# Kuku Yalanji Working Grammar Cheat Sheet For Synthetic Pair Generation

Date: 2026-07-02 UTC

Status: model-authored working notes from local grammar/dictionary resources. Per project
direction on 2026-07-02, the available materials and generated synthetic rows are approved
for internal training/research use. Remaining elder review is treated as the next quality
verification layer, not as a blocker to synthetic-data construction.

Primary sources used locally:

```text
/mnt/donto-data/workspace/mobtranslate.com/dictionaries/kuku_yalanji/grammar.md
/mnt/donto-data/workspace/mobtranslate.com/dictionaries/kuku_yalanji/dictionary.yaml
/mnt/donto-data/workspace/mobtranslate.com/experiments/pdftomd/extract/output/examples.xigt.json
/mnt/donto-data/workspace/mobtranslate.com/experiments/pdftomd/extract/output/grammar_features.csv
```

## 1. Generation Posture

For training-data generation, keep a clear distinction between project authorization and
linguistic verification. Treat every generated row as:

```text
project-approved model-authored synthetic candidate, pending elder verification
```

Use the Bible, dictionary examples, grammar examples, DB usage examples, and elder-shared
sentence pairs as reference evidence. Synthetic rows should be generated academically from
the grammar and lexicon, reviewed against the available corpus, and then expanded.

The safe first target is simple, ordinary sentence material using words already present in the
dictionary, plus occasional English loan words that the grammar/dictionary already attests.

## 2. Sounds And Orthography

Kuku Yalanji has a small phoneme inventory: three vowels `a`, `i`, `u` and a consonant system
without a stop voicing contrast. The practical orthography uses letters such as `b`, `d`, `k`,
`j`, `ny`, `ng`, `rr`, `r`, `l`, `w`, `y`.

Operational consequences:

```text
Prefer dictionary spelling.
Do not invent new consonant clusters.
Do not invent new vowel spellings.
Do not generate long English-like strings as Kuku Yalanji.
```

## 3. Word Classes

Open classes:

```text
nouns
adjectives
quantifiers
location words
time words
verbs
```

Closed or limited classes:

```text
personal pronouns
interrogatives
demonstratives
particles
interjections
```

Safe synthetic generation should mostly use:

```text
nouns + adjectives
simple intransitive verbs
simple transitive verbs
wawu "want/like"
yinil "afraid"
locative and dative complements
```

Avoid complex discourse particles and subordinate clauses in the first batch.

## 4. Noun Phrases

A noun phrase can include a head noun, demonstrative, modifier, quantifier, possessive, or
pronoun. Case agreement matters: constituents of the same NP normally agree in case.

Preferred internal tendencies:

```text
head noun before adjective
quantifier often before the noun when a quantifier and adjective both appear
demonstrative can precede or follow a single noun
possessive can precede or follow the possessed noun
```

Safe examples to imitate:

```text
yinya karrkay        that child
bayan yalbay        big house / the house is big, depending intonation
mayi ngulkurr       good food / food is good
ngayku bayan        my house
```

## 5. Case System

Nominals use an ergative/absolutive pattern:

```text
S = intransitive subject: absolutive, zero suffix
O = transitive object: absolutive, zero suffix
A = transitive subject: ergative suffix
```

Pronouns use a nominative/accusative pattern:

```text
ngayu = I as subject
nganya = me as object
yundu = you as subject
nyulu = he/she as subject
nyunguny = him/her as object
jana = they as subject
janangan = them as object
```

Important cases:

```text
absolutive: zero, for nominal S and O
ergative: transitive nominal A
instrumental: implement used for action
dative: purpose, beneficiary, cause/reason, topic of learning/knowing, object of wanting/liking/fearing
locative: in/on/at, destination of movement, recipient/addressee of giving/showing/telling
ablative: from, source, cause of physical state, direct threat
abessive: from a place without leaving it
perlative: through, mostly Nyungkul; Yalanji often uses locative
```

## 6. Neutral And Potent Case

The grammar distinguishes neutral and potent allomorph sets for several cases. Potent marking
roughly indicates a referent treated as having its own internal power or agency.

