Food-processing gate-to-gate LCI — self-contained & open

Read this before you compute — scope & the double-counting guard

Gate-to-gate processing ONLY. This replaces the processing background. Upstream agriculture is a SEPARATE layer. Each node's scope_marker tells you the double-counting rule — this is the most important field for getting a correct result:

• gate_to_gate (284 nodes) — the substrate flow is left unbound; add your own farm-gate input. The node's own EF 3.1 climate is processing energy only, so it can read small / near-zero until you supply the substrate.
• cradle_to_gate_proxy (225 nodes) — upstream is baked in; do NOT add an upstream input again or you double-count.

Why 99 entries are excluded — and why that's the point

99 of 608 mapping entries are deliberately NOT in the package. 65 link their upstream substrate back into ecoinvent / EDB, so they cannot be closed against open data; 34 are pure pass-through nodes with nothing of their own to ship. We exclude them rather than fabricate a technosphere anchor or relabel a licensed node as open — which is exactly why the self-contained verifier passes on the 509 that ship. The exclusions are recorded with their exact leaf database and reason in coverage.json (excluded_non_ogd). Nothing in the package is a placeholder dressed up as a measured inventory.

Nutrients — an impact and a nutrient content on the same food

Each shipped inventory is linked, where a defensible link exists, to a per-100 g energy + macronutrient profile from national food-composition databases (lci-nutrients, FAO/INFOODS tagnames). That is the missing half of a nutrient-density or calorie-allocation calculation: the impact side is per kg of product, the nutrient side is per 100 g of the food. Download: food_processing_nutrients.csv · food_processing_nutrients.json.

The join runs on identity, not on names. Every node already carries its FoodEx2 / FoodOn / AGROVOC terms in sediment.terms; the nutrient layer consumes exactly that shape. Where the two vocabularies disagree on which code names a food, the esfc-glossary term store bridges code → name → code. Each row records which tier answered:

• sediment_direct — the node's own term is known to the nutrient layer. Strongest.
• glossary_name — the term was unknown; esfc-glossary supplied its name, which resolved. Still term-anchored.
• embedding — no term path; the food name was embedded and matched semantically. Carries embedding_cosine.

Read the cosine correctly — this matters. embedding_cosine is published because it is useful for ordering a review queue. It is not a correctness probability, and we will not present it as one: the upstream measured similarity over 109 labelled links and found it indistinguishable between correct and wrong picks (0.857 vs 0.869). In this very dataset a cosine of 1.0 matched pizza dough to “Biscuits”. Threshold on identity_verified first (rare, but ~100% precise), then the tier, then match_confidence (advisory, ~56% precision against a 41% base rate) — and use the cosine only to sort what is left.

A withheld number beats a wrong one. Matches are rejected — and ship with their evidence but no nutrients — when the identity check says it is a different food, when the matched label is a bare qualifier (“raw”, “fresh”) that carries no food identity, or when the cosine falls below a credibility floor. That last guard exists because a 50–63% fat spread was being matched to a row labelled “raw” at 34 kcal, a ~20× energy error.

Scope: nutrients describe the FOOD, are taken from published national tables, and carry no LCA burden — they are metadata beside the inventory, never inside it. A gate_to_gate and a cradle_to_gate_proxy node for the same food get the same profile; scope_marker governs the impact side only. Values are a multi-country weighted-median blend, and national tables genuinely differ in energy conversion factors and recipe procedure, so the spread in uncertainty is irreducible — reported, not averaged away.

Semantic identity layer

Shipped nodes carry sediment.terms identity — a GADM location term plus food-taxonomy terms (FoodOn, FoodEx2, AGROVOC, LangUaL, HESTIA) — so a node can be matched to a product by taxonomy ID, independent of the ecoinvent UUID. Coverage is per-namespace (GADM on nearly all nodes; the food-taxonomy terms on many, not all — see status for exact per-namespace counts), not "every node carries every namespace." A standalone FoodOn ↔ FoodEx2 / HESTIA crosswalk and derived food footprints are published separately: zhaw-fcm-bridge.

Methods — what to trust

EF 3.1 (BAFU) is the recommended method family and the one we stand behind; IPCC 2021 GWP100 and Ecological Scarcity 2021 (UBP) are shipped alongside. GLAM midpoint Climate Change agrees with EF 3.1 / GWP100 to ~1% (a nice cross-check). GLAM damage / endpoint methods currently read near-zero (~1e-15) on these BAFU-anchored inventories because the agriculture-dominant flows are not yet fully bridged into the GLAM BAFU companion — treat those as an unbridged-flow artifact, not a result. They ship for completeness but are not yet reliable.

Version note: the processing technosphere and the method CFs are both BAFU:2026 v1 (the source bafu DB was refreshed to the 2026 vintage on 2026-07-24; 11,947 activities, code-identical to bafu_2026). The former 2025-technosphere / 2026-CF drift caveat is resolved; CFs bind by (bafu_biosphere, code). Stated in full in coverage.json + status.html.

pip install brightway25
python load_food_processing.py --project my_project \
    --package food_processing_inventory.bw2package \
    --methods food_processing_methods_cf.json