Teacher‑reviewed research from the mnemo project. Every number is a snapshot you can re‑run and disagree with — the raw notebooks live on GitHub.
Estimated energy per token across the free‑model swarm vs. a frontier model — EcoLogits + ML.ENERGY methodology, honest error bars. The swarm runs 50–200× leaner; the teacher review cuts that headline down to an upper bound and says why.
The OpenRouter catalog listed 22 free models. A one‑sentence practice test found that listed is not working — exhausted pools, silent hangs, and four things “everyone knows” that the API response contradicts.
Retrieval without embeddings, per‑task token budgets, the validator gate.