Good research starts
with questioning the data.
Catch missing values and duplicate IDs. Let a small, local anomaly model suggest unusual readings. Keep the final decision—and a record of why.
Header required · max 1,000 records / 500 KB. No upload, storage or execution of your pasted data.
Different group/unit combinations get separate models. Only one numeric measurement is modeled at a time.
Every flag needs a reason.
A high score is a suggestion to investigate, not proof of an error. Choose Keep or Exclude and explain each flagged record before exporting.
No audit yet.
Load the synthetic example, then run the audit. It includes deliberate duplicates, missing values and one unusual reading.
| CSV record / ID | Group · unit | Reading | What to inspect | Your decision | Reason |
|---|
Exports stay locked until the review is complete. Downloaded files contain your data; share them deliberately.
A model is a second look.
You are the reviewer.
One-feature Isolation Forest fits 64 random partition trees within each group and unit. Short isolation paths yield higher anomaly scores. Groups need eight numeric readings. The fixed 0.65 threshold is a prototype heuristic—not a calibrated error probability. It can miss errors and flag valid extremes. Every keep/exclude decision remains visible in the audit.