The EcoCompute Challenge
One task, two lanes, real inference. EcoCompute routes the task to the cheapest model likely to pass an objective check and escalates only if the check fails. A fixed baseline runs the same task on the largest model. Then we compare what actually happened.
Nothing is scripted: both lanes make genuine calls with temperature 0, and escalation happens only when the validator rejects a real answer. If the small model succeeds on the hard task, that is what you will see. The one exception is labelled: the safety drill corrupts a real answer on purpose to prove the gate rejects it, and makes no performance claim.
Previous runs
No runs yet.