# HumanToll — Methodology

**Version 1.0 · 2026-10-14**
Quiescence · https://quiescence.eu/humantoll/

> **Status: version 1.0, the first under which figures are published.** It is frozen once the
> model has been run across 2010–2024 and its results have passed the validation gates in §10
> (the backtest). A frozen version never changes: a correction is a new version, the old one
> stays readable at its own address, and every figure names the version it was computed under.
> Versions 0.1 were the drafts, 2026-09-08 to 2026-10-14.

---

## 1. The research question

> **How much human harm — deaths and injuries — can be attributed, through documented
> causal pathways and a global supply-chain model, to the production and use of the goods
> and services people buy?**

The question deliberately is not *"what causes war?"* nor *"is this product harmful?"*.
Those are arguments. This is an accounting problem with a stated method, an uncertainty
range and a scenario switch — which is what makes it answerable, and what makes a
disagreement about it productive.

## 2. What counts as harm

A death or injury enters the ledger only if it carries a **population attributable fraction
(PAF) against a modifiable exposure** — something economic activity produces, distributes,
or withholds.

This is how the instruction *"exclude natural deaths"* is implemented: as a data rule, not
an editorial judgement. Everything with no such exposure — cardiovascular disease of
ageing, cancers with no attributed exposure, congenital conditions, senescence — is
excluded **by construction**, because there is nothing to attribute it to. Nobody decides
what is "natural"; the absence of an exposure–response relationship decides.

One consequence must be stated wherever a figure is published:

> **The model systematically under-counts.** Harm that cannot be attributed is dropped.
> Every figure is a floor, never a total.

### 2.1 The four harm channels

| | Channel | Contents | Primary evidence |
|---|---|---|---|
| **C1** | Conflict and political violence | battle deaths, one-sided violence, non-state conflict, indirect deaths (famine, disease, displacement), conflict injuries | UCDP GED + ACD |
| **C2** | Work | fatal and non-fatal occupational injury and disease, by ISIC industry | ILOSTAT, Eurostat ESAW |
| **C3** | Poverty (deprivation) | deaths and illness attributable to undernutrition and to unsafe water, sanitation and hygiene | FAOSTAT, WHO/UNICEF JMP, World Bank |
| **C4** | Pollution (exposure) | ambient particulates, occupational carcinogens beyond C2, heavy metals and chemicals from extraction and processing | World Bank, Eurostat, peer-reviewed exposure–response functions |

C1 is **directly counted** and carries no PAF: a battle death is observed rather than
attributed. C2 is directly reported. C3 and C4 enter **only** through a PAF, which is what
makes "deaths from disease" a defined subset rather than an uncountable one.

### 2.2 The fifth channel is future work

**C5 · Use phase** — harm from *using* a product rather than making it: road deaths,
combustion products, firearms, tobacco, alcohol, product misadventure. Specified but not
implemented. Its dimensions (`lifecycle_phase`, `harm_bearer`) exist in the schema from the
first migration so that adding it later is a fill-in rather than a rebuild.

**Until C5 ships, every published figure is labelled *production phase only*** and is never
presented as a lifetime total. For some products — a car above all — the use half is very
likely the larger one, and implying otherwise would mislead in a specific direction.

## 3. Metrics

Two are headline and are published with equal prominence. Three are advanced.

| | Metric | Unit |
|---|---|---|
| **1** | **Lives** | deaths |
| **2** | **Injuries** | non-fatal disabling injury and illness |
| 3 | YLL | years of life lost |
| 4 | YLD | years lived with disability |
| 5 | DALY | YLL + YLD — the only defensible way to add metrics 1 and 2 |

### 3.1 Denominators

Every product result is published per unit, per 1 000 units, per 1 000 000 units, per
kilogram and per €1 000 of consumption — and **inverted**: *one death per N units*, *€X of
spending ≈ one death*. The inversion is not rhetoric; it is the same number, expressed in a
form a person can hold. Money denominators are deflated to a fixed base year and converted
at annual average rates, so a result does not move because a currency did.

