95% conformal prediction interval (global, n=21 calibration): 2,751,996 – 11,654,527 tCO₂. Restricted to steel-BF/BOF stratum: 4,711,522 – 6,807,400 tCO₂ (subject to small-n undercoverage).
Latest year supersedes 2022 as LODO label
iz-1 is a 2-layer browser-resident MLP with WebGPU-accelerated LoRA-shaped weights, trained on TR-MRV-Bench v0 (59 Turkish CBAM-scope facilities; 21 audit-grade Scope 1 labels with operator-PDF citations). Inputs: nameplate capacity, geo-coordinates, sector + steel route one-hot, Climate-TRACE-derived capacity factor and activity, operator-disclosed capacity factor (non-leaky in LODO), and Sentinel-5P NO₂ activity proxy. Evaluation: stratified leave-one-disclosure-out with 5 outer × 3 inner = 15 seeds per held-out facility. Headline: median 87% reduction in log-MAE vs. EU CBAM default; closed-form formula (capacity × route-EF × cf) reaches the same headline within statistical noise.
TR-MRV-Bench v0 is Apache-2.0 licensed. The benchmark, methodology, and per-facility data are public at
github.com/abgnydn/iz and iz-b0n.pages.dev.
Cite: Günaydın (2026), "TR-MRV-Bench: a public per-facility emissions benchmark for Turkish CBAM-scope
industry, with a physics baseline that beats the EU CBAM default by 87%."