Questions, answered.
What you send, what you get back, what it costs, and what we never claim.
What do I send?
Photos in folders, one folder per category (2–20 categories), plus test photos with no labels. Any size or format; JPEG or PNG is easiest. A share link you control is enough; we never ask for credentials.
Can we evaluate radar, speech, text, or a custom hardware target?
Yes. Alongside photos, we run sealed evaluations on radar chips, spoken-word audio, multi-source document streams, and compact timing-neuron models for non-backprop silicon targets. You keep the test labels; we send back predictions.csv.
Do you ever see my answer key?
No. You keep it. We send back predictions.csv and you score it with your own script.
How can I verify exact removal or two-copy merge without seeing the code?
Include a poison or audit batch in your training folders, then ask us to remove those 5,000 items (0.149 s) and compare against a clean run trained without them (10,799 of 10,800 identical answers). You can also send two split lesson orders and verify that the merged predictions are bit-identical regardless of order.
What do I get back?
Accuracy on your held-back photos, per category, with the confusions; how accuracy holds for the categories taught first after the later ones are added; the model size and the time to learn a new category, on hardware we report; and a 30-minute review call.
Can I get the code, the model or an API?
No. Our method and code stay ours, and the predictions file is the only output. No script, binary, container, model file or remote access goes out.
What does it cost?
The founding evaluation is $5,000, fixed, paid up front, and credited in full against a pilot or license signed within 90 days. The pilot is $15,000: your categories on your hardware target, with a written integration plan. See both
How long does it take?
A report in two weeks.
What happens to my photos?
They stay yours and are deleted within 30 days of the report. Both sides sign a one-page NDA.
Does it learn on a phone or a wearable?
Not today. Learning runs on a laptop now. Phone, wearable and chip versions need their own build and their own measurements.
What hardware does learning need?
Our current CIFAR-100 reference uses a MacBook Pro M1 Max with 64 GB of memory and its GPU. We report the actual hardware for each evaluation.
Something else? Write to mars@buildcivilization.com.