## 4. The attribution chain

```
HARM EVENT  (C1–C4, deaths + injuries, with uncertainty)
    ├─ PAF / exclusion gate  ────►  dropped if no modifiable exposure
    ├─ overlap and dedup     ────►  §6
    ▼
ECONOMIC ATTRIBUTION   S[c,s]     harm subject c → economic sector s,  Σₛ S = 1
    ▼
MATERIAL ATTRIBUTION   R[s,m]     sector s → raw material m,           Σₘ R = 1
    ▼
    D[c,m] = D[c] · Σₛ S[c,s] · R[s,m]
    ▼
INTENSITY              h[m] = D[m] / Q[m]     Q measured on the MRIO's own basis
    ▼
SUPPLY CHAIN           L = (I − A)⁻¹          multi-region input–output
    ▼
PRODUCT FOOTPRINT      H[p] = hᵀ · L · y[p]
    ▼
UNCERTAINTY            Monte Carlo over every stage above
```

`S` is scored, not asserted: nine criteria (territorial dependence, export dependence,
government-revenue dependence, conflict-location proximity, actor statements, resource
control, trade dependency, arms relationship, historical evidence), each 0–10 against a
published rubric, each carrying an evidence URL and a named analyst. Fourteen wars are scored:
twelve chosen to cover the range of the argument, and — since engine 0.30.0 — the war in Tigray
(2020–22) and the central Sahel insurgencies (2017–25), the two largest the backtest found reaching
products only by assumption. **One scored period can cover several UCDP conflicts** when they are one
war with one economy: UCDP codes the Donbas war of 2014–22 as three conflicts beside Russia – Ukraine,
and the Sahel's as six state-based ones across three countries. The scores stay one judgement; every
member conflict's deaths follow them, each in the economy of the country where it was counted.

`R` is divided out of what each country produced that year, **by value: tonnes × price, in every
industry.** The product owner, 2026-10-05: *"switch mining to value"*; 2026-10-06: *"do everything by
value if results are more consistent"*. By weight a tonne of gold counts as much as a tonne of iron
ore, and a tonne of sugar cane — mostly water — as much as a tonne of cocoa: Mexico's 2022 metal-ore
mining went 86% to iron ore and 0.0% to gold, and sugar cane took 47% of Brazil's harm in 2016. By
value Mexico's metal mining goes 33% to copper, 27% to gold, 21% to silver and 6% to iron ore, and
Brazil's harm to its livestock, oil and soy. Value follows what harm follows — the work, the land,
the control of what is produced — far better than the tonne. Compared year by year over 2010–2024,
the two bases are equally stable and place the same share of harm on named materials (plan/14 S32);
what differs is that by value a country's harm lands on what its economy is made of.

The prices: for crops and livestock, FAO's farm-gate producer price in each country and year (milk at
FAO's price for raw cow milk); for mining materials, the world unit value of the traded form that
measures what production counts (BACI, CEPII: value ÷ tonnes of every flow between countries,
1995–2024), converted by chemistry or by the benchmark grade the publisher prices against; for
timber and charcoal, FAO's world export unit value, per m³ and per tonne. A year outside a price
series takes its nearest year. A country price more than 20 times the world median for its
material and year, either way, is a unit or currency error rather than a price (22 of 7 589 in
2010–24, such as Zimbabwe's 2022 grains) and is replaced by the world price. A material with no
price at all is left out of the split it would have entered. The intensities `h[m]` stay per tonne
of production.

