Notes
What we learned from the systems we run.
Our current investigations. From work we are still running.
Most self-improving agent papers fail the eval traps they should already know
A methods audit of forty-one peer-visible self-improving LLM-agent papers under a five-trap checklist. Locked primary result: 39/41 (95.1%) fail two or more applicable traps. Seriora Research preprint.
What is an agent?
A lab note on the gap between next-token prediction and systems that perceive, act, and revise how they act. Model weights are not scaffolding. Autonomy is not agency.
When the score is wrong
Alignment fails quietly when models optimize proxies instead of intent. As reinforcement learning, preference models, and self-improving agents scale, that failure mode stops being a training curiosity and becomes the central reliability problem, because we still lack evaluations that can reliably tell learned-what-we-meant from hacked-the-score.
Harness-Bench preview
We ran the base Seri harness on Harness-Bench with GPT-5.6 Luna Pro. This note reports that untreated run.
How compaction works
Context is finite. Compaction is how an agent keeps the thread after it has already seen too much.