grammar-v7d — MyDeskMate English Grammar Checker (run #4, PASSED)

T5-base (220M) grammar correction model, continued from grammar-v7b lineage with a ~12% dose of exam-style calibration data (the "dose ladder" sweet spot).

Pre-registered promotion record (run #4, 26–29 Sep 2026 GPU + local verify)

Promotion rule (written BEFORE the run): keep only if (CoNLL > 47.35 or BEA > 33.90) AND JFLEG GLEU >= 62.30 AND Do-No-Harm < 5%. All four gates passed:

gate	measured	required	verdict
JFLEG GLEU (official)	62.73	>= 62.30	PASS
CoNLL-2014 F0.5 (m2scorer)	49.04	> 47.35 (live v5)	PASS
BEA-2019-dev F0.5 (ERRANT)	35.41	> 33.90 (live v5)	PASS
Do-No-Harm (clean wikitext)	2.80%	< 5%	PASS

Live serving path (ONNX Runtime int8, greedy, this repo's graphs)

Measured through the exact production engine on CPU (threads=2):

metric	value	target
JFLEG GLEU (official)	62.61	>= 60
CoNLL-2014 F0.5 (official m2scorer)	49.34	>= 65 (open)
BEA-2019-dev F0.5 (official ERRANT)	34.61	>= 70 (open)
Do-No-Harm (500 clean, ship probe)	4.20%	< 5%
Over-correction (2,000-sentence wikitext-2 protocol)	5.05% (2000 sentences)	< 5% — missed by 0.05 pp, published as measured
latency, live HTTP (prod thread pin)	102 / 95 ms mean/median	< 200 ms
latency, in-process ort_bench run	188 / 162 ms mean/median	reference

• Evidence: logs/v7d_live_latency.log (live HTTP), ort_bench_v7d.json, production_benchmark_v7d.json (battery), logs/v7d_ort_suite.log (probe).

Head-to-head (Phase 1, all systems measured in OUR lab through OUR official

scorers — published with raw JSON at mydeskmate/gec-bench)

system	JFLEG GLEU	CoNLL F0.5	BEA-dev F0.5	Do-No-Harm
grammar-v7d (this model)	62.73	49.04	35.41	2.80%
LanguageTool 6.6 (official)	47.57	25.60	20.35	98.80% FAIL
GECToR (bea-dkt+conll, kc=.3)	53.94	56.94	49.52	10.40% FAIL

Our row above = the pre-registered GPU PyTorch promotion run (beams=4). The live ONNX int8 serving run of the same model reads 62.61 / 49.34 / 34.61 with the 500-sentence probe at 4.20% — see the "Live serving path" table above. Both runs are published; competitors were measured once each through our scorers with their default settings.

• We beat LanguageTool on every axis (fluency, exam tags, and 35x safer).
• GECToR wins the two narrow academic-exam tag sets, but touches 10.40% of clean text (fails our <5% do-no-harm gate) and loses fluency by 8.7 GLEU.
• Honest note: GECToR's own README quotes CoNLL 63.2 / BEA-dev 57.6 (different pipeline/thresholds). We publish our-measured numbers on the same yardstick for every system.

Open targets (unbeaten rows)

CoNLL-2014 >= 65 and BEA-2019 >= 70 remain open — that is the Phase 3 GPU push (synthetic-data absorption from mydeskmate/gec-lab:gec_synth_500k.jsonl).

Card generated 2026-10-19 11:02 UTC.