Labels

The Labels page is the review queue for content labels. Jev labels each chunk of public content an agent reads, such as invoice or payment_fraud. Here a person checks those labels, so you can tell how often the detector is right before you trust its flags.

Review a chunk

1. To review. Chunks nobody has reviewed yet, least sure first, then newest. Each row shows the start of the chunk, its origin, the label Jev picked and how sure it was.
2. AI fallback. When Jev picks none, the worker asks your own model to label the chunk. It picks a fixed label if one fits, or suggests a new short one with a one-line reason. It shows Waiting until the worker gets to it, and Skipped when the worker has no OPENAI_API_KEY.
3. Review. Opens the whole chunk, its likeliest labels and where it came from. The right label starts at the fallback’s label when there is one, and at Jev’s otherwise. Click Approve to keep it, or pick another label and click Correct. Quard saves who reviewed it.

Chunks are stored as Jev got them: secrets removed, and emails, IBANs and card numbers masked. Only public content is labeled and stored this way.

Read the counts

Per label counts, for each label Jev picked, the chunks still waiting, the chunks reviewed and how often a review kept the label. For risky labels, Right when flagged shows how often a review agreed when the chunk was risky enough to be flagged. Check the thresholds on at least 200 reviewed examples per risky label.

Reviewed lists the latest reviews, with what Jev said, the label a person chose and who chose it.

A review never changes a guard decision that was already made. See Labels and trust for how detector labels flag content.