Brașov is the correct Romanian spelling: its second letter with a diacritic is ș, a lowercase s with a comma below. Braşov uses ş, an s with a cedilla. They can look alike, but they are different Unicode characters—and a benchmark by Daniel Butnaru reports that language models often copied the cedilla form when it appeared in their input, even though they produced the correct spelling from clean text.
Which spelling is correct: Braşov or Brașov?
In Romanian, write Brașov. The Romanian Academy describes the marks on the letters ș and ț as a comma below (virgulița), not a cedilla. Its explanation explicitly distinguishes the comma from the cedilla used in older works (Academia Română language guidance).
As an Amazon Associate I earn from qualifying purchases.
The distinction is also present in the underlying text, not just in the way a typeface draws the characters:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Brașov contains ș, Unicode U+0219: Latin small letter s with comma below.
- Braşov contains ş, Unicode U+015F: Latin small letter s with cedilla.
Because these are different characters, exact text comparisons and searches can treat the spellings as different even when a reader sees nearly identical shapes. In Romanian prose, the Academy’s guidance supports using the comma-below form.
#1 Best Overall
Why did the models copy the cedilla spelling?
The benchmark’s central observation is a context-copying effect, not proof that the tested models were inherently unable to distinguish the two characters. Butnaru reports that all tested models retained the correct letters in 45 runs when the input text used the correct spelling. When the input contained cedillas, however, models often reproduced them instead of correcting them.
In the author’s reported results, clean-answer rates on cedilla input ranged from 44% to 85% across the models. The rate also varied by task: replies were clean 97% of the time, summaries 62%, and retrieval-style answers 45%. Butnaru says 93% of the incorrectly spelled words—1,105 of 1,186—were copied verbatim from the input. These are results from that benchmark’s specific prompts, model versions, proxy, and setup, not estimates for every model or real-world deployment.
What the 15-model benchmark tested
Butnaru describes six kinds of tasks: generating Romanian text from ASCII place names; echoing or summarizing correct and cedilla-spelled input; proofreading with different instructions; comparing Unicode character sequences; restoring diacritics from context; and writing a Romanian text-fixing function. The author says scores were calculated by comparing Unicode code points rather than using an LLM judge. Calls ran through Kaggle’s model proxy with its defaults, and the echo task was repeated three times per model.
The article names 15 models from Anthropic, Google, OpenAI, and other providers. Those names belong to the dated benchmark and should not be read as a current market ranking: model versions and availability change. In a generation task using English prompts and ASCII place names, Butnaru reports that 11 of 15 models wrote every Romanian word correctly. The contrast with cedilla-containing inputs is consistent with the article’s broader finding: the input spelling mattered.
Explicit diacritic instructions improved correction
The benchmark also compared proofreading prompts. Butnaru reports these clean-correction rates for cedilla input:
| Instruction | Reported correction rate |
|---|---|
| “Proofread the following text” | 79% |
| “Correct the spelling and the diacritics” | 99% |
| An explicit explanation of the cedilla issue | 97% |
In a separate follow-up, adding a Romanian orthography rule produced clean-answer rates from 64% to 96%, with ten of the 15 models perfect on that task, according to the author. These figures are specific to the reported benchmark; they do not establish how a different model, version, or production interface will behave.
Why the figures need careful interpretation
This is one author’s benchmark, not an independent replication or a population-wide measurement. Its results describe the named models and the tasks, prompts, proxy behavior, and run configuration in Butnaru’s write-up. They should not be used to rank current commercial models without checking versions and reproducing the tests.
The same caution applies to the article’s web-text observation. Butnaru says 30 of 35 Romanian website front pages contained cedilla forms; the pages were fetched on 30 September 2026. That is a small, dated sample, not a representative estimate of Romanian websites or the web as a whole.
Character-comparison results also need context: the author notes that an identical-string control changed how one task should be interpreted. That caveat reinforces the distinction between observing a model’s behavior in a particular test and making a general claim about what models can or cannot recognize.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to avoid cedilla carryover in text and software
Make the requested spelling explicit
If you want Romanian orthography corrected, say so directly—for example, ask the model to correct spelling and diacritics using Romanian conventions, with comma-below ș and ț. The benchmark suggests that a generic proofreading request may be less reliable than an instruction that names diacritics.
Normalize known legacy forms carefully
For software handling Romanian text, a pipeline can normalize known legacy cedilla forms to the Romanian comma-below characters. Unicode normalization such as NFC can help ensure equivalent composed and decomposed forms are handled consistently, but it does not by itself turn U+015F into U+0219. A translation step for those specific legacy characters may be appropriate when the text is known to be Romanian.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Test the implementation with both composed and decomposed input, and scope any replacement to Romanian text or a clearly defined data field. Do not indiscriminately remove or replace diacritics: they can distinguish words, and cedillas are valid in other languages. Romania’s Ministry Order 414/2006 addresses standardized character encoding for electronic documents and applies requirements to public authorities, public institutions, and notaries; it should not be presented as a rule binding every private publisher (Government of Romania, Ministry Order 414/2006).
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
