What it does
Data Integrity Screen reads a research data file (CSV, TSV, TXT, Excel or JSON) and looks for patterns that are unusual in genuine measurements:
- Duplicated records: rows identical in every measurement.
- Near-identical records: rows identical except for one value, the trace of copying a row and editing it.
- Identical or exactly related columns: a column copied, or computed from another.
- Repeated sequences: a run of values that appears again elsewhere, the trace of copy and paste.
- Uneven final digits: invented numbers rarely have evenly spread last digits.
- Preference for 0 and 5: values rounded or estimated by hand.
- Equal increments: values that rise by exactly the same step, as filled-in data do.
- Values on a common step: data coarser than their decimals suggest.
- Outlying values: a data-quality check.
- Benford's law: first digits, only for data spanning three orders of magnitude.
- GRIM: reported means that no set of integer data can produce.
It produces a report that lists every finding with its exact cells, its statistic and the ordinary explanations to rule out. The report records the file's SHA-256, so anyone can reproduce it.
A pattern is not proof
The screen finds patterns worth checking against the raw records, instruments and notebooks. Every pattern has ordinary explanations, such as a record exported twice, a derived column, an instrument's resolution or values read by eye. These are listed with each finding and should be ruled out before any conclusion is drawn.
Validation
- Statistics: the χ² and binomial tail probabilities and the Benjamini–Hochberg adjustment match SciPy.
- False alarms: among 3,000 simulated clean datasets of six kinds, 0–2% got a finding to review. The six kinds were normal measurements, log-normal values recorded to 3 significant digits, counts, replicated designs, random-walk time series and a 0.5-resolution instrument.
- Detection (500 datasets each): duplicated rows 99%, pasted runs of 5 values 100%, runs of 6 equal increments 100%, a column with invented final digits 100%, a copied column 100%.
- Near-identical rows: 100% in tables with 5 or more measurement columns, but rarely in tables with 3–4, where such pairs also occur by chance.
- GRIM: no false alarm over every true mean of integer data with n < 100.
The scripts and results are in the site repository (scripts/integrity/simulate.mjs, scripts/benchmarks/integrity-2026-10-14.md).
Before you start
- The file is read in your browser and never uploaded. Only the optional AI summary sends anything, and it sends the finding descriptions, not the data.
- Put one observation per row and one variable per column, with a header row. Several sheets in an Excel file can be screened one at a time.
- Numbers in text files are read exactly as written, so the digit tests see the recorded digits. Spreadsheet numbers are read as stored; a display format such as "0.00" does not change them.
Quick workflow
- Open Data Integrity Screen and choose a file (up to 50 MB).
- Under 02 · Columns, check the selection:
- Measurement columns are screened.
- Index columns (1, 2, 3 …) and factor codes (treatment, replicate) are left out, because their regular patterns are by design.
- Tick or untick to change the selection.
- For a table of reported means, open Check them with GRIM and choose the mean column, the n column and the items per participant.
- Under 03 · Settings, keep the validated defaults or change the review level and run lengths, then select Run the screen.
- Read the findings. Show cells displays the rows involved with the flagged cells highlighted.
- Select Download report (HTML); open it in a browser and print to PDF if you need a PDF. Findings (JSON) and Flagged cells (CSV) give the same results for other software. A parameter record with the file checksums is produced for each export.
- Optional: under 05 · Optional, choose a model and select Write summary for a plain-language summary with questions to ask. It is added to the report, marked as AI-written.
How the checks decide
| Check | Needs review when |
|---|---|
| Duplicated / near-identical rows | The expected number of such pairs by chance, from each column's value frequencies, is below the review level. |
| Repeated sequence | The expected number of chance repeats of that length, over every pair of columns and positions, is below the review level. |
| Equal increments | The expected number of such runs by chance, from the column's own increments, is below the review level. |
| Final digits, 0/5 preference | The q-value (Benjamini–Hochberg over all digit tests in the file) is below the review level. |
| Identical columns, GRIM | Always (they cannot happen by chance). |
| Exact linear relation, common step, outliers, Benford | Reported as notes only: they have frequent ordinary explanations. |
The digit tests take care to read digits correctly:
- Fixed-format columns: where every value is written with the same decimals, zeros are counted.
- Mixed-format columns: where software may have dropped trailing zeros, only the digits 1–9 are used.
- Third significant digit: only digits in the third significant place or later are tested.
- Common step: columns on a coarse step (an instrument resolution such as 0.5) are not digit-tested.
Data, privacy & external services
- Reading, screening and the report happen in this tab.
- The optional AI summary sends the finding descriptions to the provider you choose; the API key stays in the tab's memory.
Limitations
- A clean result does not show that data are genuine: careful fabrication can avoid every pattern tested here.
- Digit tests need at least 50 usable values per column; small tables are screened for duplication and steps only.
- Rows are compared in file order; sorting the data before export removes the order that repeated-run and step checks use.
- Near-identical rows cannot be judged reliably in tables with fewer than 5 measurement columns.
Troubleshooting
- A measurement column was detected as a factor: it has few distinct values; tick it to screen it.
- Many findings in a column of manual readings: 0/5 preference and uneven digits are expected for values read by eye; note this in the review.
- A derived column is flagged: an exact linear relation (unit conversion, percentage) is expected if the computation is documented; untick one of the two columns.
- The file cannot be read: save it as UTF-8 CSV, or check that the first row holds the column names.
