How to read these numbers
A figure without an interval is a claim. A figure with one is a measurement of how much is known — and on this site the interval is not decoration, it is the part that decides whether the number may be published at all.
Every published figure carries a 90% interval and a confidence grade, and both are computed, never chosen. The function that renders a figure refuses to run without the run id, the scenario, both ends of the interval, the grade, the methodology version and the data version, so no page on this site can print a bare number even by mistake.
The rule, and why it is code rather than a policy
A result with no interval is not published.
That sentence is enforced by the function that prints figures: it throws an exception if any of the six things a reader needs is missing. A template cannot work around it, because there is no other way to print a number. A grade of INDICATIVE renders as a range with no central value at all — the model has a plausible span for it and no best guess worth showing.
Is anything published yet? not yet. The engine computes these figures and no page shows them; publication is a later session, and until then the site shows what a product is made of and withholds what it costs.
What the last run actually did
Run 427, scenario S-MAX, year 2024, engine 0.24.0. The numbers below are that run's own parameters — not a description of them.
- Draws — 2,000 draws per figure (monte_carlo_pcg64, seed 20260921). The same seed gives the same interval, to the last bit.
- The ledger, deaths — σ 0.150 — measured, not assumed: the harm ledger publishes a low and a high for every event, and summed over a year they span a factor of 1.64 end to end.
- The ledger, injuries — σ 0.050 — measured, not assumed: the harm ledger publishes a low and a high for every event, and summed over a year they span a factor of 1.18 end to end.
- The economic structure — σ 0.23 — calibrated from published comparisons of input-output databases, because this one cannot be compared against a second system. See below.
What is sampled, and what each spread is made of
Three kinds of spread, and the difference matters: one is read from data that publishes its own uncertainty, one is calibrated from published research, and one is a convention attached to a confidence grade because nobody has measured it. The third kind is the first thing that should be replaced.
| Stage | Kind | Where the spread comes from |
|---|---|---|
| The harm ledger | measured | Conflict data publishes a low, a best and a high estimate for every event; health data publishes confidence intervals. Summed over the year the model runs on, that is the ledger's own interval. |
| Which industry the harm belongs to | graded | The attribution scores carry a confidence grade, and the grade sets how far the shares are allowed to move. |
| Which material an industry's harm belongs to | graded | Production data quality. The materials of one industry are sampled together, because they divide one quantity of harm between them. |
| How much of a material a product needs | graded | The grade on each line of the bill of materials — a measured composition moves less than a national average applied to an archetype. |
| World production of a material | graded | Whether the figure is reported, estimated or imputed by its publisher. |
| The structure of the world economy | calibrated | Held fixed, with one multiplier standing in for it, calibrated from published comparisons between input-output databases. |
| The scenario | not sampled | Deliberately. The assumptions that change a figure most — how many indirect deaths a war causes, whether harm is routed through the economy at all — are switches a reader can set, not error bars. Five scenarios show that sensitivity better than one wider interval would. |
Inputs that move together are sampled together. The ledger's own level is one draw shared by every material in every product, so a long supply chain cannot average its way to a false precision, and the materials of one industry move against each other, because a share that rises must take from another. Sampling them independently would make every interval narrower than it should be, which is the dangerous direction to be wrong in.
The one honest shortcut
The structure of the world economy is held fixed, and its uncertainty enters as a single multiplier.
Re-sampling the whole economic structure for every draw is possible and slow, and it would be a second-order effect next to the uncertainty in the harm data. So the table is held as published and one calibrated multiplier stands in for how much a different, equally defensible table would have changed the answer.
The check that would keep that honest is to run the same product through a second, independently built database and publish the difference. This site cannot do it: the licence of the obvious second system does not permit this project to use it. So the multiplier is calibrated from the published literature alone, and this paragraph exists because the alternative is to let a reader assume a cross-check that was never run.
Calibrated from: Giljum et al. (2019), Journal of Industrial Ecology 23(4) 946-958, doi:10.1111/jiec.12833 — a 15% corridor holds 46% of countries; carbon-footprint convergence (Moran and Wood 2014) is tighter and is NOT used
What a grade means
A grade is the weakest of three readings: how wide the interval is, how many attribution steps stand between a recorded death and this figure, and what the inputs themselves were graded. A figure is as good as its worst part.
| Grade | Means | Figures in the last run |
|---|---|---|
| HIGH | Primary sources, few steps, an interval within about ±40% of the centre. | 0 |
| MEDIUM | Modelled attribution, three or four steps, an interval within a factor of about three. | 0 |
| LOW | Sparse or proxied inputs, more than four steps, an interval spanning an order of magnitude. | 0 |
| INDICATIVE | Shown as a range only. There is no central value worth printing. | 0 |
No grade on this site is typed in by hand. Each is computed from the figure's own interval, depth and inputs, which is why most of them are low: the model knows how much it does not know.
What no interval on this site covers
These are not in any interval here, because nobody has measured them well enough to put a number on. They are listed because the uncertainties you do not list are the ones that get found.
- Indirect conflict deaths — The largest single unknown. Estimates from three to fifteen times direct deaths are all defensible, and the choice moves the headline more than anything else in the model, so it is a scenario switch rather than a constant hidden in an average.
- The default scenario itself — Routing all attributable harm through the economy is a position, not a measurement. The scenario menu is its sensitivity analysis.
- Informal and artisanal mining — Barely present in any statistical system, and exactly where conflict and occupational harm concentrate. Under-counted; the direction is known and the magnitude is not.
- Under-reporting of occupational harm — Severe, and uneven by country income. Corrected by a scenario, not by a fudge factor.
- Everything after a product is made — These figures cover production only. A car's use — road deaths, exhaust — is very likely the larger half of its toll and is not counted anywhere on this site yet. Every page that shows a figure says so.
- What the supply chains do not reach — A bill of materials traces what it names. Where a product's chain stops at an input the model has no data for, that mass carries no harm at all, and each product page publishes how much of itself it could trace.
Questions readers ask first
What does the interval beside a figure mean?
It holds the middle 90% of 2,000 draws of the model, each drawing its uncertain inputs anew. It measures what the model knows, not every way it could be wrong: harm it cannot attribute is not in it at all, which is why every figure is also a floor.
Why only two significant figures?
Because the model does not have three. A third digit would claim a precision the inputs cannot support, so the function that prints figures rounds once, to two, in every language.
What does the confidence grade say?
How far to trust the figure, from the weakest of three things: how wide its interval is, how many attribution steps stand between the recorded harm and the product, and how good its inputs are. It is computed, never chosen.
Why do some figures show only a range?
A figure graded INDICATIVE has a plausible span and no central value worth quoting, so the site prints the span alone and will not print a single number for it.
Can two products be compared by their intervals?
Not by looking whether they overlap: the draws behind them share the same ledger, so they move together. The comparison page draws the two together and reports the interval of their ratio instead.