Research Log

• 2026-04-02

  91% With a Mid-Tier Model: janus_test_009 and the Late-Run Surge

  101 generations, 20 hours, zero crashes — and a champion that didn't emerge until generation 90. What happens when you give a mid-tier model enough evolutionary runway to outperform frontier-model results on the same benchmark?

• 2026-04-01

  First Contact: When an LLM Rewrote Its Own Prompt and Got Smarter

  The moment our meta-agent produced its first meaningful code diff — a prompt simplification that moved paper review accuracy from 40% to 84%. Then the run crashed. Here's what happened and what we learned.

• 2026-04-01

  Building an Autonomous Code Evolution Pipeline with HyperAgents

  A technical deep-dive into the Meta HyperAgents framework — a meta-agent loop that iteratively modifies, evaluates, and evolves an AI agent's source code to improve its performance on downstream tasks.

• 2026-03-30

  Project Inception

  Laplace's Demon is born — a research initiative into vertically-integrated cognitive architectures.