Measured against 26 LLM configurations.
First on every primary metric.

What this means for an institution's writing

A language checker in a ministry, a newsroom or a bank is judged on four things. Does it flag what is actually wrong? Does it leave correct text alone? Is the suggested fix right? And can the people accountable for the text see why? The L3 evaluation measures the first three and the product answers the fourth.

Precise flags. Of the changes L3 proposed, 84.54% landed on a stretch of text the linguist also changed. That is the number a reviewer feels: a queue of suggestions where most are worth accepting. Measured.

Few false alarms, and few misses. L3 was the only system in the field with precision above 80% and recall above 60% together. The large language models were either cautious and incomplete or thorough and noisy; none managed both on this benchmark. Measured.

Suggestions you can accept. Once a span was flagged correctly, L3's replacement matched the reference 89.79% of the time exactly and 96.38% of the time after Arabic normalization. Those rates are conditional on the flag, and we say so wherever they appear. Measured.

Explanations and ownership. Every WriteX correction carries the rule behind it, and the engine deploys on-premises, so nothing has to leave your perimeter to be checked. The evaluation does not score those; they are what the product does. Product capability.

Energy. Measured on an NVIDIA A10 24 GB, L3 used approximately 0.000014 kWh per 1,000 tokens at accelerator scope. Measured. The measurement itself says nothing about hardware requirements; on-premises deployment is a product capability, described above. Product capability.

The headline number

F0.5 is the headline because it matches how institutions experience a language checker. It combines precision and recall but weights precision more heavily, since an unnecessary change has to be reviewed, rejected and sometimes reverted by hand. L3 scored 80.29%; Claude Fable 5.1 was next at 44.68%, a 35.61-point gap. The strongest comparator changes with the metric: Claude Opus 5 was next on F1 at 34.58% against L3's 74.66%, and GPT-5.6 Sol was next on recall at 41.08% against 66.85%. L3 led on all four. Measured.

Few false alarms

The scatter is the chart to remember. Claude Fable 5.1 was the most precise large language model at 64.21%, and its suggestions were usually right, but it found only 20.15% of the reference edits; a document checked with it keeps most of its errors. GPT-5.6 Sol found 41.08% of the edits, the most among the models, at 25.03% precision; a document checked with it comes back with more changes to reject than to accept. Most other configurations sit close together in the lower left. L3 is the one point in the upper right. Measured.

For a writing team this is the difference between a tool people keep on and a tool they switch off. Reviewers tolerate a missed error more readily than a stream of wrong flags, which is exactly the asymmetry F0.5 encodes.

Suggestions you can accept

Flagging the right span is half the job. L3's replacement matched the linguist's reference byte for byte on 89.79% of correctly flagged spans, and on 96.38% after normalization, which removes diacritics and tatweel, the elongation character, and unifies alef, hamza and teh marbuta variants so a correct suggestion is not marked wrong for one vowel mark. Claude Fable 5.1 was close on this conditional measure, at 89.07% and 94.54%, but on a much smaller set of flags. High conditional accuracy with low recall means a polite tool that misses a lot; the two numbers have to be read together. Measured.

Whole sentences, not just spans

An editor receives a sentence, not a list of spans. On the strict end-to-end test, where the whole output must equal the whole reference, L3 matched on 55.96% of sentences, double the 27.98% of Claude Fable 5.1. The test is unforgiving by design: one extra diacritic or one missing period fails the sentence. It is a summary of the whole pipeline, not an error rate. L3 also kept the sentence intact while correcting it, scoring 98.84% on ROUGE-1 F1 against the reference, ahead of Claude Fable 5.1 by 2.28 points. Measured.

What the corrections look like

Four benchmark rows, verbatim: the source, the linguist's reference, L3's output and the outputs of several large language model configurations. Three are outright errors; the fourth is an editorial standard, where the source is grammatical and the reference applies the house convention. The models were told not to rephrase, so that row shows what the engine's editorial layer does, not a grammar failure, and the caption says so. Where a model matched the reference, the card says so too.

Speed

L3 delivered 52.28 standardized output tokens per second at 1.041 seconds per timed sentence, second to Qwen 3.8 27B on Groq at 66.98 tokens per second. Claude Fable 5.1, the strongest model on quality, ran at 15.07 tokens per second and 3.585 seconds per sentence. The figure is service throughput, measured end to end with request overhead included, and it is what a writer waiting for the sidebar experiences. It is not the batched, saturated rate used for the energy measurement, and the two are not comparable. Measured.

Energy, and why on-premises is realistic

We measured L3 on an NVIDIA A10 24 GB: at least 95% utilization, a 30-second warmup, a 10-minute steady state, power sampled at 1 Hz, approximately 100 W mean GPU power at approximately 2,000 tokens per second. That is 0.05 J per token, or approximately 0.000014 kWh per 1,000 tokens at accelerator scope. Multiplying by a three-times factor for host, cooling, networking and facility overhead gives a full-stack estimate of 0.000042 kWh. Measured.

The comparison figures are cited anchors, because API providers publish no power telemetry: Google's Gemini Apps disclosure gives 0.0012 kWh per 1,000 tokens, and the IFP School per-query estimates, scaled the same way, give 0.0061 for GPT-4o-mini, 0.074 for GPT-5 medium and 0.141 for GPT-5 high. On that basis L3's full-stack estimate is approximately 29 to 3,357 times lower. The result depends on the overhead factor, so the report shows the two-times to four-times range; across it, L3 stays at least 21 times more efficient than the lowest anchor. The practical point for institutions is narrower than it sounds: the measurement itself ran on an NVIDIA A10 24 GB, and WriteX offers on-premises deployment. Measured for L3; cited estimates for the anchors; on-premises deployment is a product capability.

Inside WriteX

The engine in this evaluation is the one behind WriteX's Arabic language checking in the editor, the Office add-ins and the browser extensions. Product capability. To see how it surfaces in the product: language checking covers the error categories and the rule behind each card, the technology explains the specialised components that make the engine explainable and fast, deployment covers cloud, sovereign cloud and on-premises options, and WriteX against LLM chatbots puts these numbers beside the workflow differences. The full report on lisan.com carries every table, the citation and the version history.