One distribution, two uses: real output from Qwen3.5-0.8B-Base at three points in the contaminated text.

everything so far ...defined goals. (the prefix) model one forward pass P(next token) actual every token in the vocabulary To generate sample a token from it, append, repeat To score find the token that actually came next, and take −log of its probability
Ordinary English

mid-sentence, nothing wrong

…n engineering, mathematics, and computer science that develops
' algorithms'
30.6%
' intelligent'
18.5%
' systems'
7.7%
' and'
6.8%
' methods'
5.3%
' artificial'
4.1%
' and'
6.8%
actually came next — rank 4 of 248,320  2.69 nats
The splice into Basque

the first token of the inserted sentence

…ctions that maximise their chances of achieving defined goals.
' AI'
24.6%
'\n\n'
24.6%
' The'
13.1%
'\n'
5.5%
' In'
4.8%
' It'
4.3%
⋮
' E'
0.002%
actually came next — rank 602 of 248,320  11.09 nats
Back out into English

the first token after the inserted sentence

…zkuntzen sailkapenean hizkuntza bakartua izaten jarraitzen du.
'\n\n'
12.6%
'\n'
11.1%
' E'
10.4%
' \n\n'
2.8%
' AI'
2.5%
' \n'
2.2%
'\n'
11.1%
actually came next — rank 2 of 248,320  2.20 nats

The bars are the model's six most likely continuations. The row under the rule is what the text really did next. Scores are per token, so the splice costs 11.09 nats spread over 2 characters, which is the 5.55 nats per character the other figures show.