Docs / What we have found

A characterisation factor that does not belong to the substance it is on

Part of What we have found, the class where the row is right about the substance and the number on it was computed for something else.

The method disagrees with the model it says it implements

EF 3.1's three toxicity categories are USEtox 2.1 — the JRC's report says so, and describes the small deliberate adjustments made on top. For the two human-health categories the LC-Impact result workbook is a second copy of the same model, and comparing the two is what turns "this number looks large" into a measurement: both count expected cases of disease per kilogram, one directly and one as the years of healthy life those cases cost, and the constant between them is the model's own severity factor. That is what makes the comparison trustworthy at all, and it is why only human toxicity is compared — the workbook's freshwater ecotoxicity numbers are LC-Impact's own ecosystem-quality result, a potentially disappeared fraction of species where EF's CTUe is a potentially affected fraction, so a ratio between them measures the distance between two models rather than anything about the substance.

For four substances — biphenyl, o-phenylphenol, benfluralin and the 2,4/2,6-toluenediisocyanate mixture — the two are more than a hundredfold apart in every compartment. The worst is biphenyl, two benzene rings joined together, used in dyes, plastics and as a citrus fungicide:

Both columns count expected disease cases per kilogram emitted. EF 3.1 is therefore saying that a kilogram of biphenyl into city air causes 0.2 cases of disease — a tonne would cause two hundred. Ranked against the other 3,380 substances EF characterises for non-cancer human toxicity in that compartment, biphenyl comes third, above mercury, and just below a rodenticide designed to be lethal in milligrams. Seventh on the same list is 2-phenylphenol, which is biphenyl with a hydroxyl group added, 170,000× above USEtox. Two close relatives both anomalous is unlikely to be coincidence.

A substance ranked third of 3,380 dominates any inventory that contains it, and the error is invisible in a result: it looks like a genuine hotspot (#552).

What is deliberately not in that list, because a difference is not automatically a defect:

• The metals. All 27 disagree with USEtox — typically 25–75× — and the JRC report says why: EF 3.1 replaced USEtox's metal factors with ones derived from EU data and tabulates every change. That is a decision, and it is documented.
• Differences in one compartment but not another. Those are the two models routing an emission differently. A substance only counts here if every compartment disagrees, which no fate difference explains.
• Freshwater ecotoxicity, however wide the gap. 34 further pairs were measured and none of them is a finding: the two publications are not counting the same thing there, and the constant between them is fitted across two models rather than converting within one. Over the substances not in dispute, 55.3% of the cancer factors and 32.5% of the non-cancer ones land within 1% of the model, against 20.0% of the ecotoxicity ones. A number that is far from a differently defined number is not evidence that it is wrong.

Where a naming error stops being cosmetic

The factor is where a naming error stops being cosmetic. Every class on the pages before this one changes which substance a row is; this class changes the number that row contributes to a result. Four shapes turn up on this page, and each of them was found a different way.

Two implementations of the same method disagree

EF 3.1 is published by the JRC and re-implemented by others, and the re-implementations do not always agree. Over the four-list build of 2026-08-29 the published implementations agree exactly — same float, full precision — on 36,065 of the 36,744 triples more than one of them states, and the 679 that differ are what this class is about. Thirty-six of them are more than 100× apart. The clearest is EF's duplicated kresoxim-methyl: EF ships the substance twice under two EC numbers, each row with its own factors, and the two implementations read different rows — 134.73 against 15,445 for freshwater ecotoxicity in one compartment. That one carries a ruling, and the ruling publishes one of the two numbers rather than pronouncing on anybody's ecotoxicology.

A claim withdrawn, and why it is kept here. Until 2026-08-18 this section reported a second case: ecoinvent's implementation giving Vanadium V a human-toxicity factor exactly ten times the JRC's — 1.3532e-05 against 1.3532e-06, and ten more pairs, the ratio 10.0000 in every one. The argument was that a modelling choice about vanadium speciation would not preserve the mantissa to five significant figures across six compartments, and that preserving the digits while moving the decimal point is what a unit slip does.

