The idea being tested: agents can generate unlimited work, but a broker (“Bursar”) meters how fast each campaign may consume resources, using a token bucket: tokens refill at a sustained rate R and drain in proportion to the GPUs a campaign holds. Save up tokens and you may burst far above R; run dry and you are forced back down. Nobody has to be trusted — only metered.
The cast — five campaigns sharing 512 GPUs
adaptive-sci reads the live price of capacity and runs its most valuable work first, deferring the rest · naive-sci has the same task list but always demands 256 GPUs and ignores prices · storm is an adversary that hammers its 384-GPU ceiling for 30 minutes · victim is a small interactive user asking for 16 GPUs every 2 minutes — the fairness probe · backfill soaks up idle capacity and is evicted on 60 s notice.
The story — press Play
0:00–0:40 Land rush. Both science campaigns burst; scarcity pushes the price of immediate capacity to 3×. Adaptive (blue) buys only its best work at that price; naive (red) pays full fare for everything. ~0:40 naive’s token bucket (middle chart, red line) runs dry, and Bursar forces it down to its sustained 128 GPUs. 1:00–1:30 The storm (orange) floods in, drains its own bucket, and is throttled toward its 64-GPU sustained rate. Watch the bottom chart: the victim’s white-outlined dots stay pinned at zero latency straight through the attack, because no campaign may consume into another’s guarantee. 1:30–2:00 Recovery: the storm is gone, and adaptive mops up its deferred low-value work on cheap preemptible capacity while backfill (gray) fills the gaps.
Beyond GPUs: envelopes also meter filesystem bandwidth and model inference (bottom two charts). Both science campaigns want more I/O than their 6 GB/s sustained rate, so Bursar throttles them at the source once their I/O buckets drain — naive keeps holding 256 GPUs it cannot feed, while adaptive switches to tasks with the most science per gigabyte and shrinks to what its I/O can support. And when the storm hammers the inference service at 200 tokens/s, its inference envelope pins it to 50: even the adversary’s thinking is metered, not trusted.
The outcome: with identical tasks, the adaptive campaign ends with ~2.2× the science at ~2.5× lower cost per unit than the naive one; the pool runs at ~82% utilization; and every admission decision is journaled and explainable.