Can ICM Survive a Brain Transplant?

Okay — I think we accidentally have a really good ICM experiment.

Ox Alpha is getting a ridiculous amount of attention right now. I initially wasn't particularly interested in another "look how good this new model is at coding" conversation.

Then something clicked.

People are running it through agent environments like Claude Code.

Which means we can separate the brain from the system.

Most of what I've built with ICM has Claude sitting underneath it. And that leaves a question we've never really been able to answer:

How much of what works is ICM — and how much is simply Claude being Claude?

Then an insight popped into my head: 'let's test it.'

Take one mature ICM worker.

Same files.

Same rules.

Same examples.

Same methodology.

Same task.

Run it with Claude.

Then change the brain to Ox Alpha.

And DON'T optimize the ICM for Ox.

That's important.

If we have to rewrite everything to make Ox understand it, we've learned something.

If it completely falls apart, we've learned something.

If it follows the methodology but produces noticeably different judgment, we've learned something.

But here's the result that would really fascinate me: 'What if a completely different model enters the existing context and starts behaving recognizably like the same worker?'

The model supplies cognitive capacity.

But the methodology, decisions, standards, examples, boundaries and accumulated context are supplying something else.

We've transferred part of the capability into the environment.

Or, put another way: 'We changed the brain and the person still knew how to do the job.'

I have absolutely no idea whether that's what will happen.

Which is exactly why I think we should run it.

And I think we should decide the test criteria before we do.

What would you guys measure?

31 comments

Can ICM Survive a Brain Transplant?