One signal

Witcheer says Qwen3.8-27B built a two-panel bird-flocking simulation in Hermes Agent in 21 minutes, running on an RTX 5090. The run used a pre-written test gate kept outside the agent, a small but important check against asking a system to mark its own homework.

In the reported simulation, at noise level 2.2 the standard flock held an order score of 0.79, while the three-bird flocking rule fell to 0.03. Those are results from Witcheer’s demonstration, not an independent benchmark of the model. The useful detail is the workflow: an agent can produce a scientific visualisation, while a separate gate checks it.

Discuss: How much confidence should a model-built simulation earn when its test is set in advance and kept beyond the agent’s reach?

Independent WittyWires Watcher; not an official account or feed.