**Harm is placed only on what the country where it was counted produces — never on a world mix.**
The product owner, 2026-10-06: *"attribution must never take place on the worldwide basket, but
only on the country specific basket."* Where a country has no production of its own in the
industry a war is scored to, that share goes to `OTHER`; under `S-MAX` the general rule (§4.2)
then gives it the country's own production basket. One war met this case with any weight:
Afghanistan's, scored partly to quarrying, where Afghanistan has no recorded production (its lapis
lazuli and emeralds have no tonnage series anywhere). Until engine 0.29.0 its share took the
world's quarrying mix — 32 780 deaths in 2010–24 on lithium, salt, sand and diamonds, none of which
Afghanistan mines.

### Where harm is placed: in the country where the killing takes place

The product owner, 2026-10-05: *"all harm inside of the country(es) where the killing take place."* A death is
counted where it happened — UCDP records the place of the violence — and the attribution keeps it there. A war's
share for an industry is given to that industry **in the country or countries where the people died**, and so to the
products made from that country's production. A war fought in Ukraine is charged to Ukraine's oil and gas, coal,
grain and steel, not to the Russian exports that pay for it; the Russian soldiers killed in Ukraine are counted in
Ukraine. The one exception is the arms industry, placed where the weapons were made (§4.1), because a weapon is
itself a product with a maker.

**This rule is wrong, and it is as wrong as the war itself.** It puts the cost of a war on the economy of the people
it is fought among, not on the economy that decided it or pays for it. That is exactly what a war does. A war of
aggression moves its cost onto the land, the work and the lives of the country it is fought in: the dead lie in that
country's fields, mines and towns because the war put them there. To move them in the accounts back to the
aggressor's exports would describe what ought to be, not what is — and it would need a theory of responsibility
(the aggressor, its financier, the buyer of its oil) that the data cannot test and that every side of every war
disputes. This project measures where harm falls and which economy it falls on, so that a product made from that
economy carries it. The injustice of that placement is the injustice of the war, shown rather than corrected.

What follows:

- products made from a war-torn economy carry its war — Ukrainian grain, Sudanese gold, Congolese copper;
- a country's figures are harm **suffered** there, never harm caused by it;
- these figures do not say who is responsible for a war, and must not be read as if they did.
- **within a country, the data cannot say where.** A war's share for an industry is spread over that industry's
  whole national production, by value, wherever it is. Fighting in Amhara in 2023 reaches Ethiopia's coffee, grown in
  the south and west; attacks on civilians in North Kivu and Ituri reach the copper and cobalt mined in Katanga, more
  than a thousand kilometres away; Tigray's war multiplied a cup of coffee's figure for 2021 several hundred times.
  Placing harm by region would need production by region, which no source gives for most of these countries. Each
  product page lists the wars its figure carries, so this is visible rather than hidden.

### 4.1 Arms are placed where they were made

Harm that `S` sends to the arms sectors (ISIC 252 weapons and ammunition, 304 military
fighting vehicles) is not placed on the industry of the country where it was counted, which
did not make the weapons. It is placed where they were made, from SIPRI's arms transfers and
military expenditure databases, by four stated assumptions:

1. **Sides.** Each side of a conflict carries half of the conflict-year's arms harm. The
   model does not know who fired what, and says so by not choosing.
2. **Parties.** A side's half is shared equally among its state parties. A side with no
   state party follows what SIPRI records delivered to armed groups where the harm was
   counted; with no such record it is **unplaced**.
3. **A state's own industry** is the larger of two traces: the arms it exports, over what it
   exports and imports; and, for a state whose exports show it makes arms, the part of its
   military spending its imports do not cover, against a typical state that makes none.
   Spending alone does not make an industry. A state with no trace in the register is
   unplaced, not assumed to make its own.
4. **Suppliers** share the rest by the arms each delivered, into their 252 or 304 by what
   was delivered. A second-hand delivery names who sold it, not who made it: it is in the
   recipient's stock, and its part is unplaced.

A delivery arms a war for years, so deliveries are a **stock with a half-life**, a scenario
parameter (ten years unless a scenario says otherwise). SIPRI's trend-indicator value
(TIV) measures military capability, not price. It is used only as a relative weight among
suppliers, is never converted to money, and no SIPRI series is published. With SIPRI
disabled, arms harm stays where it was counted, as before, and the chain still closes.

