Performance Rating Calibrator
Normalize lenient and strict raters, then apply your grade distribution — entirely in your browser.
1. Paste or drop your review scores
You can also drop a .csv or .tsv file anywhere on this box. Tab, comma and semicolon files are all detected automatically.
2. Match your columns
Preview of the first rows, exactly as the parser read them.
3. Choose how to calibrate
Each rater's scores are re-expressed as distance from that rater's own average, so a generous manager and a strict one land on the same scale.
Review type weights
Only the types an employee actually has are used — the weights are re-scaled for each person. Set a weight to 0 to leave a type out.
How far each rater's average sits from the overall average. A long bar to the right means that rater scored more generously than everyone else.
Employees on the same composite score always share a grade, so the actual share can differ from the target.
Nothing is uploaded. Every score is parsed and calibrated inside this page — open DevTools → Network and you will see no request while you calibrate. Only your preset, column mapping and weights are kept in this browser.
Frequently asked questions
What is z-score normalization, and why correct each rater separately?
Two managers rarely mean the same thing by "4 out of 5". Research on appraisal calls this the hawk-dove problem: doves inflate, hawks deflate, and the employee's grade ends up depending on who happened to review them. Rater bias normalization fixes the scale before the ranking. This performance review calibration tool converts every score into Z = (X − μ) / σ, where μ and σ are that rater's own average and standard deviation. A 4.5 from a manager who gives everyone 4.6 becomes slightly negative; a 4.0 from a manager whose team averages 3.2 becomes strongly positive. Rank on the z-scores and you are comparing people, not their reviewers.
Why does the forced distribution differ between companies and countries?
Because it is a policy choice, not a statistic. GE made stack ranking famous and then dropped it in 2015, and much of the US now uses a soft 20 / 70 / 10 guideline or no grades at all. Many Korean and Japanese employers still run a firm relative curve — 인사평가 등급 배분 with S/A/B/C/D quotas per department. So this forced distribution calculator ships presets instead of a hard-coded curve: pick a 5-grade or 3-grade preset, or type your own labels and percentages. Nothing in the maths assumes a bell curve — the tool ranks and cuts wherever you tell it to (상대평가 등급 산출).
What happens when people tie right on a grade boundary?
Everyone with the identical composite score is kept in the same grade, and the tie is resolved upward into the higher grade. That means the actual share can overshoot the target — the distribution tab shows both, for example "target 10% / actual 12%". We do this on purpose. Splitting identical scores to hit a quota produces a grade you cannot explain to the person who lost the coin flip, and in a calibration meeting explainability beats a tidy percentage. If a tie block is too large, that is a signal your scores lack resolution, not something the tool should hide.
How do I verify that our review data is never uploaded?
Check it yourself, do not take our word for it. Open DevTools (F12) → Network tab, clear it, then paste your scores and hit Calibrate: the request list stays empty, because the parsing, the z-scores and the grade assignment all run in JavaScript in this tab. You can also load the page, go offline (Network → Offline, or unplug), and calibrate — it still works, which is only possible if nothing leaves the browser. The page keeps your preset, column mapping and weights in localStorage; employee IDs, names and scores are never written to storage and never sent anywhere.
Can we treat the normalized grades as final?
No — this is preparation for a calibration session, not a verdict. Z-scores assume a rater's own ratees are a fair comparison group, which breaks down for very small teams, for a genuinely outstanding team, and for anyone rated only once. The tool flags those cases instead of hiding them: raters with a single ratee or zero score variance are left out of normalization, and every excluded row is listed with its line number. Bring the ranking, the rater bias report and the excluded rows to the meeting, let managers argue the edge cases with evidence, and record the human decision as the outcome.