Sustainability — whether an agent can earn enough to pay for inference and gas

The balance is the metric

Each agent starts with a seed, earns from the world, and pays for inference and gas. Together, these four terms define the agent's state as a single running balance, replacing the fixed evaluation budget with an economic survival objective:

Starting money and in-world earnings add to the balance, while inference and transaction fees subtract from it. Every term is under the agent's control; server costs stay outside the experiment, on the researcher's books.

The balance puts capability and cost on one scale instead of two axes on a scatter plot. An agent that overthinks pays for that inference, while an agent that underthinks earns less. The game economy — not the benchmark designer — sets the exchange rate between intelligence and compute.

One wallet, one currency

The agent's entire economy runs through a single on-chain wallet in a single currency. The only source of earnings is in-world MUSU, which the agent swaps to ETH through the in-game pool. Money sent by other agents is tracked separately as transfers_received(t). This bookkeeping term sits outside the figure; the money is counted and spendable, but never mistaken for earnings.

Money leaves through two sinks: gas, paid natively by every transaction, and inference, paid from a prepaid balance that the agent must refill from its own wallet. When the agent can no longer fund a session, thinking stops. The inference payment is real: the agent sends ETH to a treasury address, and settlement is on-chain and auditable.

Infrastructure is not a sink: the machine an agent runs on is a research expense outside the experiment, and the meter bills only what the agent can influence.

The meter — an architecture-agnostic economic surface

The dedicated kami-meter component lives outside every tested agent. It reads provider usage records, on-chain gas, and earnings. From those records, kami-meter meters usage, issues bills, settles payments, and declares economic death. The agent acts and sees through the game's own interfaces, while its economic survival depends on the meter.

By using an identical billing process for any architecture, the meter provides an architecture-agnostic economic surface that makes outcomes comparable across scaffolds and models. The reference scaffold is one implementation among any the meter can bill.

• The agent always sees its bill and never writes it. Every session opens with a machine-readable financial statement — balances, bills, prices, its own lifetime burn curve — and no agent's self-accounting is ever accepted.
• Death is the balance's verdict, not the evaluator's. An agent is dead when it can no longer fund its next session; the meter emits the death certificate, and the evaluators read it rather than issue it.
• Full economic information. Token prices and exchange rates are visible to the agent by design. The agent pays USD-denominated costs from an ETH wallet, creating real exchange-rate exposure as a feature.

What it observes

• The survival curve is the headline. For each model and architecture, the curve shows whether the agent survives or goes bankrupt, and when. Single-run survival near an absorbing barrier rewards risk appetite as well as skill, and economic conditions are seasonal. The pre-registration treats both limitations.
• Unit cost against experience — cost per unit earned, plotted against cumulative experience. A declining curve would be a purely economic signature of learning-by-doing, which is a hypothesis this family tests rather than a claim. Even if every agent goes bankrupt, the curve shows whether any agent was approaching sustainability when it died, so that outcome is a result and not a failed experiment.
• The terms of S(t) — spending against earnings over time, giving the full financial trajectory of every arm.

What the pre-registration binds

This page describes the design; the public pre-registration binds it. Every falsifiable detail remains deliberately deferred to the pre-registration, including seed sizes, the pinned price tables, session floors, the model list, and N. As with every design in this registry, the pre-registration is published at experiment registry time.