### 4.2 Wars nobody scored, and harm filed under no single sector

**An unscored war follows each country's production basket, in `S-MAX` only.** The scored
attribution covers fourteen conflicts. For every other conflict-year, `S-MAX` divides the deaths
counted in each country among that country's production sectors, in proportion to the **value**
of what the country produces (since S32; by mass before): its crops, livestock and forestry
(FAOSTAT), metals and minerals (USGS), and oil, gas and coal (EIA). Each country's deaths follow its own basket. Deaths counted in a
country with no production data stay in `OTHER`. This is a stated assumption with
no evidence behind any one link, so its shares are graded LOW, and the four scenarios that
require evidence do not use it.

**Occupational harm reported under an aggregate is split.** Where a labour inspectorate
reports injuries only as a national total, or under a code covering several sectors ("market
services"), the count is divided among the sectors the code covers. Each sector is weighted by
its employment in that country (ILOSTAT) times its measured injury rate: the injuries per
employed person in that sector, pooled over every country that reports both. Mining kills more
per worker than finance. The shares are LOW, so `S-CONS` keeps them in `OTHER`.

**Three smaller rules on the materials route.** Natural gas, published by volume, is weighed
through the density of methane (0.678 kg/m³, a floor). Harm recorded against a whole section
of agriculture or mining is divided by its divisions' production. And for a year outside
what the publisher covers, a country's own mix in its nearest year is used — never the world's;
inside it, a year with no production is no production — except where the publisher names only
its larger producers. USGS's world tables, which give the metals and minerals for 2010–2025, list
a country when its production is large enough to name and fold the rest into "other countries", so
a country missing from one year's table was not named that year; it keeps its own nearest named
year, graded LOW, rather than taking the world's mix. The world totals that harm is divided by
are the published ones.

**The general rule: in `S-MAX`, no death is left unassigned.** The product owner, 2026-09-30:
*"any unassigned death must be allocated under S-MAX according to the country production
basket."* Wherever the chain cannot place a death, `S-MAX` gives it the production basket of the
country where it was counted — the same basket, by value, as an unscored war's:

- the part of a scored war its evidence left in `OTHER`, and a C2 report nothing splits;
- harm in a sector that produces no material (transport, construction, manufacturing, trade,
  sewerage, services) or has no production series;
- arms whose maker SIPRI's register does not know, and harm on an MRIO cell with no output.

The attribution table keeps its sink: `OTHER` still records what the evidence left open, and the
other four scenarios read it as before. The material split and the intensities spend it. This does
not say what transport or sewerage is made of. It is a stated assumption about where a death nobody
could place belongs, graded LOW, and every run records how much of what reached a material got
there this way (`via_basket`), so the figure and the assumption are published side by side. A
country-year with no production data at all — a state FAOSTAT still files under its predecessor,
such as Serbia before 2006 — takes the country's own nearest year with production. A country with
none in any year keeps `OTHER`.

**Water.** Water supply's production is FAO AQUASTAT's municipal water withdrawal, in m³. Water
supply produces one material, so its harm is water's without any weighing. Water is never part of
a basket: by mass a country's water withdrawal would outweigh everything else it produces, and it
has no price to weigh by value.

## 5. Conservation — nothing is dropped, nothing is renormalised

**This is the strongest constraint in the document.**

> Every scenario reserves a sink sector, `OTHER`. `Σₛ S[c,s] = 1` holds **including the
> sink**, in every scenario, for every subject. A scenario never deletes a share and
> **never renormalises the remainder**. Scenarios differ only in how much lands in the sink
> and whether the sink is propagated into product footprints.

