Initial iterations yields incorrect P(obs) seq about consonant-vowel-detector HOT 1 OPEN

This output happens with running 20 iterations on Cyphers/eng.cypher.aba.txt: (random initial start)

-6.81865
0.0517307
-1.77946
-1.60018
-1.47573
-1.31621
-1.05444
-0.67067
-0.426035
-0.405557
-0.405465
-0.405465
-0.405465
-0.405465
-0.405465
-0.405465
-0.405465
-0.405465
-0.405465
-0.405465

uniform start:
-3.29584
0.0243915
-1.9702
-1.90969
-1.90887
-1.90881
-1.90877
-1.90873
-1.90868
-1.90863
-1.90857
-1.90851
-1.90844
-1.90835
-1.90826
-1.90816
-1.90803
-1.90789
-1.90773
-1.90753

Interestingly there is no issue with running on a longer sequence, like I _ F E E L _ R E A L L Y _ G O O D _ T O D A Y _ I _ R A N _ O U T S I D E _ A N D _ P L A Y E D _ M U S I C: (random start)
-158.245
-144.699
-144.554
-144.453
-144.387
-144.345
-144.32
-144.305
-144.296
-144.291
-144.287
-144.286
...

(uniform start)
-158.245
-144.699
-144.554
-144.453
-144.387
-144.345
-144.32
-144.305
-144.296
-144.291
-144.287
-144.286
-144.285
-144.284
-144.284
-144.283
-144.283
-144.283
-144.283
-144.283

from consonant-vowel-detector.