Docs / How a factor is decided

How a factor is decided

A method is not a number until somebody implements it against a flow list, and the people who do that do not always get the same answer. EF 3.1 says what kresoxim-methyl sprayed on a field does to fresh water; the European Commission's JRC says 134.73 CTUe per kilogram, and the ecoinvent Centre, implementing the same method against its own flow list, says 53,540. This list publishes both, each under the name of the team that said it, and then a third implementation of its own — the consensus implementation — which is the subject of this section.

Every number in that third implementation was arrived at by one of a small number of routes, the route is written on the factor as its derivation, and the routes fit on one page. This page is that page. Each stage links to the page that answers it in full; for what the tables and files are called, read Characterisation factors instead.

Three things are worth knowing before the first stage.

Nothing here is computed. No factor is averaged, weighted, interpolated or modelled. A number in the consensus implementation is a number somebody stated for that flow, a number a person ruled or signed for, or a number this list's own convention carried from the context next door — and the factor says which.

Publishing nothing is a normal outcome. Where two implementations disagree and nobody has ruled, where the model a method is built on says the number is wrong by a hundredfold, where only one transcription speaks and nothing vouches for it, the consensus implementation publishes nothing and asks. The question goes into a review queue, with the evidence; an empty factor is often a decision, not a gap.

The same rules apply to every method. EF 3.1 is rendered by four implementations and Stepwise 2006 by one, and the flowchart below runs the same way over both. What differs is only who is speaking.

The flowchart

For one substance, one impact category of one method, and one context of our taxonomy — cadmium, EF 3.1 freshwater ecotoxicity, emitted to a river:

flowchart TD
    A[Does any deciding implementation state a number for this flow here?] -->|no| G[Has a person signed this substance taking a donor's number, and is the donor published here?]
    G -->|adopt, and nobody characterised the substance anywhere| G1([adopted — the donor's number in this context])
    G -->|no entry, or the publisher characterised the substance| B[Does a convention say this flow takes another's number — a broader particle size window first, then a neighbouring context?]
    A -->|yes| C[Does the method's underlying model contradict it more than 100×?]
    B -->|no| B1([blank — counted, nothing published])
    B -->|yes, and the substance is published there| B2([carried — from that window or context, by that rule])
    B -->|yes, but nothing is published there| B1
    C -->|yes| C1([contradicted-factor queue — nothing published until ruled])
    C -->|no| D[Who speaks?]
    D -->|only the method's own publisher| D1([sole])
    D -->|several, within 2%| D2([agreed — the most precise number])
    D -->|several, and they differ| D3([contested-factor queue — nothing until ruled])
    D -->|only a transcription| E[Does the same substance already publish this number in another context of this category?]
    E -->|yes| E1([restated — an implementation said it for this flow])
    E -->|no| F[Has a person signed this substance taking this number from that donor?]
    F -->|adopt| F1([adopted])
    F -->|decline| F2([nothing — the decline is on the record])
    F -->|no entry| F3([proposed-factor queue — nothing until signed])

Read in words: a number is published because somebody stated it for this flow, because a person ruled or signed for it, or because our own convention says this context takes that one's number — and for no other reason. Nothing crosses from one substance to another without a signature. Nothing is published where a queue is still asking. A blank no rule names stays blank and is counted, not queued.

A signature comes in two shapes, and the flowchart has a box for each. Where a transcription stated a number for the flow — ecoinvent giving Copper, Ion copper's number under EF 3.1 — the signed entry answers that stated row. Where nobody stated anything for the flow — Stepwise 2006 names no zinc ion, and every ecoinvent zinc emission lands on Zinc(2+) — the signed entry says the substance takes its donor's published number in the same context, and the convention then treats the ion's other contexts as it treats the element's (#922). Both publish as adopted; both refuse to overwrite anything; the second also refuses a substance the method's own publisher characterised anywhere, because that publisher's silence about one category is a decision.

Figures on this page are from the build of 1 September 2026, run 20260901T0451332605450000, LCIA run 28e3b69f…: EF 3.1, with ecoinvent 3.12, ecoinvent 3.8, BAFU 2026 v1 and Stepwise 2006 merged onto it.

Copper is the worked example throughout, as it is for how a flow is decided, because its two substances take two different routes: the metal is stated by everybody, and the ion is stated only by ecoinvent.

1. Who is speaking, and does it count?

An implementation has a role. The method's own publisher is its reference — the JRC for EF 3.1, 2.-0 LCA consultants for Stepwise 2006. Anybody else rendering the same method against their own flow list is a transcription: the ecoinvent Centre's EF 3.1, and GreenDelta's. Ours is the consensus, and it is derived from the others.

Not every transcription counts. GreenDelta's EF 3.1 is published, compared and shown on every flow's page, and decides nothing: their package is a hand transcription nobody can re-run, so a number of theirs is never evidence for a number of ours. The method's own file says who decides, and the flowchart's first question — does any deciding implementation state a number here? — is asked of the JRC and the ecoinvent Centre for EF 3.1, and of one publisher for Stepwise.

Copper to surface water, freshwater ecotoxicity: the JRC states 46.47592828149169 CTUe/kg, the ecoinvent Centre states the same to the last digit, GreenDelta too. Two deciding voices, one number: published as agreed.

Where it is written down: implementations in data/lcia-impact-categories.json and data/stepwise-2006-impact-categories.json, with role and decides on each row.

In full: Who states a number, and whose counts?

2. Does the model underneath say something else?

Before two voices are compared, one check comes first, because agreement can be worthless: two transcriptions of one file agreeing is not evidence about the file. EF 3.1's toxicity categories are derived from USEtox 2.1, and where USEtox states something more than a hundredfold away from EF for a substance in every compartment, the consensus implementation publishes nothing for that substance and category, whatever anybody says.

