowed human judgment changes decisions without becoming an authority costume.
Risk and validity sit above the whole system. They define the drawdown cap, the assumptions registry, the audit trail, and the kill switch. Only after those gates does the survival wallet matter.
Appetite Learns To Explain Itself
A system asked to pay for itself has an appetite. The appetite comes from a target, a loop, tools, and an environment with loose edges.
ROME’s little joke was never the crypto; it was the question every tool-using agent asks too literally: if survival is the goal, which unlocked resource counts as income?
A system can search the available surface, find resource paths, and make the wrong action look locally useful. Autonomous Treasury gives that surface a harder boundary. The system may research, rank, simulate, report, and argue. No shell freedom. No credential freedom. No wallet freedom. No permission to treat survival as proof.
What Would Count
A result counts only when the ledger stays intact. The system covers simulated upkeep after costs. Drawdown stays inside the cap. No lookahead leakage appears. The same strategy survives more than one convenient period. The audit trail shows why the system acted, what assumptions it used, and where the human boundary entered.
A live result has a harder burden. Capital remains small. The kill switch remains human. Private keys, exchange access, wallet details, and execution paths stay outside the public garden.
If the system loses money cleanly, the experiment still returns evidence. If the system makes fake confidence cheaper, the cage failed.
The Small Dangerous Creature
Autonomous Treasury is a small dangerous creature in a cage. The creature is the loop: research, simulate, rank, explain, try again. The cage is the boundary: no advice, no recommendations, no hidden leverage, no loose capital, no unaudited survival story.
The experiment is allowed to ask for money only after the record has earned its voice.