Products

Something you can check.

The arithmetic on a card. The behaviour of a code change. The record of a training run. Each begins with the same question: did it do what it said?

Live

For people running
GPU fleets

Card-check

Check the answer, bit for bit.

Run one fixed training step on your GPU and compare its result with a recorded reference for the same hardware and software. A match means the bytes agree. A difference points to the first tensor that differs.

The published runs took about 12 seconds on an NVIDIA L4 and under 10 seconds on an L40S or H100, including model construction and hashing. Once installed, the check runs without a network connection.

Request card-check

Access is approved on Hugging Face for evaluation. The runner, reference files, and terms are there.

What a match tells you

The reference pins the card model, driver, PyTorch build, SM count, and MIG mode. The check compares the step’s parameters, optimizer moments, and loss, with separate hashes for its tensors.

This checks the arithmetic exercised by that workload; it does not certify every path in a GPU. Keep your memory, thermal, and link checks. An unrecorded configuration returns NO-REF, which is not a hardware failure.

Live · Free sample

For coding-agent
builders and researchers

Reed

What a code change actually did.

Five hundred executed behavioural records from 141 open-source repositories. Each records a real code change, its environment, the small test programs run before and after it, and the outputs they produced.

Use the sample to study or evaluate code behaviour. Follow a row back to its source, rebuild its environment, and replay the checks with git and Python. You do not need Reed’s own pipeline to examine the result.

Get the Reed sample

Free sample; request access on Hugging Face. The schema, reproduction guide, and licensing are included.

What the sample covers

There are 493 Python records and 7 JavaScript records. Every row was selected because its probes observed a change; this is not a representative sample of all pull requests.

A record also identifies the smallest set of patch hunks found to cause that observed change. Its conclusions cover the supplied probes, not every possible behaviour of the program.

Research

For people asking
how a model was trained

Training receipts

A claim you can go back and examine.

We are developing cryptographic proofs of training steps under exact integer rules. Each receipt commits to a step’s computation and links its result to the next. The evidence page collects selected recorded experiments and the boundaries of what they show.

See the evidence

A research record, not a production training-audit service.

Bring a question from your own machines.

A fleet to check, a repository to understand, or a result you need to reproduce.