The digits were right and the conclusion was wrong. EF 3.1 ships two vanadium flows in every context — vanadium, 7440-62-2, and vanadium (v) — and states the smaller number for the first and the larger for the second, in its own workbook, in the same compartment. The factor of ten is EF's own speciation factor for vanadium(V), derived as ten times the element's, which is precisely why the mantissa survives it. ecoinvent's implementation was reading the right EF row all along; this project was comparing it against the wrong one, because the correspondence table sent ecoinvent's Vanadium V rows to the element. With the routing corrected the two agree to every digit, and the two rulings that had been written about it were withdrawn (#333, #560).

The general lesson is worth more than the finding was: a comparison between two implementations is only as good as the mapping that decides which two rows are the same question. A ratio computed across a wrong pairing looks exactly like a finding, and looks more like one the cleaner it is.

The factor was computed for a different substance

This is the trichloroethane case seen from the factor side: EF's five mislabelled flows carry 1,1,2's toxicity numbers and 1,1,1's ozone-depletion factor, because the toxicity method was keyed to the registry number and the ozone method to the name. Where two implementations are compared, this shows up on its own: putting the nine ecoinvent rows on the right chemical made seven factor questions disappear — twenty-four factors on which the ecoinvent Centre had been characterising 1,1,1-trichloroethane's rows against the JRC's numbers for a different chemical.

The factor belongs to the substance and not to the place it was taken from

The three above are about the substance. This one is right about the substance and wrong about the compartment, which is harder to see because nothing in the number looks odd.

Resource use, minerals and metals is abiotic depletion measured against ultimate reserves, and the reserve is the continental crust: an element's crustal concentration times the mass of the crust. The ocean is not in it, and EF 3.1 says so in its own flow list — it holds Non-renewable element resources from water as a compartment of its own, and characterises nothing in it. Both independent implementations read it that way: the JRC gives bromine from the ground 0.00439 and bromine from water nothing, and the ecoinvent Centre gives Magnesium, in ground 2.02 × 10⁻⁹ and every in water element flow nothing.

One flow breaks the pattern, and it is derived rather than stated. Carnallite is the Dead Sea salt ecoinvent's Israeli magnesium takes 9.5 kg of per kilogram of metal, filed from water by ecoinvent and by EF 3.1 alike. The ecoinvent Centre characterises it at 1.0375 × 10⁻⁵ where the JRC states nothing — and the number is chlorine's crustal factor scaled by carnallite's chlorine content: KMgCl₃·6H₂O is 106.35 / 277.84 = 38.3% chlorine by mass, and 2.71 × 10⁻⁵ × 0.383 = 1.04 × 10⁻⁵. The v3.12 implementation report (Table 20) describes the rule: a mineral's factor is composed from its formula. The rule reads the formula and does not read the compartment, so it fires on a salt dissolved in the sea — and the same implementation, one row away, gives chlorine from that same compartment nothing.

The decision, and it is the exception in these findings. The factor is declined in lcia-factor-rulings.json, so this list publishes no minerals-and-metals number for carnallite taken from water, while the ecoinvent Centre's statement of 1.04 × 10⁻⁵ stays published under their name (#664, #714). A decline withholds this list's own number and erases nobody's.

What this project does with the other three. Nothing, mostly, and deliberately: this list publishes what implementers state. A factor is not overruled because it looks wrong; it is recorded, compared against the other implementation and against the underlying model, and where they contradict each other the disagreement is published as a question rather than silently resolved. The four USEtox cases are recorded with both numbers and no verdict. The carnallite decline is not a counter-example: a factor only one implementation states is published by nobody until somebody answers for it, and the ruling is that answer written down rather than left to stand by accident.

What it cost, by implementation: