A receipt records its inputs inside the hashed preimage, and every kernel is deterministic, so a receipt is already a complete test case. This page walks through how the replay page uses that: it reads your files in memory, re-runs them, and lists the results that differ. Every number in the report is a pair of hashes you can recompute.
The page recomputes each receipt's own hash first, so a file that does not match itself takes no part.
The hash comes from the one shared function the kernels use, so the page has no second way of writing the preimage. Files are read into memory with the browser's file reader and dropped when the page closes. Nothing is stored and nothing is sent.
Which of my kept receipts still match their own hashes?
Condition A uses the published kernel as it is today. Condition C changes the parameter values you name.
The page loads each kernel module from this site on first use and keeps no state between receipts. A node that is not on this site, a receipt whose inputs are not in the file, and a private-input node are reported as unresolved with the reason. They are never reported as the same.
What would last quarter's receipts have said if this threshold were 5 percent lower?
Each line shows the recorded and the new execution hash, and the exact places in the output that differ.
A changed result means the output differs under the stated condition. It does not say the recorded receipt was produced incorrectly; whether an earlier kernel needed correcting is the claim of a published erratum. The report is built in memory and saved only when you press an export button. A selection is capped at 2,000 files and 20 MB, and a larger one is refused before anything is read.
Which of these 400 receipts change under the stated condition, and where?