Real runs, real numbers.
Everything below is actual output from the engine that powers the tool, reproduced with its measurements. Nothing was hand-polished afterward.
English: a productivity blog post, score 16 to 87
| check | before | after | |
|---|---|---|---|
| AI-inflation vocabulary | 12 | 0 | fixed |
| stock AI phrases | 2 | 0 | fixed |
| participle tails | 3 | 0 | fixed |
| summary closers | 2 | 0 | fixed |
| significance inflation | 2 | 0 | fixed |
| em dashes | 1 | 0 | fixed |
| sentence-length stdev | 3.3 | 2.4 | still failing |
Diff excerpt. Deletions struck through, insertions on green. The one check still failing after the rewrite is right there in the table: honesty over polish.
العربية: منشور تقني، من 73 إلى 96
| الفحص | قبل | بعد | |
|---|---|---|---|
| العبارات الاحتفالية الجاهزة | 3 | 0 | أُصلح |
| علامات لاتينية / تطويل | 2 | 0 | أُصلح |
| الترقيم الآلي (إرشادي) | 1 | 0 | أُصلح |
مقتطف من الفرق. لاحظ الربط بـ«ثم» و«و» بدل الترقيم الآلي:
The whole eval suite, since one example proves nothing
The repo carries a committed evaluation set: 20 texts across genres and lengths in both languages, run end to end against every engine and prompt change. Latest run, on the engine this site actually serves: mean gain +24.1 points, 17 of 20 texts improved, zero run errors. The worst English fixture went 16 to 87; Arabic fixtures finished between 90 and 98. Two drafted candidates failed the faithfulness guard mid-run and were discarded before they could win, which is its whole job.
Three texts did not beat their originals, and the tool said exactly that instead of shipping them quietly. That honesty is a feature, not a bug to file.