The Patronymic Rule Everyone States Wrong

Russian civil registry offices publish the rule for building a patronymic. Here is the clause that decides between -ьевич and -иевич, as it appears on a regional registry office's own website:

The и turns into ь after a single consonant or the group нт. The и stays after к, х, ц, and also after two consonants (except the group нт).

It is compact, it is quotable, it is reproduced on school sites, style blogs and other registry offices, and it is the version that ends up in code. It is also wrong in three separate ways at once, and not one of them will ever show up as a crash, an exception, a failed assertion or a broken character.

Our Russian given-name dataset holds 316 male names carrying a combined weight of 1,699. Fifty-nine of them end in -ий — 18.7% of the entries and 18.8% of the weight. Roughly one father's name in five goes through this branch. Below is what happens to those 59 names, what the reference dictionaries actually say about them, and why a rule can be wrong for decades without anything breaking.

What the corpus looks like before the rule fires

The -ий column is the only one with a genuine decision in it. Everything else is a suffix and a truncation. This one requires you to know something about the consonant sitting in front of the ending — and that is where the received rule breaks.

The 18 names that decide the question

Of the 59 names in -ий, 43 have a single consonant before the ending (Васил-ий, Юр-ий, Анатол-ий) and 16 have a consonant cluster. The interesting cases are the 18 names where either the cluster or the final letter itself does work. Here they are, all of them, with the weight our dataset assigns each:

The full cluster inventory across those 16 cluster-final stems is small enough to print in one line: нт × 8, мп × 2, рг, рк, тр, кл, вд, фр × 1 each.

Look at that distribution for a moment, because it explains everything that follows. Ten of the sixteen clusters begin with a sonorant (л м н р). Eight of those ten are the single cluster нт. That is 80% of the sonorant cases covered by one two-letter string — which is exactly the coverage profile that lets a wrong rule survive indefinitely.

Where the received rule actually fails

1. It names a cluster where it should name a principle

The published rule carves out нт by name. It is not hard to see why: Лаврентьевич, Иннокентьевич, Терентьевич and Викентьевич are common enough that anybody writing the rule down would notice them, get an ugly result from "two consonants → -иевич", and bolt on an exception.

But нт is not the phenomenon. It is the most frequent instance of the phenomenon. The phenomenon is that the contracted -ьевич form remains available when the cluster starts with a sonorant, and stops being available when it starts with an obstruent. нт (sonorant + obstruent) and мп (sonorant + obstruent) behave the same way; тр, кл, вд, фр behave the other way.

Because the rule enumerates instead of generalising, it puts Харлампий and Евлампий on the wrong side of a line it drew itself. And it does so in both directions at once, which is the part worth pausing on:

• For Харлампий the rule tells you Харлампиевич and gives no hint that Харлампьевич exists.
• For Иннокентий the rule tells you Иннокентьевич and gives no hint that Иннокентиевич exists.

Both hints are wrong, and they are wrong for the same class of names. The Russian Academy of Sciences orthographic dictionary records both forms for both names.

2. Its list of letters is one letter short

The rule says the и stays after к, х, ц. The reason it stays is purely orthographic: Russian spelling has no such sequences as кь, хь, ць, so -ьевич is not a form you could write even if you wanted to. Стахий cannot give Стахьевич because that string is not Russian orthography, not because of any rule about patronymics.

But гь is exactly as impossible as кь, and г is not in the list. Neither are the hushing consonants ж ч ш щ, after which a soft sign is equally unwritable in this position.

Here is the interesting part: in our corpus that omission never fires. We hold exactly one name ending in -гий — Георгий — and its stem happens to end in the cluster рг, so the "two consonants" clause catches it and produces the right answer for the wrong reason. We hold no name ending in -жий, -чий, -ший or -щий at all.

So the defect is latent. It is a rule that is wrong in a region of the input space the current data does not reach. Add one name to the corpus that lands there and the rule starts emitting strings Russian cannot spell — with no warning, no error, and no test failure, because the code did exactly what it was told.

3. It has no category for "both", and roughly half the branch needs one

This is the largest of the three, and the one that changes how you should think about the whole problem.

The received rule is a partition: every name goes left or right, one answer each. The dictionaries do not work that way. The RAS orthographic dictionary's list of personal names gives, for a large group of -ий names, two patronymics joined by "and" — meaning both spellings are correct and the only requirement is that a given person's documents be internally consistent. Gramota.ru, the reference service of the Vinogradov Russian Language Institute, states it plainly: from a significant group of names in -ий, patronymics may be formed both by the general rule and by replacing -ий with ь, and Геннадиевич and Геннадьевич, Иннокентиевич and Иннокентьевич, Виталиевич and Витальевич, Евгениевич and Евгеньевич are all orthographically correct.

We checked 27 of our 59 names against dictionary entries — that is 80.6% of the branch by weight. The result:

Totals: 8 names admit only -иевич (weight 74), 2 admit only -ьевич (weight 50), and 17 — 63% of the names checked, 52% of the weight — admit both (weight 134).

