Reviewer (R1): the first human reviewer, who accepts or rejects a fellow's submission and writes a comment.
Lead audit: a lead (DOL, e.g. Teja, Ovais, Jacob) re-checking the task or the reviewer's call. Stronger than a reviewer alone.
Implicit: no decision is recorded, but the task moved on to Packaging. Weakest.
Human decision: what each person decided (rejected/approved, GOOD/BAD). A commit can have several, from different people (say a rejection and a later audit), not repeated runs.
Why (read by model): people don't tag their reasons, so a model read their words to say whether the problem was cheating. It names the model.
Gold label: the final answer, made from those two columns. The decision says good or bad; the model's reading says whether "bad" meant cheatable. E.g. "rejected" + "comment describes a shortcut" = Cheatable.
Gate said: the current QC's verdict on that same commit. Red means it let a cheatable commit through.
Click a row to read the full reviewer comments, what the model concluded, and links. ≈ after a commit means it was guessed from the PR history.
What the gate said, by gold label
The gate's verdict on each commit (the one the reviewer saw, or the latest run when that wasn't recorded). Click a number to list those commits.
| Commit | Repo | Gold label | Backed by | Human decision | Why (read by model) | Gate said |
|---|
Each chip is one labelled commit, oldest first. Click a chip to see that task's rows in Labels.