Journal · October 21, 2026
First runs: what we measured.
NEXUS v1 combines a base model with memory files that define its behavioral principles. No additional preference tuning is applied. The goal is to test how much of a model's behavior can be shaped through this setup, and where that approach stops working.
The first runs showed that voice and register were more consistent than instruction-following. The model could maintain a recognizable way of responding, but it did not always hold the question itself. Looking into the training pipeline revealed a mismatch: the two-pass scaffold used at inference was not represented in training.
The next step is v2. I'm working on a cleaner match between training and inference, followed by an evaluation battery that separates task performance from behavioral consistency. Until those tests are in place, these are observations from early runs, not conclusions about what the design can achieve.