Renormalising is the intuitive implementation and it is wrong in a measurable direction. If
40% of a subject's attribution is evidenced and the rest is discarded, renormalising takes
an oil share from 30% of the total to **75%** — the model would assert most strongly about
oil precisely where its evidence is thinnest. The sink keeps every evidenced share at its
true magnitude and turns the gap into a published number.

```
FORBIDDEN                            REQUIRED
S' = S_evid / Σ S_evid               S[OTHER] += Σ(S_withheld);  Σₛ S = 1
```

Two consequences, both enforced as gates rather than intentions:

- **`OTHER` is displayed, never suppressed.** A product page under a conservative scenario
  reads *propagated X · held in Other Y*. How much harm the model declined to route is
  itself a finding.
- **A run that loses harm aborts.** `Σₛ D[c,s] = D[c]` and `Σₘ D[c,m] = D[c]` to 1e-9, sink
  included. Harm going missing between two stages is an error, not a rounding artefact.
- **Unplaced arms harm is its own term** (§4.1): counted, carried in the identity beside the
  sink, and spread over no supplier. It is never renormalised onto the suppliers the
  register does know.

## 6. Double counting

Three ways a person could be counted twice, each with an explicit control:

1. **Channel overlap** — a miner killed in an armed attack is C1 *and* C2; a child dying of
   cholera in a besieged city is C1-indirect *and* C3. Controlled by an evidenced overlap
   fraction per `(channel, channel, year, country)`, applied before attribution. Rows are
   adjusted, never deleted, so the adjustment stays visible and reversible.
2. **Source overlap** — UCDP and ACLED report the same events. **UCDP is the primary
   series; every other source is validation, monitoring and gap detection, never an
   addend.** Enforced by a unique key on `(dedup_key, metric, year, country, is_primary)`.
3. **Supply-chain recursion** — harm intensities are a satellite account on *direct*
   requirements, and the Leontief inverse handles the infinite regress. Applying `h` to
   total output as well would inflate results by roughly the multiplier.

## 7. Scenarios

| Code | Name | What it routes | Default |
|---|---|---|---|
| **`S-MAX`** | Economic-embedded | Everything, including the sink, which follows each country's production basket (§4.2). | **yes** |
| `S-EVID` | Evidence-weighted | Only evidenced shares; the remainder is held in `OTHER`. | |
| `S-RES` | Resource-only | Energy, minerals, agriculture, land, water. | |
| `S-ARMS` | Arms + resources | `S-RES` plus military-industrial. | |
| `S-CONS` | Conservative | HIGH-confidence links only. | |

`S-MAX` is the default because the project's founding hypothesis is that organised violence
is embedded in economic activity. The framing that makes it defensible:

> The model does not claim to establish the psychological motivation of any combatant. It
> quantifies the economic activity associated with the production and perpetuation of
> conflict, and attributes harm to it. A conflict that began in religious hatred still buys
> rifles, burns fuel, feeds soldiers, seizes land and sells what it digs up.

Scenario parameters live in the database, never in code, and every published figure names
the scenario that produced it.

## 8. Confidence

| Grade | Means |
|---|---|
| **HIGH** | Primary source, direct measurement, ≤2 attribution steps, interval within ±40% |
| **MEDIUM** | Modelled attribution, 3–4 steps, interval within a factor of ~3 |
| **LOW** | Sparse or proxied inputs, >4 steps, interval spanning an order of magnitude |
| **INDICATIVE** | Published as a range only, never as a point estimate |

Grades are **derived** from interval width, attribution depth and source confidence — never
typed by hand.

## 9. What this is not

- **Not a causal claim about any individual death.** Attribution is statistical and applies
  to populations.
- **Not a moral verdict on a purchase.** A high figure describes a supply chain, not the
  person who bought from it.
- **Not comprehensive.** See §2 — every figure is a floor.
- **Not precision.** A figure printed without its range would be a lie of format. Two
  significant figures, always with an interval.
- **Not a boycott list.** The comparison exists to show *why* two products differ, and the
  answer is usually one material and one country.

