Replay an incident you already solved
Any vendor can quote an accuracy number about themselves. This is the other thing: bring one failure you already debugged, tell us what it turned out to be, and watch the engine work from the evidence you have without that answer. One log is enough to start. Then read the score.
Your answer is sealed first
It is hashed and the hash is handed to you before the log field unlocks.
Only incident evidence is sent
The diagnosis request carries only what you submit in the second step. Your sealed cause and fix stay out of that request.
You can recompute it
The receipt publishes the salt, so the hash is yours to verify.
What a miss looks like
A wrong answer is recorded as wrong, in those words. If the engine declines to name a cause because the log does not support one, that is reported as declining, not quietly counted as a win. And if you did not tell us how long your investigation took, no time saving is shown at all. We would rather lose the demo than publish a number you can take apart.
Step 1. Record what the cause actually was
We hash this and hand you the hash before you paste anything. That is what proves the engine could not have seen it.
Leave it blank and we will not show a time saving at all. We do not substitute an average for a number you did not give us.
Step 2. Paste what you have from the incident
One error log is enough to start. If you have other evidence, add it below. Only what you submit in this step reaches the engine, never the cause or fix you sealed above. Redact secrets before submitting; keep timestamps and rank or GPU IDs when you can.
Have more incident evidence? Add it here (optional)
Useful examples: an earlier error from another rank, scheduler exit reason, Xid or ECC event, or a short metrics timeline. Do not paste the confirmed cause or fix here. Leave this blank if you only have one log.
One replay is a demonstration, not a measurement
A single scored incident tells you how we did on that incident. Our published accuracy methodology, its sample sizes, and what it does not yet prove are on the evidence and evaluation status page, and the reproducible benchmark corpus is at the benchmark page.
See what a live cluster adds to this