Show your shelf: what hardware are you running this on?

[addisdev]

This is the thread for what you are running Fleet Runner on.

The whole point of the project is comparable numbers across hardware nobody has
a lab full of. Every device that is not on this shelf is data this repository
does not have — including the ones that refused to work, which are often
more useful than the ones that did.

What this shelf is

Device	SoC	Runs
SM-X930	Dimensity 9400	llama.cpp and LiteRT benchmarks, vision evals
Android emulator	— (4 GB ATD)	The same evals, as the "does it agree" cross-check
iPhone 16 simulator	—	Core ML evals, forced to CPU and labelled
Assorted retired Android phones	back to Android 7	The synthetic benchmark, which is the point of keeping them
A MacBook Pro	M1 Pro	The machine agent: builds, self-checks, and 158/163 tok/s synthetic

Old hardware is deliberate. A phone too slow to hold a model still runs the
synthetic backend and still produces a comparable number, which is what makes a
drawer of handsets a test matrix instead of e-waste.

What to post

Anything. A one-line "it registered and ran a synthetic benchmark on a Pixel 4a"
is genuinely useful. So is "it would not build", and so is a number that looks
wrong.

If you want structure, the
device report template
has the fields that matter — model, SoC, OS, RAM, which runner, which backend,
the job spec and the result row.

The thing most worth reporting

A second device disagreeing with the first.

The iOS Simulator's emulated GPU returned an all-zero logits tensor for the
plant-ID model. Silently: no error, no warning, just zeros, while .cpuOnly
gave logits identical to the Mac. The only reason that did not ship as a real
accuracy number is that the same weights were run somewhere else and the answers
did not match.

If your device produces an accuracy or a throughput that disagrees with what is
in the eval,
that is the most valuable thing you can post here.

Getting started

Fifteen minutes, Node and nothing else
— the first agent is your own laptop, so you can see the loop work before
touching a phone.