Regression Test Suite
Full item-level results from the 600-item regression test suite. The suite checks calculation behaviour and, where a comparable baseline is available, rating stability and CO2 reproducibility across three test cases. These are internal checks, separate from the CAB’s independently drawn live sample.
Loading the run identity from portal-versions.json…
Suite v2.7 contains 600 declarations drawn against the released edb-to-bafu scope of v1.12.2 and carries all 965 base products reachable there (4 of the 969 are unreachable by design). v2.7 answers the review of v2.6: 243 caterer recipes and all 150 full-specification products were re-issued under new ids, each naming its v2.6 predecessor. Caterer recipes were re-read so that each reads as a coherent dish, full-specification products got plausible undeclared amounts, origins and names, and base products that no item carried were added where they fit. Representativeness: the 250 caterer recipes are drawn from real food-service production data (74 customer namespaces, 2024 onwards, deduplicated on composition) and stratified by namespace, √-weighted with an 8% cap per namespace — the case spans the recipe population Eaternity receives rather than reproducing its frequencies, and no single customer exceeds 7.2% of it. Every ingredient string is replaced by a different real term and amounts are jittered, so no item carries its source’s strings or amounts. The 200 label and 150 full-specification items descend from disjoint draws on a frame of openly published CH/DE/AT product declarations, stratified to a Swiss convenience-retail assortment and reconstructed with fantasy brand names and synthetic GTINs.
The four ratings under the scheme
Every item of the suite is rated in all four categories of the scheme: climate and water scarcity against their benchmarks (A–E), rainforest and animal treatment against fixed criteria (A/B/E, see the benchmark page). The nutrition score is not part of the scheme. Distribution in this run (600 items):
| Rating | Scale | This run (600 items) |
|---|---|---|
| Climate CO2e per Daily Food Unit, against the benchmark | A–E | A 239 · B 202 · C 140 · D 15 · E 4 |
| Water scarcity scarce litres per Daily Food Unit, against the benchmark | A–E | A 479 · B 73 · C 29 · D 13 · E 6 |
| Rainforest criteria-based: certified and uncertified soy/palm content | A/B/E | A 316 · B 0 · E 284 |
| Animal treatment criteria-based: certified and uncertified animal-derived content | A/B/E | A 251 · B 9 · E 340 |
This run is the state under review, so it is compared with no earlier state. From the next release on, the comparison and the materiality reading (ESFC EmpCo Rulebook v1.1, Section 12.4) cover all four ratings: rating changes for every category, and the value criterion for the quantities (CO2e and scarce water per Daily Food Unit).
Local Oracle ↔ EOS Parity
The offline Local Oracle reproduces the production EOS climate score across the 600-item suite.
Vintage: This parity measurement predates suite v2.4 and was made on the v1.2.0 population, not the 600 items shown above; it is evidence about the oracle, not about this run’s recipes. Parity was measured against the full production run of 2026-08-28 (a separate production reference, pre-re-baselining; not the 2 September validation-environment run) with the oracle's EOS-pick cache rebuilt from that same run. Known production defects at this vintage, stated rather than hidden: poultry by-products (chicken shank, chicken edible offal, paste broth) resolve to a ~22 g CO₂/kg, 0 kcal target — a matching defect (eos#21) whose fix will be the first post-baseline change record; and production changed its data state on 26.8. (EDB package 2026-05-04 → 2026-06-08 plus ~954 new name→term matchings), so the 2026-08-25 run measured the May package (butter 11.75, milk 1.95 kg/kg) and this run the June package (9.0 / 1.43, matching the July anchor; eos#22 tracks whether a cache recompute contributed) — which is why the validation baseline was re-pinned to this run on 2026-08-28. Ratings are compared across two ladders — the oracle rates on the declared 2026 benchmark ladder, production still rates on the pre-2026 anchor — so the rating-match figure reads as a declared-vs-live gap, not as oracle error. The re-pin against the re-baselined stack on the validation environment follows the final run.
Median absolute divergence by case
Largest divergences
| Recipe | Case | EOS (g) | Oracle (g) | Δ |
|---|
These are the worst-case items in the resolved sample — the median across all 598 resolved items is far smaller. Rows where EOS reports a small whole-gram value (e.g. 1–3 g per portion) show a large percentage for a sub-gram absolute difference; the genuine disagreements are the high-gram rows where a single ingredient resolved to a different background than EOS picked.
Items not used for the median
| Item | Case | Reason | Why excluded |
|---|
| ID ▲ | Title ▲ | Case ▲ | CO2 (g) ▲ | Climate ▲ | W · R · AT | Computed ▲ | Deviation ▲ |
|---|