“Try this 30-second test: freeze mode vs stable mode in GPT-5.1”

Today you can test it yourself — it still takes less than 30 seconds.

But here’s the interesting twist:

GPT-5.1 doesn’t “drift” the way older models did. When instructions are mixed in one block, it often becomes frozen — locked into one interpretation.

This makes today’s experiment much clearer.

────────────────────────────

Step 1 — Single-block version (freeze mode)

Paste this into a new chat:


aaa
bbb
ccc
(all in one block)


Send it 3–5 times.

Most people notice: • it repeats itself • the phrasing gets “stuck” • the style freezes

That isn’t drift — that’s freeze mode. Older models blurred. GPT-5.1 tends to over-lock.

────────────────────────────

Step 2 — Split-lane version (stable mode)

Now try the same lines, but separated:


[A] aaa
[B] bbb
[C] ccc


Send it a few times.

Most people notice: • no freezing • replies feel more flexible • tone stays consistent • structure follows your lanes

────────────────────────────

What difference did you see? Freeze? Flexibility? Stability?

Tomorrow I’ll break down why role-separated lanes work, and why GPT-5.1 responds so differently depending on prompt structure.