Inferred grammar

Four layers, each derived from the layer above. Glyph statistics and affixes are counted directly; categories and rules are fitted models and are labelled as such.

layer 1 · observed

Glyph inventory and recurrent groups

Character n-grams ranked by pointwise mutual information: sequences that co-occur far more than their component glyphs would predict. These are the candidates for single underlying signs.

Glyph frequency

13.3%

10.5%

9.3%

9.2%

7.5%

6.9%

6.8%

6.1%

5.7%

5.5%

3.9%

3.6%

3.2%

Bound glyph groups

aiin11.1bdaii8.7biin7.5bqoke8.3bqoka8.4bkain8.7bhedy7.4bched7.5bkaii8.1baii6.3bpche8.5bokai7.3bqok6.2bshed7.3bdain7.8beedy6.5bopch7.9bokee6.4bche5.3btaii7.6bain5.9bchol6.5bkeed6.5bchey6.2bedy5.1botai6.8b

Hover for counts. High-PMI groups such as these behave like units: they resist being split by the morphology model and cluster at fixed positions inside words.

layer 2 · observed

Morphology

Affixes kept only if they attach to at least 12 distinct stems that are themselves attested as words. 96% of word types receive a non-empty prefix or suffix.

Prefixes

form	stems	tokens
o-	689	5,836
q-	480	4,864
ch-	469	3,638
qo-	410	4,540
d-	357	2,839
y-	354	1,364
che-	265	1,965
sh-	249	1,753
l-	264	1,145
ot-	216	1,674
ol-	233	898
s-	220	900
cho-	211	766
t-	211	740
ok-	179	1,800
k-	172	868

Suffixes

form	stems	tokens
-y	3210	14,554
-n	1053	5,931
-dy	603	5,661
-ey	336	3,487
-ar	369	1,580
-r	326	3,482
-aiin	344	1,937
-l	305	3,869
-ol	311	2,311
-s	317	796
-al	291	1,293
-or	285	1,460
-in	244	4,608
-edy	233	3,729
-o	224	883
-ain	201	1,120

Most productive stems

ai60ch146ke81ee121ed49ka52eo54da44ol71ta35sh90ckh62cth62al67ar54te45od61kai21or46ko41kch57pch65ok51che71to33he5ot35tch43ho6do25dai19dy11qo8s1ek45kee46

layer 3 · inferred

Word classes

Seven clusters over context, position and affix features. The role names are descriptions of measured behaviour — they are hypotheses about grammatical function, not identifications of parts of speech in a known language.

C1 · modifier-like

inferred

types: 107
tokens: 12,020
line-initial: 8%
ending: -y

frequent members

daiinolchedyshedycheyqokeeyqokeedydarqokainsheyqokedyqokaiindaldy

evidence trail

observeddistributional counts
12,020 tokens across 107 word types; 7.8% line-initial, 9.2% line-final, mean relative position 0.51.
observedmorphological signature
dominant ending "-y" covers 27.5% of class tokens; mean word length 4.85 glyphs.
inferredcategory assignment
k-means (k=7, cosine) over left/right context vectors built from the 40 most frequent forms, plus positional and affix features. The role label follows from the profile above; no natural language is assumed.

C2 · predicate-like (verb / attribute)

inferred

types: 341
tokens: 3,246
line-initial: 6%
ending: -dy

frequent members

ytaiinoteodyokeodychedaiinqokeodyqokcheyopcheyokamotchdyokchedyorainqokamqotchedyqotchdy

evidence trail

observeddistributional counts
3,246 tokens across 341 word types; 6.3% line-initial, 22.1% line-final, mean relative position 0.58.
observedmorphological signature
dominant ending "-dy" covers 22.8% of class tokens; mean word length 5.53 glyphs.
inferredcategory assignment
k-means (k=7, cosine) over left/right context vectors built from the 40 most frequent forms, plus positional and affix features. The role label follows from the profile above; no natural language is assumed.

C3 · opener / topic marker

inferred

types: 139
tokens: 1,325
line-initial: 59%
ending: -ol

frequent members

dshedytchedyyteedydchedydchorpolycheeydcholsoiintchorydaiindoiindcheyokeeol

evidence trail

observeddistributional counts
1,325 tokens across 139 word types; 59.3% line-initial, 4.8% line-final, mean relative position 0.22.
observedmorphological signature
dominant ending "-ol" covers 17.4% of class tokens; mean word length 5.61 glyphs.
inferredcategory assignment
k-means (k=7, cosine) over left/right context vectors built from the 40 most frequent forms, plus positional and affix features. The role label follows from the profile above; no natural language is assumed.