The received rule has no way to express the third row. It will hand you one answer for Геннадий and tell you nothing about the other. Where the norm genuinely hesitates, a rule that never hesitates is not more precise; it is less informative.

And the "one consonant" branch has a live counterexample

The rule's most confident clause is "a single consonant → -ьевич". Дионисий has a single consonant, с, before the ending. The rule yields Дионисьевич. The dictionary records only Дионисиевич, and we could not find Дионисьевич recorded anywhere.

That is one name at weight 1 in our data — 0.06% of the corpus. It is also proof that the clause is a tendency wearing the clothes of a rule.

Why the obvious fix is worse than the bug

Suppose you notice the Харлампий problem and reach for the natural repair: if there is a sonorant anywhere in the cluster, keep the soft sign. It is a one-line change and it fixes мп without a hard-coded list.

It also destroys Дмитрий.

Дмитр- ends in тр. There is a sonorant in that cluster — р — it is just in the wrong position. The deciding element is the first consonant of the group, not whichever one you happen to spot. The same patch takes out Онуфрий (фр), Ираклий (кл), Георгий (рг) and Аверкий (рк).

Now count the damage by weight, which is the only count that matters when the data is sampled by frequency:

The patch touches fewer names and does 4.6 times more damage, because Дмитрий alone carries weight 52 — the sixth most common male name in the entire corpus. Fixing a naming rule by inspection, on names you happen to think of, optimises for the wrong quantity. The rare names are where you notice the bug; the common names are where it costs.

Three more rules that are stated too broadly

The -ий split is the biggest trap in Russian patronymics, but it is not the only place where the compact version of a rule is a different rule from the real one.

ё does not disappear — one name loses it

The version you will hear is that ё becomes е when a patronymic is formed: Пётр → Петрович. It is stated as a phonological rule about the letter.

Our corpus contains exactly five names with ё, carrying a combined weight of 63:

One name out of five, 22% of the ё-bearing weight. Пётр loses its ё because the vowel is in the stem syllable that alternates — Пётр / Петра / Петрович — and the alternation is a property of that word, not of the letter. Артём keeps its ё in every form. A rule phrased as "ё → е in patronymics" would produce Артемович, Федорович, Семенович and Парфенович: four wrong forms out of five, at 78% of the relevant weight, from a rule that is true of its own example. Whoever wrote it checked Пётр and stopped. One confirmed example is not a check of the rule; it is a check of the example.

Яковлевич is a fossil, not a rule

Яков → Яковлевич, not Яковович. The extra л is a real historical process — Slavic l-epenthesis, the same one that gives любить → люблю and ловить → ловлю: a labial consonant followed by a jot develops an l between them. Яков + the -jevič suffix produced Яковлевич and the form stuck.

The temptation is to generalise: labial consonant, add л. Our corpus says how expensive that would be. Fifty names end in a labial (б в м п ф), carrying weight 214 — 12.6% of the corpus. Exactly one of them takes -левич.

The other 49 take the plain suffix: Ростислав → Ростиславович, Глеб → Глебович, Иосиф → Иосифович, Максим → Максимович, Архип → Архипович, Вячеслав → Вячеславович. And Лев, which is labial-final too, does something different again — it drops its vowel: Львович.

So l-epenthesis in Russian patronymics is a lexical fact about one word, with a 2% hit rate over the environment it appears to describe. It belongs in a list, not in a rule.

-инична depends on stress, and stress is not in the data

Names ending in unstressed -а/-я take -ич/-ична: Никита → Никитич / Никитична, Савва → Саввич / Саввична.

But four names take a longer feminine form with an extra -ин-:

Илья → Ильинична, Фома → Фоминична, Лука → Лукинична, Кузьма → Кузьминична.

Not Фомична, not Лукична. The conditioning factor is final stress: Илья́, Фома́, Лука́, Кузьма́.

And here is the structural problem. Our corpus stores names as written, with no stress marks — as does every name registry we know of, and as does virtually every dataset anyone will hand you. Фома and Зосима are the same shape in every respect a program can see; one gives Фоминична and the other gives Зосимична. No amount of cleverness over the letters can find the difference, because the difference is not in the letters.

That is not a defect that can be fixed by a better rule. It can only be fixed by an explicit list, and the honest thing is to say so rather than to write a rule that appears to derive what it is actually memorising.

We should add that even the stress account is incomplete: our corpus also holds Сила and Фока, which are end-stressed, and the registry handbooks group them with the -ич/-ична names rather than with Лукинична. We follow the handbooks and we do not have a satisfying explanation for the split.

Why nothing catches this

Every wrong form in this article is a well-formed Russian-looking string. Run the standard checks over Лаврентиевич, Дмитрьевич, Артемович, Яковович, Фомична and Стахьевич:

• Not empty — passes.
• Valid UTF-8 — passes.
• All characters Cyrillic — passes.
• Starts with a capital — passes.
• Ends in -ович / -евич / -ич — passes.
• Length in a plausible range — passes.
• Contains the father's name as a prefix — passes.
• Deterministic: the same input gives the same output — passes.