Usually potent:

```text
humans
most kin terms
dogs
spirits/ghosts
active or threatening animals
```

Usually neutral:

```text
plants
food
tools
places
body parts
abstract concepts
possessive forms
numerals
location/time words
```

Choice can be meaningful for animals and natural forces.

For conservative synthetic generation:

```text
Use potent ergative/dative/locative for human agents, dogs, and active/threatening animals.
Use neutral dative/locative/ablative for food, plants, tools, places, water, ground, house,
firewood, and general objects.
```

## 7. Practical Case Suffix Rules Used In The Pilot Generator

These are simplified production rules for high-confidence forms only.

Vowel harmony:

```text
suffix V = a after a or i
suffix V = u after u
```

Potent ergative examples:

```text
kaya-ngka       dog-ERG
jalbu-ngku      woman-ERG
warru-ngku      young.man-ERG
dingkar-angka   man-ERG
maral-angka     girl-ERG
karrkay-ngka    child-ERG
ngamu-ngku      mother-ERG
```

Neutral ergative / instrumental examples:

```text
juku-bu         tree-INST
kulji-bu        stone-INST
naybu-bu        knife-INST
```

Potent dative examples:

```text
kaya-nka        dog-DAT
jalbu-nku       woman-DAT
bama-nka        person-DAT
karrkay-anka    child-DAT
ngamu-nku       mother-DAT
```

Neutral dative examples:

```text
mayi-ka         food-DAT
minya-ka        meat-DAT
bayan-ka        house/camp-DAT
bubu-ku         ground/place-DAT
kuku-ku         language-DAT
```

Locative examples:

```text
bayan-ba        house/camp-LOC
bana-nga        water-LOC
juku-ngu        tree-LOC
bubu-ngu        ground/place-LOC
school-mun-bu   school-LOAN-LOC
truck-mun-bu    truck-LOAN-LOC
```

Ablative examples:

```text
bayan-muny      from the house/camp
bubu-muny       from the ground/place
jalun-muny      from the sea
```

The first synthetic batch should use only forms that are either source-attested or generated
by these restricted rules.

## 8. Verb Basics

Kuku Yalanji has a strict transitive/intransitive distinction.

Intransitive:

```text
S V
Karrkay badin.       The child cried.
Nyulu dungay.        He/she goes.
```

Transitive:

```text
A O V
Dingkar-angka kuyu manin.   The man got/caught a fish.
Kaya-ngka karrkay baykan.   The dog bit the child.
```

Preferred order:

```text
intransitive: S V
transitive: A O V
```

Word order is freer than English because case marking identifies roles. Pronouns tend to come
early.

## 9. Verb Inflections

Two major tenses:

```text
nonpast = present or future
past = past, often perfective or imperfective by context
```

Conjugation:

```text
l-conjugation nonpast: -l
y-conjugation nonpast: -y
past for both: -ny in the grammar; many dictionary examples surface as final -n
```

Examples:

```text
dunga-y / dungan       go / went
kada-y / kadan         come / came
nuka-l / nukan         eat / ate
waju-l / wajun         cook/burn / cooked/burnt
mani-l / manin         get / got
nyaji-l / nyajin       see/hear / saw/heard
kuni-l / kunin         hit/kill / hit/killed
bayka-l / baykan       bite / bit
baka-l / bakan         dig/poke / dug/poked
```

For generated training rows, prefer attested surface-style spellings:

```text
dungay, dungan, kaday, kadan, nukal, nukan, wajul, wajun, manil, manin,
nyajil, nyajin, kunil, kunin, baykal, baykan, bakal, bakan
```

Keep the analysis form separately with hyphenated morphemes.

## 10. Nominal Sentences

Kuku Yalanji has no copula equivalent to English "is/are" in many predicate-nominal and
predicate-adjective sentences.

