pflt-fsot-sample / docs /CROSS_LINGUAL_FSOT.md
dappalumbo91's picture
v0.2.1: FULL verification source (not metrics-only)
29d94b3 verified
|
Raw
History Blame Contribute Delete
8.2 kB

Cross-lingual analysis under FSOT

Built: 2026-07-21T16:50:31.075216+00:00 Law: S=K(T1+T2+T3) · pin D1D38A

Unique stack (already shipping)

Intrinsic free-parameter FSOT model — competitors have no law pin, no offline classical+visual densify stack, and no cert-gated converse under the same constitution.

Cross-lingual mapping is a surface of FSOT intelligence: densify knowledge without rewriting seed law. English gloss is a temporary pivot for analysis; meaning is law-grounded densify, not a cloud API.

Catalog solidification

Metric Value
Language codes 113
OPEN-SET form→gloss 99.99%
PRODUCT form→gloss 99.99%
All langs ≥95% (n≥20) True
Solid (n≥50 & OK) 106
Thin (n<50) 7
Gap to 100 codes 0

Breadth vs competitors

System Languages (order of magnitude)
Google Translate ~249
NLLB ~200
DeepL 30–100
PFLT (FSOT) 113 (56% of NLLB, 45% of Google)

English-gloss hub (cross-lingual connectivity)

  • Gloss keys sampled: 40058
  • Concepts shared by ≥3 languages: 1762

Top shared concepts

Gloss # langs Sample langs
order 12 arc, de, es, fr, it, ja, nl, pl, pt, ru, sga, tr
say 12 de, es, fr, it, ja, la, nl, orv, pl, pt, ru, sv
help 11 ar, cu, es, fr, it, nl, osx, pt, ru, sv, tr
see 11 cu, de, es, fr, it, ja, osx, pt, ru, san, sv
work 11 de, es, fr, grc, it, nl, non, pt, ru, sga, sv
bear 11 cu, es, fr, grc, it, ja, nl, pl, pt, ru, san
sound 11 ang, es, fr, it, ja, nl, ru, san, sv, tr, xcl
leave 11 de, es, fr, got, grc, it, ja, pl, pt, sv, tr
female given name 11 de, es, fr, it, ja, nl, pl, pt, ru, sv, tr
remain 10 cu, egy, es, fr, it, ja, la, nl, pt, sv
kill 10 cu, de, es, fr, it, nl, pl, pt, sga, sv
throw 10 cu, de, es, fr, he, it, pt, san, tr, xcl
sing 10 ang, cu, de, es, fr, it, osx, pl, pt, sv
sin 10 ar, cop, es, fr, got, grc, la, roa-opt, sga, sv
raise 10 arc, de, es, fr, it, ja, nl, pt, san, sv

Top language pairs (via shared gloss)

Pair Shared concepts
ru-sv 298
fr-it 269
nl-sv 245
fr-sv 236
fr-nl 230
it-nl 225
nl-ru 225
it-sv 223
es-it 221
it-pt 219
de-ru 216
de-nl 215
fr-pt 210
ja-ru 204
it-ru 200

Multi-source translation probes (form→EN gloss path)

Lang Coverage Source Hyp (EN surface)
es 1.0 hola mundo hello world
fr 1.0 bonjour monde hello world
de 1.0 hallo welt hello world
it 1.0 ciao mondo hello world
la 1.0 aqua lingua water language
pt 1.0 ola mundo hello world
nl 1.0 hallo wereld hello world
ru 1.0 привет мир hello world
ja 1.0 water
zh 1.0 water
ar 1.0 ماء water
hi 1.0 नमस्ते hello
sw 1.0 jambo hello
tr 1.0 merhaba dünya hello world
pl 1.0 cześć świecie hello world
ko 1.0 안녕 hello
he 1.0 שלום peace
el 1.0 γεια hi
ga 1.0 dia dhuit Genderless alternative
eo 1.0 saluton mondo hello world

M6 sentence snapshot (not parity claim)

BLEU-4 2.44 · B1 43.87 · U-F1 34.31 · cov 90.33%

Climbing offline toward competitive sentence bars; law remains master.

