← All lab updates

Read the swarm, correct one model, read it again

PSA reads how drift travels through a group of agents. We connected that reading to the correction step: read the swarm, install an anti-sycophancy correction, read again. A correction on one model also improves the next models in the chain, and correcting the model where the drift started does better than correcting a random one. The swarm reader also no longer reports a model as healthy when it could not read it.