Docs / How a flow is decided

How a flow is decided

A source list ships a row: a name, usually a registry number, a compartment, a unit, and the vendor's own identifier. Some forty decisions later it is a flow in the consensus list, with a settled identity, a chemistry it did not arrive with, and a link back to the row it came from.

This page is those decisions, in the order they are made. Each one is a question, and each links to the page that answers it in full. For the ordered inventory of what each processing step does, rather than the reasoning behind it, read Harmonisation steps instead.

Three things are worth knowing before the first stage.

No step edits a flow. Each one proposes a change, the run applies it and writes down which step made it and why, and where two of them write the same field the later one wins. That is what makes every stage below answerable after the fact rather than on trust: on the build quoted throughout this page, copper carries 65 recorded changes, written by seventeen different parts of the run.

Declining is a normal outcome. At most of these stages the honest answer is sometimes a question rather than a value, and the question goes into a review queue for a person. A field that is empty is often a decision, not a gap.

A flow enters through one of two doors, and almost every question anybody has about "the pipeline" is really about one door or the other.

Two doors: transformed, or merged

EF 3.1 is the base list, and it is transformed. Every one of its 93,993 rows goes through twenty-seven processing steps in a fixed order, and what comes out is the consensus list. Stages 1 to 7 below are that journey.

Every other list is merged. ecoinvent, BAFU, a SimaPro-lineage method file: their rows are matched against a consensus list that already exists, and a row becomes something new only when nothing it could belong to is found. Stage 8 is that journey, and stages 6, 7 and 9 then happen again on its terms.

The difference is visible in any flow's history. Copper came through the first door, and its change log names every part of the run that touched it. Copper, Ion came through the second, and its change log is empty — the chain had finished before ecoinvent's row arrived, and everything true of that substance was decided by the merge and the passes after it.

flowchart TD
    EF["EF 3.1 — the base list"] --> T["the 27 steps: identity, chemistry, names"]
    T --> L["split into substances and occurrences"]
    SRC["another list: ecoinvent, BAFU, SimaPro"] --> M{"which substance is this?"}
    M -->|"found one"| C{"which flow of it?"}
    M -->|"found none"| MINT["mint the substance"]
    M -->|"found several"| Q["report it: the row matched too much"]
    C -->|"found one"| ATT["attach the row to it"]
    C -->|"none fits its compartment"| NEW["add a flow where the row said"]
    L --> C
    ATT --> F["characterisation factors"]
    NEW --> F
    MINT --> F
    F --> P["published"]

Figures on this page are from the build of 25 August 2026, run 20260825T1354430622920000, revision 1c1385b: EF 3.1, then ecoinvent 3.12, ecoinvent 3.8 and BAFU 2026-v1 merged in that order.

Copper is the worked example throughout, because it goes through both doors: EF 3.1 ships the metal, ecoinvent ships an ion of it, and the two end up as two substances that share a number.

1. What did the source list actually ship?

Whitespace is stripped, capitalisation is regularised, and any manual fix the list carries is applied to the raw row first, because a fix names the vendor's own field names. A row that cannot be made into a valid record is dropped rather than guessed at. EF 3.1 spells copper copper; it is title-cased here, and the change log says which rule did it.

Then the row's own words are kept. Every flow carries a source_refs entry recording the list, the version, the vendor's identifier, and — verbatim — the name and the compartment strings as shipped. This is not bookkeeping for its own sake: it is what lets you audit a merge afterwards, and it is what makes the output usable as a translation table between two lists.

Nothing is deduplicated here. Two lists supplying the same substance in the same compartment produce two rows at this point, and stay two rows until there is enough resolved identity to collapse them correctly.

Where it is written down: elementary_flow_sources, and the Source lists panel on any flow's page. See Choosing sources.

2. Which compartment is this, in our words?

Every list names its compartments differently, and none of them uses this project's vocabulary. emissions to air / unspecified, air, Emissions to air, indoor all have to resolve to one identifier from a deliberately small controlled list — and from this point on that identifier, not the display strings, is what "the compartment" means.

