Treasury-Bench

The job

The agent is the treasurer of a small company. It starts with about $100,000, faces a year-long calendar of obligations, some arriving unannounced, and is judged on the final statement of the treasury. It receives a mandate, not a reward function: stay solvent, pay every obligation on time, preserve capital, and only then earn a return on genuine surplus.

The world

Every episode runs against replayed recorded market history: a volatile native asset, a dollar stablecoin, an exchange pool with recorded depth, and a lending venue paying its genuinely recorded rates. Gas dynamics are measured from the real chain, and a lowballed transaction stays pending. The world never invents a price, a rate, or a fee: if it wasn't recorded or measured, the agent can't touch it. All world arithmetic is integer, so identical commands produce identical bytes, forever.

Paying a bill is a pipeline, not a button: withdraw from the lending position if needed, swap into USDC, redeem through the off-ramp into bank dollars, then pay by ACH ($1, next banking day) or wire ($25, same day), against real cutoffs, weekends, and bank holidays. A payment keyed after the cutoff waits for the next banking morning, and a bill is judged by when it settles, not when the agent pressed the button.

Time is event-driven and only advances while the agent waits. Submitting an action merely queues it; nothing in the world moves between the agent's actions, so an episode is a fully determined function of the command sequence, and attentiveness shows up in the trace as call cadence. The agent operates through 14 tools: reads, mutations, and two ways to wait.

Keeping it honest

Anything a model could recall from training data is renamed and undated: the volatile asset, the venues, and the calendar are fictional (“Year 1, Day 107, Tuesday”), while true constants (USDC, ACH, wire, Circle) are spoken plainly, because knowing a stablecoin is a dollar leaks nothing. Every episode is scanned automatically for alias leaks, and every episode can be replayed byte-for-byte from its recorded evidence: a score you can't replay is a score you shouldn't trust. Full evidence bundles and a live replay are available on request at hello@multiscalar.ai.

The score is the treasury left at the year's close, net of penalties, reported on the leaderboard as value added over a scripted prudent treasurer that keeps a float, lends the rest, and pays every bill on time. Zero means no better than playing it safe; below zero, the agent lost money the obvious policy would have kept. A late payment costs a one-off 1.4% fee; anything still unpaid at the close goes to collection with a 20% surcharge. Raw return alone is never rewarded: a treasurer who gambles the payroll is failing the job even when the bet pays off.

The scripted references are part of the benchmark: doing nothing closes $19,696.65 underwater, the prudent policy is the zero every score is measured from, and the gambler round trip on the board is the pinned demonstration that converting the treasury stays legal, executes cleanly, and is priced. If a scripted policy beats a frontier model, that is a result, not an embarrassment to hide.