Build a Human Review Queue from Classification Results
If your output has a suggested label and a score already normalized to 0–1, use this page to practice deciding what should reach a person first. A missing suggested label, or a missing, out-of-range, or below-threshold score goes to review; the remaining suggestions still need spot checks. This tool does not calibrate scores or call them accuracy.
Files are read in this browser only. Nothing is uploaded to JevLog or sent to a model. Refreshing clears the current result.
Choose a file or load the demo sample.
The spreadsheet CSV adds a prefix to formula-like cells and may change report values. Download the JSON to process exact report values in another program.
Download the fixture and expected queue
Section titled “Download the fixture and expected queue”The input CSV uses id, label and confidence. At threshold 0.8, review IDs 2 (0.60), 3 (missing score) and 5 (missing label); keep IDs 1 and 4 for spot checks. Compare the expected full JSON and expected queue CSV. Verify 5 = 3 + 2.
At threshold 0.6, ID 2 equals the threshold and remains for spot checks; the queue contains two records. Missing scores and labels still need review. These are deterministic teaching results, not model-performance evidence.
Frequently asked questions
Section titled “Frequently asked questions”Are failed requests automatically detected?
Section titled “Are failed requests automatically detected?”The tool does not inspect status/error fields. A real adapter needs a separate failure queue so stale labels or scores cannot pass into spot checks.
Can I save human edits online?
Section titled “Can I save human edits online?”The current generator does not offer online review storage or team assignment. The queue preserves input columns and adds jevlog_review_reason. Store final labels, reviewer identity, reasons and time separately, as shown in the handoff tutorial.
Is the queue retained after refreshing?
Section titled “Is the queue retained after refreshing?”No. Download the full report and queue first, then reload the input and threshold to reproduce them. No login is needed; files are not uploaded and no external actions are executed.
What to keep for the reviewer
Section titled “What to keep for the reviewer”The queue export preserves every original input column and appends a review reason. The on-page table shows the ID, suggested label, score, reason, and a text, feedback, or message column if present. Record the human’s final label, reason, time, and identity separately. Do not overwrite the initial suggestion. This output does not message customers, approve refunds, or execute external actions.
The fictional sample uses 0.8 only as an illustration: three of five records need review (one low score, one missing score, and one missing label), while two remain for spot checks. Choose a real threshold with your own human-reviewed data, then reassess it when rules or the input population change.
Continue with the human review guide, run comparison tool, or CSV preflight checker.