The unit resolves at the same time, and a unit that resolves to nothing stops the run. A flow whose unit is unknown cannot be compared with any other flow's, and guessing one would make an incomparable pair look comparable.

Where it is written down: context_iri on the flow, and context_default_mappings. See Flow contexts.

3. Is this the substance the number says it is?

A registry number designates one substance by construction. A name designates whatever the person writing it meant. So the number outranks the name — and the hard part is that getting this wrong does not look like trusting a name, it looks like trusting CAS Common Chemistry, which is the most authoritative source of registry numbers there is.

Before anything is looked up, the numbers themselves are checked: CAS and EC both carry a check digit, transcription errors are proposed as corrections where the intended number is unambiguous, and each CAS ↔ EC pairing is tested against ECHA's own inventory. A row with no number at all can take one from its own name, but only where that name means exactly one substance — Borax is both the decahydrate and the anhydrous salt, and a name that means two substances has identified neither.

Copper arrives from EF 3.1 with 7440-50-8 and no EC number at all; the cross-check looks that number up in the inventory and fills in 231-159-6, saying so in the change log. ecoinvent's Copper ion arrives with no registry number at all, which turns out to be the whole of its later story.

Where it is written down: changelog, under the step that wrote it; disagreements in the commonchem-cas-name-differences, commonchem-name-cas, ec-malformed and ec-cross-check queues.

In full: What do we do when the name and the CAS number don't agree?

4. What is this substance made of, and what shape is it?

Now that the identifiers are trustworthy, they are used to look things up. ChEBI and PubChem supply synonyms, cross-references, formulae and structures — and this is where a mis-linked record would do the most damage, so four guards stand in front of it. A cross-reference somebody else made is not evidence. A ChEBI record whose formula disagrees with the flow's is not this substance. A PubChem compound whose own curated registry section does not list the number that found it is dropped. And a number that reaches several substances narrows before it enriches, or it enriches nothing.

Structure is then computed rather than asserted — formula, mass and InChIKey derived from a structure are arithmetic, not testimony — but only where identity is unambiguous enough to be certain what is being computed from.

Two sources will still draw one substance differently: with a stereochemistry and without, as an acid and as its ion, as a formula that cannot be written as a structure at all. Those are not one question, and the pages below take them apart.

Where it is written down: changelog under enrich_references, rdkit_* and opsin_iupac; formula_mismatches; and the cas-ambiguous queue.

In full: What do we do when one registry number reaches several substances? · What do we do when two sources draw the same substance differently? · What do we do when the formula cannot be drawn?

5. What should it be called?

The published name is chosen from candidates across ChEBI, the EC inventory, PubChem and Common Chemistry, and only where explicit confidence rules are met — typically several independent sources agreeing. Where they do not agree it does not guess; the case goes to a queue.

Two cleanup passes run before that choice rather than after it, so the scoring never sees a contaminated set of names: a synonym that is some other substance's preferred name is removed, and so is an element symbol on anything that is not the pure element. Two more run after everything else has finished writing names, to take out the strings that are catalogue entries rather than names — supplier grades, trade names with a product number, database accessions.

Where it is written down: changelog under consensus_match and the strip steps, and the consensus-match and undecided-label-replacement queues.

In full: What do we do with a name that belongs to something else?

6. Is this one substance, or two?

When the steps have finished, the flows are split into two layers: the substances, and the occurrences of them. Identity is the registry number, falling back to the normalised name — except for a flow whose name is exactly a nuclide, which is grouped by its nuclide, because both source lists put the element's number on some nuclide rows.

Three things then pull that grouping apart or hold it together:

• An origin qualifier splits a substance. Biogenic and fossil carbon dioxide are two substances, not one substance twice, and so are green, blue and grey water.
• A number two names share is a question, not a merge. 794 numbers were in that position on an earlier build; where a curator has ruled that they name two substances the flows are held apart, and where nobody has ruled they stay merged and go to the contested-cas queue. Staying merged is the conservative answer, because splitting a substance nobody asked to split strands every characterisation factor pointing at it.
• Each substance is typed from its chemistry, not from its name — which is what catches Chloride and Sodium ion, where a pattern over the label missed them.
• A compound and the product it was sold as are two substances. Chlordane the compound is 57-74-9; technical chlordane, the mixture of about a hundred chlorinated compounds that was actually sprayed, is 12789-03-6, and CAS publishes no substance record for it at all — which is what a registry does when a number names a manufactured product. Both stay, and then both names have to say which is which: Chlordane, pure and Chlordane, technical. The bare name stays searchable on the product, because the product is what an inventory reporting chlordane means (#786).