Thin languages (fill next)

Lang n Open% Product%
roa-opt 47 100.0 100.0
orv 36 100.0 100.0
osp 34 100.0 100.0
mga 25 100.0 100.0
pro 17 100.0 100.0
pal 10 100.0 100.0
hit 3 100.0 100.0

Per-language (n≥20)

Lang n Open% Product% OK
ko 14646 100.0 100.0 Y
la 11360 100.0 100.0 Y
pl 8037 100.0 100.0 Y
sv 6452 100.0 100.0 Y
ja 6342 100.0 100.0 Y
de 6002 100.0 100.0 Y
hi 5731 100.0 100.0 Y
ru 5564 100.0 100.0 Y
it 5428 100.0 100.0 Y
tr 5162 100.0 100.0 Y
es 5028 100.0 100.0 Y
pi 4765 100.0 100.0 Y
ca 4619 100.0 100.0 Y
fr 4566 100.0 100.0 Y
pt 4356 100.0 100.0 Y
nl 4305 100.0 100.0 Y
vi 4079 100.0 100.0 Y
grc 3966 100.0 100.0 Y
id 3814 100.0 100.0 Y
ang 2789 99.96 99.96 Y
egy 2044 100.0 100.0 Y
ar 1319 100.0 100.0 Y
got 1133 99.91 99.91 Y
he 917 100.0 100.0 Y
san 910 100.0 100.0 Y
non 806 100.0 100.0 Y
fa 740 100.0 100.0 Y
sga 603 99.83 99.83 Y
cu 600 100.0 100.0 Y
akk 544 100.0 100.0 Y
cop 504 100.0 100.0 Y
bn 500 100.0 100.0 Y
cs 500 100.0 100.0 Y
el 500 100.0 100.0 Y
fi 500 100.0 100.0 Y
hu 500 100.0 100.0 Y
ms 500 100.0 100.0 Y
ro 500 100.0 100.0 Y
sw 500 100.0 100.0 Y
ta 500 100.0 100.0 Y
th 500 100.0 100.0 Y
uk 500 100.0 100.0 Y
zh 500 100.0 100.0 Y
arc 498 100.0 100.0 Y
osx 464 100.0 100.0 Y
peo 415 100.0 100.0 Y
syc 409 100.0 100.0 Y
sum 404 100.0 100.0 Y
da 400 100.0 100.0 Y
gl 400 100.0 100.0 Y
eo 400 100.0 100.0 Y
mk 400 100.0 100.0 Y
bg 400 100.0 100.0 Y
nn 400 100.0 100.0 Y
tl 400 100.0 100.0 Y
ga 400 100.0 100.0 Y
cy 400 100.0 100.0 Y
lt 400 99.75 99.75 Y
is 400 100.0 100.0 Y
sq 400 100.0 100.0 Y
ka 400 100.0 100.0 Y
hy 400 100.0 100.0 Y
te 400 100.0 100.0 Y
ceb 400 99.75 99.75 Y
gd 400 100.0 100.0 Y
az 400 100.0 100.0 Y
sk 400 100.0 100.0 Y
eu 400 100.0 100.0 Y
ur 400 100.0 100.0 Y
ml 400 100.0 100.0 Y
pa 400 100.0 100.0 Y
et 400 100.0 100.0 Y
yi 400 100.0 100.0 Y
kk 400 100.0 100.0 Y
km 400 100.0 100.0 Y
my 400 100.0 100.0 Y
af 400 99.75 99.75 Y
be 400 100.0 100.0 Y
gu 400 100.0 100.0 Y
sl 400 100.0 100.0 Y
mn 400 100.0 100.0 Y
oc 400 100.0 100.0 Y
yo 400 100.0 100.0 Y
mr 400 100.0 100.0 Y
jv 400 100.0 100.0 Y
haw 400 100.0 100.0 Y
am 400 100.0 100.0 Y
ne 400 100.0 100.0 Y
qu 400 100.0 100.0 Y
zu 400 100.0 100.0 Y
xh 400 100.0 100.0 Y
ha 400 100.0 100.0 Y
so 400 100.0 100.0 Y
ps 400 100.0 100.0 Y
si 400 100.0 100.0 Y
lv 400 100.0 100.0 Y
enm 400 99.5 99.5 Y
fro 400 100.0 100.0 Y
frm 400 100.0 100.0 Y
goh 400 100.0 100.0 Y
gmh 400 100.0 100.0 Y
nb 400 100.0 100.0 Y
sh 400 100.0 100.0 Y
en 228 100.0 100.0 Y
xcl 175 100.0 100.0 Y
uga 96 100.0 100.0 Y
roa-opt 47 100.0 100.0 Y
orv 36 100.0 100.0 Y
osp 34 100.0 100.0 Y
mga 25 100.0 100.0 Y

Next under FSOT

  • Continue Kaikki to100 until catalog≥100 with all n≥50 solid
  • Boost thin historical codes (hit/sum/syc/pal/peo)
  • M6 CJK BPE / larger phrase mass for ja/zh
  • Optional Ada converse smoke on multi-lang probes