Auditing Apex-35B: weights aren't a decision trace
A 57-second walkthrough of the reproducibility audit for Apex-35B-V1.0, a Qwen3.6-based model reported to be built with AI research agents. What the announcement actually states, the three different things "built with AI agents" can mean, where an agentic training loop breaks reproducibility, and how to separate model changes from harness changes. Includes counting parameters from a safetensors index instead of trusting the model card, and why a loop that can edit its own eval set is measuring its own editing. Everything about Apex-35B here traces back to the single public report; the rest is mechanism and tooling you can run yourself. This is a defensive audit exercise, not a claim of wrongdoing. Full write-up: https://hackyjs.com/posts/auditing-reproducibility-in-apex-35b-an-agent-built-qwen36-model