Princeton University
Affiliation of the authors of science-of-agent-reliability (sayash-kapoor,
arvind-narayanan, Rabanser, Kirgis, Liu, Utpala), and — the part that matters to this spoke — the
maintainer of the HAL harness (Holistic Agent Leaderboard, princeton-pli/hal-harness), the
evaluation infrastructure that paper's 14-model, five-runs-per-task sweep was executed on.
That makes Princeton a node supplying an instrument rather than a result: HAL standardizes agent execution and tracks cost per run, which is what makes a repeated-run protocol affordable enough to publish at fourteen models. The spoke's benchmark pages mostly record what a harness measured; this is one of the few places it can say whose harness.
Nothing beyond these sources is claimed here.
Connections
- affiliation of sayash-kapoor and arvind-narayanan
- maintainer of the HAL harness used by science-of-agent-reliability