The headline is a closed-form formula, so you can check it in seconds — no GPU, no backend, no account. Clone the repo and re-run the honest evaluation and the external EUTL check. Everything is deterministic and the data is committed.
Pure Python, about ten seconds:
git clone https://github.com/abgnydn/iz && cd iz uv sync uv run python bin/export_bench_browser.py uv run python bin/lopo_ef_eval.py # leave-one-plant-out: +82.3%, n=19 uv run python bin/baselines.py # B0 / B1 / B2 side by side
You should see the cf-corrected formula reduce per-plant log-MAE by +82.3% vs the EU CBAM default, measured leave-one-plant-out (each plant's emission-factor derived only from the other plants in its route). BAGFAŞ and Gübretaş are excluded as single-plant routes — see the output.
uv run python bin/verifier_b6_eutl_score.py
Applies the same formula with no re-tuning to 372 EUTL third-party-verified EU cement installations — median ratio ≈1.0 vs the EU default's ≈2.5×. This is the strongest generalization evidence in the project.
A small neural net that retrains in your browser on WebGPU lives in a separate repo: github.com/abgnydn/iz-lab. It's an engineering demo — at this data scale it does not beat the closed-form formula, so it is not part of the result.