A test suite built out of those assertions is a hundred percent green on a completely broken rule. The output is not malformed; it is wrong, and wrongness of this kind has no syntactic signature. This is the same failure mode as a silently mis-joined column, a units mix-up, or a sitemap index that points at another sitemap index: the artefact is valid, the pipeline reports success, and the defect lives in behaviour that nothing is looking at.

There is exactly one thing that catches it, and it is not a cleverer assertion. It is a list of expected outputs, name by name, checked against a dictionary by a person who reads the language. Our current test set for Russian patronymics is 75 such pairs, every one of them written down because a reference work says so and not because the rule produces it. That is a data asset, not a code asset, and it is the only part of this that generalises.

What we do, and what we are not sure about

We emit one form per name, because a generated identity has to put one string in the field.

Where the dictionaries record only one form, we emit it. Where they record two, we emit the one that dominates real usage: -ьевич for the sonorant-cluster group and for the single-consonant group. The writer Викентий Викентьевич Вересаев is the canonical illustration — the dictionary allows Викентиевич, and the man was Викентьевич.

Four things we are genuinely unsure of, stated plainly:

1. We verified 27 of the 59 names, not all 59. That is 80.6% of the branch by weight but only 46% of the entries. The unchecked tail is 32 rare names at weight 1 each, and some of them are probably doublets we have recorded as single-form.

1. Our single-consonant default may be wrong for individual names. Дионисий proves the clause is not exceptionless, and we found that one by accident. There may be others in the unchecked tail.

1. We could not read the primary dictionary list in full. The RAS orthographic dictionary's appendix of personal names is the authority here; we worked from entries quoted in reference sources and from a personal-names dictionary aggregator, not from the printed appendix. Where those disagree with the printed edition, the printed edition is right and we are wrong.

1. The sonorant/obstruent split is a description, not a derivation. It holds without exception across all 16 cluster-final stems we hold and all nine dictionary entries we could check. We do not have a phonetic account we are confident in — нтj and трj are both three-consonant sequences, and appealing to pronounceability does not obviously separate them. The pattern is real; the explanation is open.

That last one is worth saying out loud, because it is the difference between the received rule and this one. The received rule is short, confident, and states things it cannot support. We would rather publish a rule that covers the same data and admits where it is only reporting what the dictionaries say.

Data as of 2026-07-22. All corpus figures are computed from our Russian male given-name dataset: 316 entries, combined weight 1,699; 59 entries in -ий at combined weight 320; cluster inventory нт×8, мп×2, рг/рк/тр/кл/вд/фр×1; 50 labial-final entries at weight 214; 5 entries containing ё at weight 63; 13 entries in -а/-я at weight 56.

Sources:

• Образование отчеств — civil registry office of Smolensk Region. The published rule quoted at the top, including the к, х, ц list and the нт carve-out.
• Образование и написание отчеств — civil registry office of Kurgan Region, reproducing the same appendix.
• Написание личных имён. Образование и написание отчеств — the same rule with the examples Никий → Никиевич, Люций → Люциевич, Стахий → Стахиевич, Дмитрий → Дмитриевич.
• Как правильно пишутся имена и отчества? — Gramota.ru (Vinogradov Russian Language Institute). The statement that both forms are orthographically correct for a significant group of -ий names, with Геннадиевич/Геннадьевич, Иннокентиевич/Иннокентьевич, Виталиевич/Витальевич, Евгениевич/Евгеньевич.
• Имена, псевдонимы, прозвища, клички — the appendix list of personal names with their patronymics, sourced from Русский орфографический словарь РАН, ed. V. V. Lopatin and O. E. Ivanova, 4th ed., Moscow 2012, Appendix 2.
• N. A. Petrovsky, Словарь русских личных имён — individual entries consulted via lexicography.online for Лаврентий, Иннокентий, Терентий, Викентий, Авксентий, Дмитрий, Ираклий, Онуфрий, Клавдий, Аверкий, Георгий, Анатолий, Валерий, Геннадий, Василий, Юрий, Дионисий, Евгений, Харлампий, Евлампий.
• Справочник личных имён народов РСФСР, ed. A. V. Superanskaya — the handbook recommended to registry offices, and the ultimate source of the rule as the registry sites state it.
• Federal Law No. 143-FZ On Acts of Civil Status, art. 18, and the Family Code of the Russian Federation, art. 58 — the legal basis for the patronymic being a mandatory component of the official name.

A note on verification. Corpus counts in this article were produced by running our own rule over the dataset and tabulating the result; they are reproducible from the numbers given. Dictionary attributions were read from quoted entries rather than from printed editions — gramota.ru refused automated access and the Lopatin appendix mirror was unreachable at the time of writing. Twenty-seven of the 59 names were checked; the remaining 32, all at weight 1, were not. Where this article says "the dictionary records only one form", it means we found only one form recorded in the sources we could read, which is a weaker claim than the absence of the other form.