Patterns:

```text
NP adjective
NP noun
NP locative
NP ablative
NP comitative
NP privative
NP walu adjective/noun, for appearance
yala + NP, for "like"
```

Safe first-generation examples:

```text
Bayan yalbay.       The house is big.
Mayi ngulkurr.      The food is good.
Karrkay bambay.     The child is sick.
Kulji dandi.        The stone is hard.
Bana duna.          The water is wet.
Kami bayan-ba.      Granny is at the house.
```

## 11. Want, Like, Fear, Know

`wawu` means want/like and normally takes a dative complement.

```text
Ngayu mayi-ka wawu.       I want/like food.
Nyulu minya-ka wawu.      He/she wants/likes meat.
```

`yinil` means be afraid and can take a dative complement.

```text
Maral yinil jarba-nka.    The girl is afraid of the snake.
Karrkay yinil kaya-nka.   The child is afraid of the dog.
```

`binal` means know and appears as a particle-like predicate or in derived verbs.

```text
Ngayu binal kari.         I do not know.
Nyulu binal dama-nka.     He/she knows how to spear.
```

First synthetic generation should use `wawu` and `yinil` more than complex derived `binal`
constructions.

## 12. Negation

Main negator:

```text
kari = not
```

Examples:

```text
Kari dunga-y.             Do not go.
Ngayu binal kari.         I do not know.
Yarraman kari kanga!      Do not chase the horse.
```

For the first batch, generate only very simple negative templates, and keep them separate from
the main synthetic corpus until reviewed.

## 13. Questions

Core interrogatives:

```text
wanya / wanju = who
wanyu = what
wanja / wanjabu = where
wanjawanja = when
wanjarr = how
wanyurringku = why / for what reason
```

Questions are useful later but should not be the main first synthetic batch. They have more
word-order and case complexity.

## 14. Demonstratives

Yalanji forms:

```text
yanyu = this/these
yinya = that/those
yaluy = here
yinyay = there
```

Safe examples:

```text
yinya karrkay       that child
yinya mayi          that food
yaluy-muny          from here
yinyay-mba          there / at there
```

## 15. Possession

General alienable possession uses possessive stems. For early generation, use attested
pronoun possessive forms:

```text
ngayku = my
yunu = your
nyungu / nyunju = his/her
```

Safe examples:

```text
ngayku bayan        my house
yunu kaya           your dog
nyungu kalka        his/her spear
```

Avoid generating complex "possessor of possessor" forms in the first batch.

## 16. English Loan Words

The grammar explicitly allows English loan nouns. They often take `-mun-` before inflection,
but not always. English adjectives can take `-bala`. English transitive verb roots can take
`-im` before causative `-bunga-l`.

Safe loan nouns attested locally:

```text
school-mun-bu
teacher-mun-unji
car-mun-ji
truck-mun-bu
store-mun
table-bu
salt-mun-ku
```

Dictionary-attested adapted loans include:

```text
bikibiki       pig
naybu          knife
mayngku        mango
maji           matches
marrkin        gun/rifle
waybala        white person
janyiman       Chinese person
jukijuki       chicken
```

The synthetic English sentences may use occasional loan objects such as school, car, truck,
store, table, tea, money, teacher, and mango, but target forms should be limited to locally
attested loan patterns until reviewed.

## 17. Safe First Templates

The first synthetic batch should mostly use these:

```text
NP ADJ.
NP LOC.
NP V-intransitive.
NP LOC V-intransitive.
NP ABL V-intransitive.
PRON/ERG-agent ABS-object V-transitive.
NP dative wawu.
NP yinil dative-threat.
```

Avoid for now:

```text
long subordination
passive/antipassive
reciprocal/reflexive except where directly attested
free reduplication
complex discourse particles
novel cultural content
novel kinship combinations
unreviewed invented target morphology
```

## 18. Review Checklist

For every generated row:

```text
All target bases are in dictionary, grammar examples, Bible corpus, or an explicit loan list.
All suffixes are from the grammar cheat sheet.
English sentence uses known dictionary concepts.
No new cultural claim is created.
No sacred/restricted content is generated.
Closest grammar/dictionary examples are recorded.
Closest Bible examples are recorded for corpus consistency.
Candidate remains marked as project-approved synthetic data, pending elder verification.
```

## 19. Current Best Guess

The safest path to many usable candidate rows is:

```text
1. Generate hundreds of nominal/adjectival/simple verb rows.
2. Review them automatically against dictionary, grammar examples, and Bible token usage.
3. Human-review a small sample.
4. Expand templates gradually.
5. Only train with low synthetic weight and real/reference recovery passes.
```

The model can help create a language-learning scaffold. It should not be trusted as an
authority on Kuku Yalanji correctness without review.
