Flagship Benchmark · Live · v1.0

Traditional Chinese Quality Benchmark

Measures how clean a tool's Traditional Chinese output is against Taiwan usage norms: simplified-character residue, mainland-China wording, and terminology consistency. Fixed 24-segment corpus (marketing, UI strings, game dialogue, customer support), deterministic scoring script, raw outputs archived per run.

Latest results — run 2026-11-21

Tool / EngineScoreScript purityWordingTerminologyVersion testedDate
Gemini99 (Grade A)100.097.4100.0gemini-3.5-flash-lite2026-11-21
Microsoft Translator98 (Grade A)99.094.8100.0web UI; model not exposed, recorded by access date2026-11-21
Groq96 (Grade A)100.091.897.2openai/gpt-oss-120b2026-11-21
DeepL95 (Grade A)99.092.591.4not exposed by web UI; recorded by access date2026-11-21
Google Translate89 (Grade B)99.082.383.3not exposed by endpoint; recorded by access date2026-11-21

Conditions disclosure (part of the measurement)

AI/LLM engines received the corpus with one shared neutral prompt ('Translate into Traditional Chinese as used in Taiwan' plus fixed formatting instructions). Machine-translation engines received the source text directly, with no prompt. The groups were measured under different conditions; this disclosure travels with the scores wherever they appear.

Supplement & Bing condition

Supplement (same run, 2026-11-21 00:52-00:57 CST): DeepL and Microsoft Translator (Bing) were measured afterwards through a real-browser route - free web translators, not logged in - after their scripted channels proved unreachable; the failed attempts remain archived. Bing condition disclosure: the translator's tone preset was left at its default 'casual' setting, so its output uses the informal address form throughout with colloquial phrasing. The benchmark's three metrics (script residue, mainland wording, terminology) do not score register or tone; read Bing's score with that condition attached.

Seven further first-batch tools are test pending with documented reasons on their pages (account-gated platforms, blocked free channels, rate limits). Full per-segment evidence for this run: run archive 2026-11-21. Method: methodology.

Run history (append-only)

DateBenchmark versionCorpusCoverage
2026-11-21v1.0Corpus v1.05 engines scored

Get the full benchmark report — and every future score — by email

The public pages show the headline scores. The email edition carries the complete dataset behind them:

  • The full per-segment benchmark dataset for every run (all 24 segments, every engine, every flag)
  • New-tool test alerts — the day a pending tool gets its first measured score
  • Quarterly re-test results, including score changes and what caused them

Get the benchmark report →

Free. One list, no spam, unsubscribe anytime. You will receive a confirmation email first.