The governance map
Every scheduler is a point. The dial is a curve.
Each point below is one full simulated day of the 320-GPU cluster — how much energy it spent per job against how many jobs it let slip. The sweep runs live in your browser; pick a workload and watch the frontier bend.
The headline result: a governance weight tuned on a synthetic workload (λ=0.65, 1.6% violations) produces 76.5% violations when transplanted onto the production trace — static weights do not transfer between workloads. The sweep below reproduces this live in your browser.
In 30 seconds: each point is one simulated day at a different setting of the energy-vs-reliability dial. The right setting depends on the workload, the workload changes, and a controller that re-tunes the dial beats any number a human picks once — that is the product.
Energy × reliability — one point per scheduler
indigo points = governed runs, labeled by governance weight λ · star = self-tuned controller · ring = the knee, the last λ that holds violations under 15%
All runs — one simulated day each
| Scheduler | λ | Energy/job (Wh) | SLA violations (%) | Migrations/epoch | Carbon/job (g) |
|---|---|---|---|---|---|
| first run landing… | |||||
Simulator and numbers reproduce the IEMCON 2026 paper's reference implementation bit-for-bit (39 pinned parity tests). Violation rates are conservative by construction; read them as paired comparisons, not production SLA figures.