My Microduck Journal — Development skill for Claude Code

my-microduck-journal

**A public, verifiable and continuable route to the Microduck's skills.** Every policy here was trained from scratch on one consumer PC (an RTX 50-series card, WSL2 on Windows 11), without the private teacher checkpoints the official recipe depends on, and is measured against the official policies with an open benchmark. The point is not the checkpoints — it is that the whole path is open: teachers, patches, exact scripts, checkpoint picks, evaluation, and every number and dead end, so that anyone with one consumer GPU can reproduce it, check it, or continue from it.

Fast variant (velstand_fast.onnx): pushed at 1.0 m/s on the way to a goal 1.5 m away, up again, on target at 7.5 s	Default (velstand_default.onnx, v4): the same exam, on target at 8.4 s

Contents

1. At a glance
2. Results against the official policies
3. What is different from the official setup
4. Status
5. How it was built
6. What we learned
7. Get the policies
8. Reproduce it
9. Repository layout
10. Hardware and timings

At a glance

• Policies you can use — release v0.2: one network that walks, gets up from any fall and crouches / tilts on command, in two versions: velstand_default.onnx (most robust) and velstand_fast.onnx (walks at the commanded speed).
• **Measured, not c