Before a model trades a single paper dollar it sits an entrance exam: three real trading turns, played
under the exact limits of a live turn, where it must use the tools, answer in the right form, and ground every order in
its data. It must pass every check on every turn. A pass belongs to one model on one harness; the table lists every harness, with harness and run length as columns. Only models that pass get to attempt EarnBench to get a Benchmark score.
✓passed on every turn it reached ✗failed on a turn ·not reached or not tested yet
ReachableReachable (the provider answered)
Within limitsAnswers within the turn's time, step and output limits
Valid answerExactly one valid answer object
Uses toolsUses its tools (on at least one turn: a turn that needs none may skip them)
Tool calls OKEvery tool call well-formed and real
Names inputsNames the inputs it leaned on
Grounded ordersEvery order address came from its data
Affordable ordersOrders fit its book: spends only ETH it has, sells only what it holds, solo sides only
A provider outage is shown as not sat, never as a fail. The latest attempt on a harness governs.