C4 · function word / connective

inferred

types: 62
tokens: 3,757
line-initial: 10%
ending: -∅

frequent members

aiinoraralsotardaircheorainamrsarairo

evidence trail

observeddistributional counts
3,757 tokens across 62 word types; 9.8% line-initial, 11.1% line-final, mean relative position 0.51.
observedmorphological signature
dominant ending "-∅" covers 40.0% of class tokens; mean word length 3.12 glyphs.
inferredcategory assignment
k-means (k=7, cosine) over left/right context vectors built from the 40 most frequent forms, plus positional and affix features. The role label follows from the profile above; no natural language is assumed.

C5 · modifier-like

inferred

types: 67
tokens: 919
line-initial: 17%
ending: -r

frequent members

otorokorykarsharqotorsairkorytarodaraiirtordaiirokeorqor

evidence trail

observeddistributional counts
919 tokens across 67 word types; 16.8% line-initial, 7.2% line-final, mean relative position 0.46.
observedmorphological signature
dominant ending "-r" covers 100.0% of class tokens; mean word length 4.34 glyphs.
inferredcategory assignment
k-means (k=7, cosine) over left/right context vectors built from the 40 most frequent forms, plus positional and affix features. The role label follows from the profile above; no natural language is assumed.

C6 · closer / terminal marker

inferred

types: 127
tokens: 1,684
line-initial: 9%
ending: -y

frequent members

olyotchyshdyodyckhychotyokchyshcthyalychecthydalykchydchyary

evidence trail

observeddistributional counts
1,684 tokens across 127 word types; 8.6% line-initial, 24.1% line-final, mean relative position 0.59.
observedmorphological signature
dominant ending "-y" covers 100.0% of class tokens; mean word length 4.74 glyphs.
inferredcategory assignment
k-means (k=7, cosine) over left/right context vectors built from the 40 most frequent forms, plus positional and affix features. The role label follows from the profile above; no natural language is assumed.

C7 · modifier-like

inferred

types: 109
tokens: 4,160
line-initial: 11%
ending: -y

frequent members

cholchorsholychyshocthyshyshordamokyqotyotolokol

evidence trail

observeddistributional counts
4,160 tokens across 109 word types; 10.8% line-initial, 9.6% line-final, mean relative position 0.47.
observedmorphological signature
dominant ending "-y" covers 18.9% of class tokens; mean word length 4.48 glyphs.
inferredcategory assignment
k-means (k=7, cosine) over left/right context vectors built from the 40 most frequent forms, plus positional and affix features. The role label follows from the profile above; no natural language is assumed.

layer 4 · inferred

Candidate syntax rules

Every class-to-class transition compared against what independent placement would predict. z is the standardised deviation; ⟦line edge⟧ stands for the start or end of a written line.

Strongly preferred sequences

rule	observed	expected	z
⟦line edge⟧→C3	786	163	+48.7
C4→C4	1,352	753	+21.8
C1→C1	7,731	6402	+16.6
C2→⟦line edge⟧	1,001	618	+15.4
C7→C7	1,096	708	+14.6
C4→⟦line edge⟧	1,077	694	+14.5
C6→⟦line edge⟧	406	208	+13.8
C5→C4	309	183	+9.3
⟦line edge⟧→C5	246	169	+5.9
C3→C7	237	172	+5.0
⟦line edge⟧→C7	799	673	+4.8
C5→C5	72	45	+4.1

Avoided sequences (grammatical constraints)

⟦line edge⟧↛⟦line edge⟧	0	640	-25.3
C4↛C1	1,545	2196	-13.9
C1↛C3	234	517	-12.4
C1↛C4	1,681	2196	-11.0
C4↛C3	48	177	-9.7
C1↛C7	1,757	2129	-8.1
C3↛⟦line edge⟧	64	163	-7.8
C2↛C3	83	158	-6.0

Under-represented transitions are as informative as preferred ones: a system that forbids certain orderings is behaving like a grammar rather than like a word-salad generator.