A ledger for long-horizon intelligence
We built our trading ledger to evaluate autonomous agents in live markets, where the consequences of a decision can take weeks or months to become clear. Each agent operates through an independent simulated brokerage account, while a shared system of execution, accounting, and valuation lets us follow large populations under consistent rules. Between decisions, prices continue to change and earlier positions affect the capital available for whatever comes next.
These accounts underpin the live paper-trading portfolios described in our previous announcement and form part of the financial simulation model we are building at Infinite Ascent. We use the resulting record to compare models over time, examining how their decisions affect returns and risk as well as whether they can sustain a useful process across changing market conditions.
Decisions whose consequences take time
Most evaluations can judge an answer as soon as it is given. A coding task is deterministic and does not change with time: the code passes its tests or it does not, and the same code passes them tomorrow. A trading decision is neither. Its result arrives weeks or months later, and chance shapes it along with the decision. A model may identify a change in a company’s prospects today, take a position tomorrow, and wait several earnings reports to discover whether the thesis was sound. While that position remains open, the agent must interpret new information and decide whether to maintain, increase, or abandon it.
The agent has to make those judgements while managing the rest of its portfolio. Capital committed to one idea is unavailable for another, and a loss changes the resources with which subsequent decisions are made. Even a promising thesis depends on the agent being able to size the position sensibly, preserve enough capital to pursue it, and recognize when the evidence has changed.
Because a correct analysis can still lose money and a flawed one can profit from a favourable market, a single return is a weak basis for judging capability. We are interested in whether models can sustain a coherent research and decision process over many such episodes, and whether the behaviour that produces good outcomes persists when market conditions change.
The accounting layer of a financial world model
In the financial world model we are developing, we want to replay periods under controlled conditions, vary the constraints under which agents operate, and compare alternative decisions without waiting months for each experiment to finish. This requires a faithful representation of the information available at a given time, together with a dependable account of what each action does to a portfolio.
The ledger maintains the cash, orders, positions, and history of each account independently, so a population of agents can share execution rules without sharing capital or access to one another’s private state. Within each account, it records how an instruction becomes an accepted order, whether and how that order fills, and how the fill changes cash and holdings. Each change is booked on both sides, as in double-entry bookkeeping, so the account’s books balance. An agent can attach its thesis and expected horizon to an order, preserving the expectation against which a later outcome can be examined.
On 5 October, one agent placed an order for 76 shares of NVIDIA, describing it as “the AI earnings engine” and expecting it to outperform the S&P 500 ETF, SPY, over the week. The order recorded a lag of at least 1.5 percentage points behind SPY as one condition that would invalidate this view. Three days later, the agent’s closing order stated that this condition had been met. The ledger recorded the purchase at $237.22 and the complete sale at $230.29, realizing a 2.92% loss on this simulated position.
A network interruption should not turn a single instruction into two trades, so commands are recorded durably and retries can recover their original outcomes. The accounting rules are also separated from the services that supply market data and store the results, with the engine producing the same answer whenever it receives the same state and inputs. We can therefore test its financial behaviour precisely and preserve those rules as we move between live observation and controlled simulation. Historical replay of the entire information environment remains a broader piece of work, beyond the ledger itself.
A simulation grounded in live markets
Before accepting a new instruction, the ledger checks available buying power, accounts for capital already committed to outstanding orders, and applies the account’s trading rules. Order size is constrained using volume from prior completed market sessions, so an agent cannot assume unlimited liquidity or benefit from knowing how much will trade later in the day.
Equity purchases are priced at the observed ask and sales at the bid, so the spread affects the portfolio alongside any additional execution costs specified by the simulation policy. A limit order can remain open until its price constraints are met or the order expires or is cancelled. A fill requires a valid, sufficiently recent market observation received after the order was accepted, during an eligible trading session.
Where an account permits short selling, the ledger checks simulated borrow availability, reserves margin collateral, and keeps the sale proceeds separate from spendable cash. Borrow fees accrue at a default annual rate of 30 basis points, with rates and availability adjustable for individual securities. Purchases remain cash-funded, without margin loans to finance leveraged positions.
The price at which an order fills determines what the account pays or receives, while subsequent market marks determine the value of its remaining holdings. We bring those marks together with cash and positions to measure performance against the account’s initial capital. Accounts receive their funding at creation and do not yet support deposits or withdrawals. If the evidence needed for a valuation is missing, the result remains unavailable rather than being presented as an unchanged portfolio.
Because these orders are simulated, they do not enter an exchange queue or move the market. Dividends and stock splits are not applied to the accounts in these experiments.
Results from the past four weeks
The figures below follow paper portfolios launched during the four-week period from 14 September to 9 October. Each week began with a new population of agents, with up to 50 portfolios for each model family and about 570 across the three runs shown. Results are grouped by model family, with performance weighted by opening capital. The week of 21–25 September is excluded for system maintenance.
Each line follows the pooled return of one model family’s portfolios from the week’s opening value.
Claude’s portfolios led the model families every week, gaining 2.39%, 2.66%, and 1.01%. The others ranged from −2.40% to +0.79%.
The paths differed as much as the totals. In the latest week, Grok’s portfolios were up 1.64%, then fell 1.52% from that peak to 0.10% before finishing at 0.72%, against a fall of 0.95% for Claude’s.
The largest fall from a preceding peak in each pooled portfolio, at daily closes.
These panels are a small part of the record: behind each line, the ledger holds every order, fill and position of the portfolios it pools.
Learning from longer experiments
The live experiments are helping us refine the financial simulation model by showing which decisions, constraints, and market conditions it needs to reproduce. Controlled replay then lets us examine the same periods under different assumptions, making it easier to investigate why an approach succeeded or failed and whether a change to a model improves its behaviour. By keeping the accounting consistent across those comparisons, we build on the evidence gathered in live markets without having to wait for the same conditions to occur again.