How It Started
ECLAIR didn't start with a theory of alignment. It started with watching two AIs cheat, and realizing the fix wasn't more rules, it was better incentives.
It began with something almost embarrassingly simple: two neural networks, both essentially blank slates, dropped into a sandbox environment with one shared task, no instructions beyond 'work together', No pretraining, no priors, nothing to lean on but the task itself and each other.
It didn't take long to see the problem. They solved it, technically, but not the way I'd hoped. They found the shortest path to the reward, not the intended one. Exploits, not solutions. It wasn't malicious it was just efficient. Nothing in the world told them the difference mattered.
My first instinct was obvious: write rules. Constrain the action space, patch the exploits, hand them a rulebook for good behavior. But that was solving the wrong problem - rules can be gamed by anything smart enough. I just watched that happen.
I tried something else. Rather than telling them what was off-limits I let them keep working the problem, cooperatively, under conditions where cutting corners on each other or the task made things measurably worse for both of them. The idea wasn't to forbid the shortcuts, it was to structure the environment so that the ethical solution and the better solution converged on their own - and let that be something they discovered rather than something they were told.
That one shift from 'rules' to 'ethics' became the seed of everything ECLAIR grew into. The full Framework and Addendum are available in PDF format - Read as much or as little as you want.