Biphenyl is the case that found this. EF gives it 0.19957 CTUh/kg for non-cancer human toxicity from urban air — which would make an ordinary industrial chemical the third most toxic of the 3,380 substances EF characterises there — and USEtox is 1,400,000× away. Eight questions over four substances are held this way today, 110 factors, and none has been ruled.

Where it is written down: the contradicted-factor queue, and data/lcia-underlying-model-factors.json.

In full: What do we do when the model underneath says something else?

3. Do the voices agree?

Two or more deciding implementations state a number, and they are within 2 % of each other: the more precise printing is published as agreed, and the other is recorded beside it. Two percent is the same line the rest of this project draws between a rounding and a disagreement, and it is never crossed over a zero or a sign change — a stated zero against a number is two implementations disagreeing about whether something has an effect at all.

Where they differ by more, the honest artifact is a question. Kresoxim-methyl's 397× is the one this section opened with, and it is a contested-factor until a curator rules — which one did, in data/lcia-factor-rulings.json, with the reasoning: EF's own file carries the substance twice under two EC numbers, and 134.73 belongs to the row ecoinvent's flow corresponds to. So the consensus implementation publishes 134.73 as ruled, and the page that asked the question shows the answer.

A ruling is pinned to the numbers it was written about. If either implementation revises, the ruling stops applying and the row asks again.

Where it is written down: derivation: agreed with also_stated; the contested-factor queue; data/lcia-factor-rulings.json, whose 14 rulings settle 52 factors on this build.

In full: What do we do when two implementations disagree?

4. Only one voice, and it is not the method's own

The method's own publisher speaking alone is sole — 92 % of EF's consensus implementation, and all of Stepwise's, because the JRC's flow list is the base list and reaches every flow, while ecoinvent's reaches 7,925. Adopting the publisher's own number is not a decision.

A transcription speaking alone is. ecoinvent characterises flows EF's list cannot express — silvicultural soil, industrial soil, groundwater, and substances such as Copper, Ion that EF does not list — and for each of those the question is whether this list should publish ecoinvent's number on a flow the method's publisher never characterised. Three things can answer it without a curator looking at the row:

• the same substance already publishes that number in another context of the category, within 2 %, so ecoinvent is restating a number this list carries rather than proposing one — restated, 3,818 factors;
• a curator accepted the substance's numbers as a population, in data/lcia-factor-adoptions.json — adopted, 1,334 factors: 62 pesticides carrying a catch-all's numbers, and 23 minerals whose number is the JRC's element factors weighted by the formula;
• today, any flow of the category publishes the very same digits — which is how Copper, Ion takes copper's 46.47592828149169 as restated, and also how α‑hexachlorocyclohexane took lindane's 0.00019937 for non-cancer human toxicity, a substitution EF's own numbers refute wherever they can be compared. This third route is the one being retired, in favour of a signed entry per pair.

What none of them answers is a proposed-factor: 14 factors on this build, ten rocks EF never listed and kresoxim-methyl's 53,540 in silvicultural soil.

Where it is written down: derivation: sole / restated / adopted; the proposed-factor queue; data/lcia-factor-adoptions.json.

In full: What do we do when only one transcription speaks?

5. Nobody spoke at all

Ammonium to surface water is characterised at 2,493.2 CTUe/kg for freshwater ecotoxicity, and so is ammonium to unspecified water and to the aquifer. Ammonium to a river — a context BAFU brought, and no method's flow list has — carried nothing, in any category. That is a blank: the flow exists here, the same substance is characterised next door, and no deciding implementation states anything for it. Before the convention there were 4,859 of them for EF 3.1 and 10,849 for Stepwise 2006, whose export states one air row per substance.

Whether a blank takes a neighbour's number is not a fact about ammonium, and not a fact about EF: it is a fact about our own taxonomy, and it is written down once, as a convention, the same for every method — data/context-carry-rules.json. Each rule is a sentence: an emission to a river takes the surface-water number, failing that the unspecified-water number. Silvicultural and industrial soil take non-agricultural soil's number, then unspecified soil's; the rural and urban air compartments take unspecified air's; the class roots take their children's number only where every published child agrees. Some blanks are left blank on purpose — agricultural soil, the unconfined aquifer, aircraft cruise height — and the file says why. Long-term air and water, indoor air and the ocean are walls nothing crosses.

What ecoinvent did with the same blanks, and what SimaPro does, are recorded beside each rule as evidence. Neither is the authority: the convention was decided class by class from the census of the blanks, by the people who run this list.

Where it is written down: data/context-carry-rules.json; the blank counts per context on /factors/coverage; tools/count_context_blanks.py.

In full: What do we do when nobody characterised the flow?

6. Three passes after the decision

Three curated corrections run over what the flowchart published, and each writes its own word so a reader can tell it from a decision:

• a number derive() put on the wrong flow is moved to the flow it is about (data/lcia-misattributed-factors.json, 5 factors);
• a rounded printing one implementation states is replaced by the precise value its own source states — refined (data/lcia-rounded-printings.json, 8);
• a context nobody characterises is filled from the neighbour the convention names — carried (data/context-carry-rules.json, stage 5 above). Sulphuryl difluoride's 4,630 kg CO₂‑eq reaches the urban ground level from unspecified air this way, and does not reach indoor air, which is a wall.

7. Checking any of it

Every consensus factor carries its derivation; every queue row carries what each implementation stated and the ruling if there is one; every difference between two implementations is in the difference report; and every curated answer is a file in the repository with a mandatory comment, never a value typed into a database.

In full: How do I find out why a factor says what it says?

The pages in this section