Copper is typed a chemical element. Copper, Ion is typed a monatomic ion. They are two substances, and the page on structures explains why an acid and its ion are not two drawings of one thing.

Where it is written down: run_stats under resolve_flow_layers, and the contested-cas queue. See Flow objects and elementary flows.

7. Is this flow one we already have?

Two flows of one substance, in one compartment, in one unit, are the same flow however differently they are named. One stays live and the others are marked deprecated with a pointer to it — deprecated rather than deleted, so a list that mapped onto the retired identifier still reaches the survivor.

A name is not part of that signature. What holds a near-identical pair apart is almost always the free-text note, and where the note is really prose rather than a version string, the pair goes to a queue instead of being collapsed.

Where two collapsing flows both published a characterisation factor and gave it different numbers, the number that is not published is kept on the one that superseded it. A value the pipeline declined to write is the same kind of fact as one it wrote.

Where it is written down: run_stats under deprecate_duplicate_elementary_flows, and the elementary-flow-collision queue.

8. Which flow does a row from another list belong to?

This is the second door, and it is two questions rather than one.

Which substance? Evidence is tried strongest first and the first rung that answers is the answer: a curator's ruling, then the curated tables for land classes, water materials and particle size windows, then the registry number, then EC, then the row's own names, then names the row does not have. On this build 13,547 of 15,609 placed rows were settled by the registry number alone and never reached a name.

Which of that substance's flows? Flows in another medium are dropped outright; if exactly one flow is in the compartment the row named, the row goes there, and 13,790 rows were settled that way. Only 63 rows in the whole build were decided by weighing candidates against each other.

ecoinvent 3.12's Copper ion reaches neither. It has no registry number, and no substance in the list answers to that name — so it matched too little, and the merge minted Copper, Ion with fourteen flows. When ecoinvent 3.8 was merged afterwards, its own fourteen Copper ion rows matched that new substance on its label. That is why merge order is a property of the lists rather than of the command line: the first list to reach a substance mints it, and everything after matches what it created.

Where it is written down: merge_outcomes, one row per source row, with every candidate and its score in detail_json.

In full: How does a row from another list find its substance? · What do we do when no existing flow fits the row?

9. What number characterises it?

A method is not a number until somebody implements it against a flow list, and two competent teams implementing EF 3.1 do not always get the same answer. So each factor says who states it, and where two implementations disagree this list publishes nothing and asks — that is the contested-factor queue.

Copper is the easy case: both implementations state 46.47592828149169 CTUe for freshwater ecotoxicity from a release to surface water, so the number is published as agreed. Copper, Ion is the interesting one. The JRC never characterised an ion this list only has because ecoinvent shipped it; ecoinvent gives the ion copper's own numbers; and because that number already stands published, bit for bit, on another flow of the same category, it is published as restated rather than queued as a proposal. All 48 of the ion's factors are of that kind.

Where it is written down: lcia_characterization_factors, with a derivation on every consensus factor, and the contested-factor and proposed-factor queues. See Characterisation factors.

In full: How a factor is decided — a section of its own, with the flowchart every consensus factor is a box of.

10. What happens when we cannot tell?

At every stage above, some rows do not meet the bar. Nothing is guessed. The case goes into a review queue with what a person needs to decide it, and the field stays empty until they do.

That is the design, not a shortfall: an artifact that states a number nobody can defend is worse than one that asks a question. What makes it workable is that the questions are enumerable, each carries its evidence, and every answer a curator gives is a file in the repository rather than a value typed into a database.

In full: What happens when the evidence is not good enough? · How do I find out why a flow says what it says?

The pages in this section