Docs / How a flow is decided

What do we do when no existing flow fits the row?

Part of How a flow is decided, and the second half of bringing in another list. The first half found the substance; this page is about which of that substance's flows the row is really about — and what happens when none of them is.

A substance is published once. It occurs many times: in air and in water, at ground level and at stack height, in kilograms and in cubic metres. Those occurrences are the elementary flows, and a source row names one of them by its compartment and its unit.

Figures are from the same build as the previous page: 25 August 2026, run 20260825T1354430622920000, revision 1c1385b. Of the 15,609 rows the algorithm placed, 13,790 were settled by the row's compartment alone, 1,756 ended in a new flow being created, and 63 were decided by scoring — which is worth knowing before reading the scoring rules below, because they are the part of this that almost never has to run.

A row is never crossed into another medium

Before anything is compared, every flow of the substance in a different dimension or a different medium from the row is dropped. An emission to water is never answered by a flow in air, however well the rest of it matches.

ecoinvent's Benzoylprop-ethyl in forestry soil is a plain example: the substance has thirteen flows, ten of them are in air or elsewhere, and only three survive to be considered at all.

If the row's compartment is one this list holds, the row goes there

13,790 rows — 88% of everything the algorithm placed — are decided by this one sentence. The row named a compartment, exactly one flow of the substance is in it, and there is nothing to weigh.

The short-circuit declines when two flows share that compartment, because then the compartment has not identified anything. That is not hypothetical: EF 3.1 holds water vapour in rural medium-stack air twice, once in kilograms and once in cubic metres.

Otherwise the candidates are scored

Four terms, and their relative sizes are the argument:

The unit is worth more than the compartment, and the water vapour case is why. ecoinvent reports Water to rural medium-stack air in cubic metres. Both EF flows are in exactly that compartment, so the compartment term is a tie at +1 and cannot separate them; the unit does, and the cubic-metre flow wins with a score of 7 — two for the unit, four shared words, no words unaccounted for, and one for the compartment being exactly right.

The word counting is the reason the compartment term exists at all. Counting shared words measures how deep two compartments agree, not whether they are the same one: Air / Unknown and Air / Long-term and Air / Aircraft cruise height all share two words with a row that said Air and differ by one. Only the compartment's identity can say which of those is the place the row named.

Coarsening is honest; contradicting is not

Publishing a release to a lake as a release to water, unspecified, loses detail and says nothing false. It is the right answer when the list has no lake.

Publishing it as a release to groundwater is a different act. Groundwater is not a vaguer way of saying lake, and the score cannot tell the two apart, since all it counts is shared words — for BAFU's lead-210, the groundwater flow outscored every honest alternative on the strength of "an environmental release to water" plus a matching unit (#432).

So a winner that contradicts the row is thrown out and the choice is made again among the candidates that do not. This fired on 89 rows.

A row's answer must not depend on when it was asked

Coarsening onto the nearest compartment that exists has a second problem: the set of flows to be nearest to grows as the merge runs.

ecoinvent's Aerosols, radioactive, unspecified is the demonstration. The identical row, in air / urban air close to ground in both releases, landed on Air → Unknown during the 3.12 pass, where that flow won 3 to 2. By the 3.8 pass a third flow of the substance existed, two candidates tied at 3, and the tie put the row in its own compartment. One intervention, two consensus flows, and the only difference was the order (#566).

So a row whose compartment this list holds gets a flow in that compartment, whenever no candidate is already there. Twelve rows on this build. A row whose compartment the list does not hold is left alone, because there is no compartment to create it in.

Four situations create a flow, and one deliberately does not

1,756 rows ended with a new elementary flow rather than an existing one:

Benzoylprop-ethyl is the first row of that table. Its substance is in agricultural ground, non-agricultural ground and unspecified ground; the row is in forestry soil, which resolves to silvicultural ground; two of the three tie, and the row gets a silvicultural flow of its own.

What does not create a flow is the case one layer up — a row whose substance could be several. Those two look alike and are opposites. Several candidates mean the row matched too much, and creating something would be choosing by another name. A row that matched too little has a gap in front of it, and the gap is what gets filled.

Where it is written down

merge_outcomes.detail_json, per row. It records the compartment the row named and what it resolved to, how many flows of the substance there were, how many survived the medium filter, and — for every candidate that was scored — its score with each of the four terms separately. A vetoed winner is named too, under rejected_contradicting_context_iri, so a report shows what the decision chose between rather than what it first considered.

That is enough to re-run a matching decision offline, against a build that has already happened, without rebuilding anything.