Climate Political Landscape Index · Methodology

Reading party manifestos by machine

How an automated classifier turns 407,000 lines of party manifestos into structured measures of where parties stand on climate and energy — explained for a first-time reader.

1The raw material

The project has gathered the election manifestos and platform documents of major parties across 22 countries — 367 party-elections in all — and split them into roughly 407,000 individual statements, one short claim or pledge each. A statement might be “We will legislate net zero by 2050” or “Scrap the carbon tax that is punishing working families.”

A large language model reads every statement, one at a time, in its original language — no translation — and fills in a fixed set of fields. Those fields, together, are the “codebook.” The model never invents a final score: it makes many small, reliable judgments per statement, and a transparent formula adds them up afterward.

2Two separate readings — and why

Every statement gets a first reading; the climate- and energy-relevant ones then get a second, richer reading, by a different set of instructions:

The core position reading — where does the party stand, and how firmly? This is the headline. It runs first, on its own, over all 407,000 statements.
The commitment & argument reading — what exactly does the party commit to, what reasons does it give, and whom does it name? A second pass that runs only on the statements the first reading flags as relevant — roughly one in ten.

Why split them? Two measured reasons. When everything was asked in one combined reading, the extra load made the model measurably worse at a core position judgment — so the headline numbers are kept on their own, frozen reading. And the richer reading only earns its cost where there is something to read, so it runs on the relevant statements alone, not the whole corpus.

3Reading 1 — the core position codebook

Filled in for every one of the 407,000 statements.

Field	What it captures	The allowed answers
Relevant?	Is this about climate, energy, or fossil extraction at all? Deliberately over-inclusive — better to over-catch and filter later. This flag is also what selects a statement for the second reading.	yes / no
Channel	Is it about using energy and cutting emissions, or producing it?	demand / supply / both
Object	The main thing named.	fossil / clean / policy / actor / other
Direction	The party’s stance toward the thing it names — not whether it helps the climate. Backing a fossil fuel counts as “for,” even if dressed up as “clean.”	for / against / mixed
Firmness	How binding the commitment is — a rising ladder.	rhetoric → aspiration → pledge → quantified target → target + date → target + mechanism
Backlash	Is the party trying to reverse climate progress? “Denial” is reserved for an explicit claim about the science — either rejecting it or casting doubt on it (“the science isn’t settled”). Objecting to a policy’s cost, or calling it futile, is not denial.	none / rollback / denial
Confidence + quote	A self-rating, plus the exact text kept as a receipt.	high / medium / low + verbatim

4Reading 2 — the commitment & argument codebook

Filled in only for the climate- and energy-relevant statements.

The “what” — the substance of the commitment

Where the first reading records that a party takes a position, the second reading pins down the detail:

Field	What it captures
Sector	Which part of the economy the commitment covers — power, transport, buildings, industry, agriculture & land, waste, or the whole economy.
Energy sub-type	The specific technology named, at fine grain — coal vs. gas, solar vs. nuclear, green vs. blue hydrogen, waste vs. crop biofuel — the cases where “clean” and “fossil” are genuinely contested.
Timing	How soon the party plans to act — near / medium / long / none given.
Distribution	Whether it addresses who pays and who benefits, and whom it names as cost-bearers and beneficiaries.
Adaptation relation	If it discusses adapting to impacts, how that relates to cutting emissions — complement / independent / substitute, the last being a subtle form of delay.

The “why” — 16 appeal cues

For each reason a party invokes, the model records which of these it is:

cost · national competitiveness · jobs · energy & climate security · development & energy access · health · agriculture · moral duty · fairness & justice · freedom · sovereignty · technology · nature · climate outcomes · effectiveness (“will it even work?”) · corruption & capture (plus “other”)

For each cue, it also records how it’s used — the party endorses it, rebuts it, or merely reports it — and, where it applies, whether it’s framed as a loss or a gain.

The “who” — named actors

Every organization or interest the party names — a firm, union, NGO, government body, industry association, and so on — is recorded with which side it’s on (fossil or clean) and the party’s posture toward it: ally, adversary, neutral, or just-mentioned.

5One statement, all the way through

“Labor’s carbon tax is driving up power bills for working families, and we will repeal it.”

Relevant?: yes
Channel: demand (a climate policy)
Object: policy (the carbon tax)
Direction: against (opposes the policy)
Firmness: pledge (“we will repeal”, no date)
Timing: none given
Backlash: rollback (repealing a climate policy)
Distributive: cost-bearer = “working families”
Appeal cue: cost — endorsed, framed as a loss
Named actor: “Labor” (a party) — adversary

Rolled up, this one statement pushes the party’s clean-energy position toward against, adds to its backlash score, and contributes a cost / loss argument to how it makes its case.

6From statements to scores

Small judgments, composed. Each field is a coarse, reliable call — a direction, a rung on a ladder — the kind a model makes consistently. A fixed, transparent formula then combines the many per-statement judgments into each party’s continuous scores (its position, its ambition, and so on), each carrying a margin of uncertainty. The model is never asked to guess a precise final number.
Normative calls live outside the classifier. Judgments like “does nuclear count as clean?” or “does stated denial imply an intent to reverse policy?” sit in a separate, swappable reference table — so they’re explicit and can be changed without re-reading the corpus.
Verbatim, never paraphrased. Exact quotes and target numbers are copied character-for-character, so every score traces back to real text.
Silence is data. A party that says nothing about fossil fuels is recorded as “no stated fossil position” — an honest reading — not scored as neutral or opposed.

7The guardrails on the run

The instructions are version-pinned; every run records exactly which instruction set and which model produced it.
The position reading’s wording is frozen and identical across runs — a single-word paraphrase once caused a measurable drift, so it’s held stable to the character.
Before the full run, the coding is validated — a stronger model and a small hand-checked sample confirm the labels hold up.

8What comes out

From the two readings, the index derives — for each party, then rolled up to each country by seat share — a family of measures.

The core measures

Position	Where the party stands on the energy transition — clean vs. fossil, as two separate scores, never merged.
Ambition	How ambitious its emissions-reduction program is, and whether that ambition targets the sectors that actually pollute.
Backlash	How much intent it carries to reverse the transition — including any outright denial of the science.
Producer ties	Whether it acts as a vehicle for fossil vs. clean producers.
Adaptation	How seriously it treats adapting to unavoidable impacts.

Captured and reported as context context

Timing	Whether the party plans near-term action or delay.
Distribution	Who it says bears the costs and reaps the benefits of the transition.
Declared coalition	Whom the party names as allies and adversaries.
Framing convergence	How similarly rival parties frame the issue.