Anyone can request a bench run — pick a model, a bench, a harness. Every request is reviewed, then dispatched to the single GX10 one at a time, only when the box is free, so nothing ever collides with what’s already running. With time estimates, re-ordering, and a full result you can open up.
The pipeline — request to result
What the queue gives you
estimate
ETA per run
Drawn from real past-run durations, plus a total backlog time for everything approved.
control
Reorder & pause
Bump the run you care about to the front; pause and resume without losing its place.
observe
Drill into results
Every finished run opens to a per-bench grid — pass/fail, tok/s, steps, and the run duration.
schedule
Run later
Queue for a future time so heavy runs land overnight, out of the way.
The guardrails
cost
Local models only
The public form offers only local models — cloud stays internal so no one can run up token credits.
accountability
Name + reason
External requesters say who they are and why; nothing runs anonymously and unreviewed.
safety
The wheel
Requests queue freely, but only dispatch when the box is explicitly handed over — the operator keeps the keys.
The queue turns "can you run X for me" into a self-serve, reviewed, observable pipeline. You ask, it estimates, someone approves, and the bench runs itself — then hands back a result you can take apart bench by bench.
Why it exists: a benchmark that only its author can run is a bottleneck. This makes the GX10 a shared, orderly resource — anyone can propose work, the operator stays in control of what actually spends the box’s time, and every run leaves a clean, inspectable trail.
IMR R&D · pipeline note · the request-a-run queue with ETAs, reorder, and result drill-in, July 2026.