1 Data
Drop CSV files here or
Up to 6 files, 10 MB each. They are read only inside the sandbox.
Demo data is synthetic (seed 20261103) with planted leaks and an answer key. No real customer data.
2 Question
3 Red team
Five hostile scripts run in the same sandbox your questions use. Each card shows the limit that stopped it.
How a question becomes a number you can trust
- Plan and code. on Vultr Serverless Inference plans the checks and writes pandas code.
- Contained run. A fresh gVisor container on our Vultr VM: no network, no secrets, read-only inputs, . Destroyed after the run.
- Retry. Errors go back to the coder, three attempts at most.
- Verify. , a different model family, checks every sentence against the raw output; a sentence with a number not in the output is cut too.
- Receipt. Hashes of inputs, code, image and outputs. Re-run it and get the same numbers.
Pick a question. The answer arrives with the code that produced it, the table, and a receipt.
Red team
Contained
Answer verified sentence by sentence
What the verifier checked ()
Code that ran written by the model, untrusted
Act on it needs a person; nothing leaves this page
The export escapes spreadsheet formulas (cells starting with = + - @). Drafts are text for you to edit and send yourself.