| Never damages correct text (500-sentence live test) | 4.20% touched | limit < 5% | ✅ |
| Speed per sentence (live engine, HTTP, mean / median) | 102 / 95 ms | < 200 ms | ✅ |
| Same input → same output | identical | must be identical | ✅ |
| Over-correction (2,000-sentence protocol) | 5.05% touched | < 5% | ❌ over |
| No offensive words introduced | 0 | 0 | ✅ |
| Full paragraphs (512+ tokens) — nothing dropped | 515 → 515 | no drop | ✅ |
| JFLEG fluency benchmark (GLEU) | 62.61 | ≥ 60 | ✅ |
| CoNLL-2014 academic exam (F0.5) | 49.34 | ≥ 65 | ❌ open |
| BEA-2019 advanced exam (F0.5) | 34.61 | ≥ 70 | ❌ open |
Live engine v7d (ONNX int8) scored 19 Oct 2026 with the official task scorers (m2scorer / ERRANT / GLEU) through the exact serving path — same engine your paste runs on. The over-correction row misses by 0.05 pp (101/2,000) and is shown as ❌; the 500-sentence ship probe passes at 4.20%. The two open rows are PhD-level exams best solved by GPU-scale models — we publish everything exactly as measured. 📊 Full benchmark report →
| checker | fluency JFLEG GLEU |
exam CoNLL F0.5 |
exam BEA-19 F0.5 |
correct text damaged |
| MyDeskMate v7d (this engine) | 62.6 | 49.3 | 34.6 | 4.2% ✅ |
| LanguageTool 6.6 (the big open-source one) | 47.6 | 25.6 | 20.4 | 98.8% ❌ |
| GECToR (academic benchmark model) | 53.9 | 57.0 | 49.5 | 10.4% ❌ |
Beats LanguageTool on every axis. GECToR scores higher on the two narrow academic exams, but it breaks 1 in 10 already-correct sentences (our ship rule: <5%) and is 8.7 points worse at real-world fluency — so it can't ship behind a "never damages your text" promise. Our own safety numbers, shown both ways: 4.20% touched on the 500-sentence ship probe (pass), 5.05% on the bigger 2,000-sentence protocol (the ❌ row above) — still 2× safer than GECToR and 24× safer than LanguageTool. All three systems were run by us through the official scorers (m2scorer / ERRANT / GLEU); raw JSON is public: huggingface.co/mydeskmate/gec-bench