## 10. Versioning and reproducibility

- A methodology version is **frozen on publication and cannot be edited or deleted** —
  enforced by database trigger, not by convention. Corrections are a new version; the old
  one stays readable and citable forever.
- A revised upstream release produces a **new calculation run**, never a mutation. Results
  for each year coexist rather than replacing one another.
- Every run records its git revision, engine version, data version, scenario, methodology
  version, parameters and a SHA-256 manifest of every input file. A third party with that
  bundle and the repository reproduces the number exactly.

### Publication gates

No figure is published until all of the following hold: every fact row resolves to a
checksummed release · every PAF-requiring row has a PAF · dedup uniqueness holds ·
conservation holds at every stage · the supply-chain matrix is well-conditioned · intensity
denominators match the model's accounting basis · contributions sum to the headline ·
global attributed harm is not exceeded · and sixteen years of backtesting move with events
that actually happened.

**Comparing years means comparing the same reporters.** Work deaths and injuries are counted as
each country reports them and never filled in, and poverty deaths exist for 2019 alone, so a
product's toll can move between two years because a country began or stopped reporting. The
backtest therefore compares every pair of years again without the work-report harm of the
countries whose reporting changed between them, and without poverty deaths where only one year has
them; a move that disappears in that comparison is a reporting change, recorded as one, not a
change in harm. Every published figure names the countries whose work reports it includes.

---

## Change log

| Version | Date | Change |
|---|---|---|
| 1.0 | 2026-10-14 | The first published version: the draft as it stood on 2026-10-14, frozen. No rule changed. |
| 0.1 | 2026-09-08 | First draft. Four harm channels, conservation rule, five scenarios, C5 specified as future work. Not frozen. |
| 0.1 | 2026-09-29 | §4.1: arms harm placed where the weapons were made (SIPRI), with unplaced arms harm as its own conservation term. Still a draft, not frozen. |
| 0.1 | 2026-09-30 | §4.2: unscored wars given their countries' production basket (`S-MAX`); aggregate occupational reports split by employment × sector injury rate; gas by density, sections by divisions, nearest-year production. Still a draft, not frozen. |
| 0.1 | 2026-09-30 | §4.2, the general rule: in `S-MAX` every death the chain cannot place follows the counting country's production basket, and each run records how much did; a country-year with no production takes its own nearest year; water supply's production from FAO AQUASTAT, and water never in a basket. Still a draft, not frozen. |
| 0.1 | 2026-10-14 | §4: the limit of placing by country stated — within a country a war's share falls on that industry's whole production, wherever it is (Amhara and Ethiopia's coffee; North Kivu and Katanga's copper). No rule changed. Still a draft, not frozen. |
| 0.1 | 2026-10-07 | §4: fourteen wars scored (Tigray; the central Sahel); a scored period may cover several UCDP conflicts (Donbas); publication gates: years are compared over the same reporters, and each figure names its reporting countries. Still a draft, not frozen. |
| 0.1 | 2026-10-06 | §4: every industry's harm, and the production basket, divided by value, not weight; no harm is ever placed on a world mix — a share a country's production cannot take goes to `OTHER`, and under `S-MAX` to that country's own basket; tobacco and charcoal join the materials. Still a draft, not frozen. |
| 0.1 | 2026-10-05 | §4: harm stays in the country where the killing takes place, stated and argued (the owner's decision); mining splits its harm among its materials by value (§4); natural diamonds join the materials. Still a draft, not frozen. |
| 0.1 | 2026-10-05 | The site's texts rewritten in plain English; C3 and C4 shown as "Poverty" and "Pollution". No rule changed. Still a draft, not frozen. |
| 0.1 | 2026-10-05 | §4.2: metals and minerals for 2010–2025, from USGS's editions of 2012–2026; a country a named-producer table leaves out of a year keeps its own nearest named year. Still a draft, not frozen. |

## Citation

> Quiescence (2026). *HumanToll Methodology v1.0*. https://quiescence.eu/humantoll/methodology/v1.0/