Docs / How a factor is decided

What do we do when two implementations disagree?

Part of How a factor is decided. Two deciding implementations state a number for the same flow, category and place — this page is what happens when the numbers are not the same.

Vanadium to air, ionising radiation aside, is as plain a metal as EF 3.1 characterises. The JRC states 1.3384 × 10⁻⁶ CTUh/kg for non-cancer human toxicity from unspecified air; the ecoinvent Centre states 1.3384 × 10⁻⁵. The mantissa is preserved to five significant figures and the decimal point has moved — in five compartments and both halves of the category, always by exactly ten (#333). Nobody disagrees about vanadium. Somebody slipped a unit.

Figures are from the build of 1 September 2026, run 20260901T0451332605450000.

Two percent is agreement; more is a question

Where every deciding implementation's number is within 2 % of the others', the implementations are saying one thing and the most precise printing is published as agreed, with the other recorded on the factor as also_stated. 0.000118 and 0.00011755 are one number written twice, and this keeps the second — the same judgement the merge makes about two rows that turn out to be one flow. Two percent is the line the whole project draws between a rounding and a disagreement, and it is not crossed over a zero or a sign change: a stated zero against a number is not a rounding of anything, it is two teams disagreeing about whether something has an effect.

Where they are further apart, the consensus implementation publishes nothing and the row goes to the contested-factor queue. Taking the method's own publisher by default would publish numbers nobody had looked at under a label saying a decision was made; the queue publishes a question instead. On this build that is 38 questions over 152 factors, 32 of them still open.

Grouped by substance and category rather than by flow, because that is the unit a person decides in: kresoxim-methyl's freshwater ecotoxicity is one question asked about two compartments, not two questions. The queue row carries every implementation's number, the ratio, whether the same disagreement holds in every context — which is what makes a broad ruling possible — and where the flow came from.

What the disagreements turn out to be

Almost never toxicology. The ones ruled so far are all the same kind of finding:

• kresoxim-methyl, 397× — EF's own file carries the substance twice, under one name, one CAS number and two EC numbers, with different factors on the two rows. This build kept one row's number and the ecoinvent Centre kept the other's, in every compartment and never a third number. Ruled to the JRC's 134.73, because that is the row ecoinvent's own flow corresponds to (#377).
• vanadium, 10.0000× — the unit slip above. Ruled to the JRC's.
• sulphuryl difluoride, climate change — ecoinvent states a global warming potential of 4,630 kg CO₂‑eq where the JRC states none; ruled to publish, with IPCC AR6 cited on the ruling (#728).

And two still open that are real, in the sense that the numbers describe different things. Fenpropimorph in soil differs by 47× between the two implementations, because EF lists the fungicide twice under two registry numbers and the two teams characterised different copies; fenoxycarb differs by 740–800× in 22 compartments for the same reason (#645). Both are identity questions wearing a factor question's clothes, and the answer to each begins in the merge rather than in this queue.

What a ruling is

An entry in data/lcia-factor-rulings.json, keyed on the queue, the substance and the category, with a mandatory comment — every ruling is a statement about somebody else's science, and one with no reasoning is unreviewable.

Two verdicts, and neither names a source list. publish names the implementation whose number this list takes; decline takes nobody's. The first design had a verdict per publisher, which reads well with exactly two and cannot express a third; publish with an implemented_by says the same thing about any number of them. It also makes one verdict sayable that the old shape could not: decline on a contested question is "neither of them, deliberately".

A ruling is pinned to the numbers it was written about. ruled_about records what every implementation stated at the time, and if either revises — or a third implementation appears — the ruling stops applying and the row asks again. A ruling that the staleness check could never fail would be the same as having no check. Fourteen rulings settle 52 factors on this build, and none has gone stale.

A factor a ruling publishes carries derivation: ruled, distinguishable from agreed (the pipeline found nothing to decide) and from sole (nobody else spoke). The queue row stays on its page marked ruled, with the verdict and the comment: an answered question is a record, not a deletion.

Where it is written down

• lcia_characterization_factors.derivation — agreed with also_stated, or ruled.
• /queue/contested-factor — the open questions, worst ratio first, and the ruled ones after them.
• data/lcia-factor-rulings.json — the answers, with ruled_about and the comment.
• /factors/differences — every disagreement, ruled or not, banded by ratio.