Pools

A job’s pool says whose machines may run its units. Without one, a job is public:

job.yaml

name: product shots
template: image/sdxl-turbo@1
input: prompts.jsonl
pool: verified

Pool	Runs on	Price per unit
public	Any machine that can run it: ours, yours, and other organizations’ but for their private ones	The template’s price, or your container’s quote
verified	Ours, yours, and other organizations’ machines of trust 80 or more (trust and verification)	The same
platform	Ours alone	The same
own	Yours alone	Free

Price and speed

public, verified and platform cost the same per unit: a template’s price, or for your own container the quote its calibration made. What changes is how many machines can take the work, fewer in verified and fewer again in platform, which runs on our machines alone, so a job there can wait longer for a machine. A unit one of your own confirmed machines does is free to you, whatever the pool; in platform your machines do none.

own is free: the job is quoted at 0, needs no credit, and goes as fast as your machines do. Your machines give a job no network, so a job of pool: own must need none: give it network: none, or, for a template’s job, name every unit’s input by an upload.

On machines other organizations supply

• No network. A job that needs the network runs on our machines alone, behind our gateway. A template job whose every unit names its input by an upload needs none, since the machine’s agent fetches the input before the job sees it.
• Your own container runs only on machines whose owners accept other organizations’ images. Such a machine pulls the image from your registry, which sees its address.
• Checkpoints. A checkpoint a supplier’s machine made resumes on that machine alone; elsewhere its units resume from an earlier checkpoint of ours, if there is one, or start again. A machine that cannot checkpoint takes only jobs whose units take at most 10 minutes or that never checkpoint.
• What the machine is told. An alias of your job, never its id (your worker’s GRID_JOB_ID is the alias), and not where you keep results: what it makes goes into the job’s own folder in your organization’s space, and unitfabric results and the dashboard fetch it from there like the rest.
• Checked work. Outputs from these machines are checked by sampling: some of a machine’s units are checks whose right output we know, and a share of its units run again on our machines, whose output replaces a wrong one (trust and verification). A wrong output the checks miss is refunded, as credit, when you report it: write to the support address shown on your receipts with the job’s id and the ids of the units (guarantees). For work that must stay on machines we run, use pool: platform.