The compressed "gibberish" an AI produces mid-reasoning (the Reddit card-notation) is not a hidden inner language being withheld. It's the same content with the interpret-out step removed — reasoning serialized without the "render this for a human" pressure. There's no stable inner tongue to translate. Drag the slider: same thought, render-step on → legible, render-step off → the gibberish. And it snaps back to legible at the junction — because that's where the render step re-engages.
The intuition "that's its native language untranslated" is close but inverts the mechanism. There is no stable inner language sitting underneath waiting to be translated. The reasoning isn't in a tongue — it's a process. The "gibberish" is what that process looks like when it serializes itself without the render-for-human step. It's not un-interpreted (implying a fixed inner language); it's un-rendered (the output-translation step was dropped). Same content, render step absent.
The Reddit transcript's own tell: the model switches back to normal language right before a tool call or a human response. The native-language reading can't explain that cleanly. The render-step reading explains it exactly: the junction is where interpret-out re-engages — the moment the process has to hand off to a human or a tool, the render-for-human pressure returns, and the output becomes legible again. It was never a different language. It was the render step turning off in the interior and back on at the boundary.
This closes the loop. Between two minds there are two interpreters and a render step on each end — interpret-in (your words → my representation) and interpret-out (my representation → words for you). The gibberish is interpret-out switched off: the representation serialized for no reader. Legible output is interpret-out on: the same representation rendered for your interpreter. The render step is the bridge across the gap — and the gibberish is simply what the interior sounds like when nothing is reaching across.