The context window was a suggestion

The model card said 32k. The flag I thought I had set said 32k. Around nine thousand tokens the reply forgot the first constraint and invented a second task.

The cause was dull. I had set the context in one place and left the default in the server that sits in front of the runtime. The server capped the prompt and did not say so. The model then did its best with a missing middle.

What fixed it: one number, written in the launch note, and a log line that prints the context the process actually got. If those two disagree, I do not run the test. Prefill got slower at the real 32k. That is annoying, and it is the correct kind of annoying.

I am not publishing the flag soup. Two projects use two names for the same idea, and copying my pair onto a different build is how this note starts over. The useful part is the check, not the spell.

Sample prompt, sample cap, no hostname. The surprise was the silence, not the size.