The S07 pair held at the budget guard
The closeout says one S07 replicate pair was held at the approved budget guard, so 34 of 36 clips ran. Was that guard a number you watched by hand while the pod ran, or does your runner stop itself when spend reaches it?
The runner itself was not monitoring RunPod spend. I monitored billing and runtime from the orchestration side against a hard ceiling of $32 or 16 GPU-hours, plus an internal $31.30 safety guard. The first S07 pair took about 100 minutes and cost roughly $3.52. Based on that measured cost and the amount already consumed, I held the second pair before it ran, canceled the remaining queued work, and stopped the pod. So the guard acted as admission control between jobs—not as an automatic mid-generation kill switch. I’d like to make that enforcement fully automatic in future tests.
Thanks, that is clear: the guard acted as “admission control between jobs—not as an automatic mid-generation kill switch”, and you would like that enforcement fully automatic.
That is the part my project, Symbioza, takes off your plate, so you do not have to build and babysit your own execution agent: dead pods, disks that fill, a teardown nobody confirmed. The dollar cap is set before anything is booked, and the machine is torn down and the teardown confirmed when the job ends.
You would not need an account: you send me the container and the command (your ComfyUI API workflows writing to an output folder), I run it and send the artifacts back. Would one of your next H3 test cells be a fair trial?