# Offline MoE candidate planner

This directory is intentionally CPU/stdlib-only. It parses a copied config
excerpt and prints candidate expert counts; it never loads tensors, downloads
weights, chooses experts, or rewrites a checkpoint.

```sh
cd work/ornith-research-20260923/moe-examples
python3 -m unittest -v
python3 plan.py fixtures/config-ornith-observed.json --prune 0.20 --prune 0.30 --prune 0.50
```

The observed config has 256 routed experts and top-8 routing. Integer rounding
explains why a nominal 20% removal is 205 survivors (51 removed = 19.92%),
while a nominal 30% candidate is 179 survivors (77 removed = 30.08%). These
are planning numbers, not published checkpoints.
