Method
How we score
Every evaluation on Argupedia follows a fixed, open procedure: currently scoring version 1.3, based on the Global Debate Evaluation Standard (GDES). No gut feeling, no editorial taste: every argument gets a number, and every number has a derivation you can read. This page explains the procedure as far as you need it to read the scores; the full constants are on the three reference pages, Baseline (impact), value catalogue (value) and Plausibility (certainty), and the foundations in the whitepaper. The entire stock is available for download under Data; changes to the procedure are listed with dates, and the most common questions are answered in the FAQ.
The method · one argument, six steps
From sentence to number
Every score on Argupedia is built the same way. Here one real argument from the speed-limit evaluation runs all the way through.
Fewer road deaths and serious injuries
With a 130 km/h limit about 49 fewer people die on motorways each year and about 560 fewer are seriously injured. Without it, today's level persists.
An argument is a testable prediction: with the measure the world goes one way, without it another. What is scored is the difference.
| Deaths avoided | 49 × 35 years | 1,715 | |
| + | Serious injuries avoided | 560 × 3 years | 1,680 |
| + | Minor injuries avoided | 1,070 × 0.15 years | 161 |
| + | Suffering of relatives | (1,715 + 1,680) × 10 % | 339 |
| = | Life-years gained per year | 3,895 life-years |
First the natural quantity: not euros but life-years, every line from a named source.
All effects land on one shared scale via fixed anchors. Life carries neither an income nor a triviality discount.
V follows the good at stake, never the topic: life sits at 10 in every debate, a euro spent at 5.
- Evidence type: studies at most 9
- Identification: experiment at most 8.5
- P 8.5
P is the certainty the evidence allows at most. The ceiling comes from the evidence, not from wishful thinking.
1.56 impact × 10 value × 8.5 plausibility ÷ 10 = 13 of 100
this argument: 13
Effect medium (+1.7) · 66 % of the effect is positive
What counts is not the ratio of the sums but the net effect, pro minus con, drawn 20,000 times: each impact varies around its card value, each argument counts with its plausibility. In 95 out of 100 runs the speed limit comes out ahead.
Three questions for every argument
The six steps above come down to three questions. For us, an argument is always a testable prediction: with the measure the world develops one way, without it another, and what is evaluated is the difference between these two futures. To do so we ask three questions; in the standard the answers are called V, I and P:
1. What is at stake? Value (V, 0–10)
Not every consequence weighs the same. The value register has eight classes: life and health are at 10, the constitutional core at 9, livelihood security and educational opportunity at 8, participation and the environment at 7, functioning systems and prosperity at 6, money at 5, comfort at 4, marginal benefit at 3. The 5 is the fixed point of the scale: every cost argument hangs on it, and a euro spent weighs the same in every debate. What is evaluated is always the concrete effect, never the topic; a cost argument in a health debate remains a cost argument. Within a class, V may deviate from the anchor by at most one point, and only with a reason. The register is global and identical for all debates: no debate can talk up its own importance. Which value dimensions exist and which class each belongs to is listed in full in the value catalogue.
2. How large is the effect? Impact (I)
Effects come in different units: tonnes of CO₂, life years, homes, euros. To make them comparable, we convert them into a common, unitless quantity, the Normalised Impact. It is derived from sources and a disclosed calculation, never from a feeling, and as the quantity that actually occurs on average: a harm that occurs in only one case in ten enters the calculation at a tenth, as a smaller harm, not as a large one with a small probability. Rare events with large harm, an accident, a crisis, therefore count with their expected value: rate of occurrence times harm, per year, with the rate taken from a named source. The scale is open at the top: a large effect may look large, a small one stays small. This is exactly what makes evaluations comparable across completely different topics; the conversion anchors are in the Baseline.
3. How certain is it to happen? Plausibility (P, 0–10)
How well is it supported that the predicted effect actually occurs? P divided by ten reads as the highest probability of occurrence the evidence supports, assessed on the basis of studies, experience and examples from other countries, and independently of how large or important the effect would be. Big claims get no advance of trust. Where the rate of occurrence is already part of the impact, as with the rare accident, P only says how reliable that calculation is; the rarity is not deducted a second time. How far each type of evidence carries is set out in How we calculate plausibility.
The argument score is the product of the three answers: V × I × P ÷ 10, in the example above 10 × 1.56 × 8.5 ÷ 10 = 13. The pro and contra sides are summed separately; the totals are shown side by side as shares (“Pro 72%”). In the standard this is called the debate score.
The balance of a measure answers three questions with three numbers. Whether the measure is better is shown by a simulation: every debate is computed 20,000 times, and each effect varies around its card value as far as its derivation is uncertain; its range covers eight out of ten runs. Plausibility reduces every value by a fixed amount: an argument with plausibility 7 counts at 70 percent. The share of runs in which pro is larger than contra gives the category: from 90 percent “very likely better”, from 70 percent “likely better”, up to 30 percent “likely worse”, up to 10 percent “very likely worse”. In between it reads “unclear”. A measure that is ahead on average is never called “worse”, and one that is behind on average is never called “better”.
How much the measure achieves is shown by the effect: pro minus contra in Normalised Impact, comparable across all debates and graded in steps of ten: below 0.05 very small, below 0.5 small, below 5 medium, below 50 large, above that very large. How much of the effect is positive is shown by the share of pro in pro plus contra. Money that only changes hands (a transfer from tenants to landlords, from the federal government to a state) appears on both sides as a card of its own, but counts in effect and share only with the difference the change of hands makes. A result might read “Very likely better · effect small (+0.2) · 92% of the effect positive”: the direction is certain, the effect small, and there are hardly any downsides.
Where the numbers come from
No number stands without evidence, and every piece of evidence shows its type on the argument. In descending order of strength:
Scientific studies carry furthest, most strongly when several independent studies reach the same result. A single study alone cannot establish high certainty. And the number of studies is only half the question; the other half, whether the evidence shows the effect at all, follows further below.
Direct precedents: the same or a very similar measure has already been implemented elsewhere, and the results are observable. This always comes with an honest note on what is different about the context: a neighbouring country is not a federal state, and 1995 is not today.
Clear mechanisms: traceable causal chains whose steps can be checked one by one, for example “a higher tax rate raises revenue as long as economic output stays stable.” The “as long as” is mandatory: the certainty of a mechanism argument is measured precisely by how robust its condition is, and by which counterforces were examined. Direct legal consequences are the borderline case: if a law cuts a budget item by two billion, the cut itself is certain; only its consequences are uncertain.
Weaker foundations, projections without a disclosed method or mere plausibility considerations, are allowed but capped: they cannot claim high certainty, and the load-bearing arguments of an evaluation must not rest on them alone. More than that: an effect backed by neither a study nor an example nor a clear mechanism counts as rather unlikely until evidence arrives. It lands in the lower half of the plausibility scale, and the argument may state which evidence would raise it.
Evidence may also add up: a single study and an independent practical example pointing in the same direction carry more certainty together than either alone, up to just above the individual limit, never beyond. The limits themselves are modelled on established scales: the likelihood bands of the Intergovernmental Panel on Climate Change (IPCC) and of US intelligence analysis (ICD 203), and the evidence grades of medicine (GRADE). Where we deliberately deviate, for example in how we treat clear mechanisms, we say so and give the reason.
Causality: does the evidence show the effect at all? That two things occur together does not prove that one causes the other. That is why every argument that attributes an effect to the measure states, besides the type of evidence, how the load-bearing evidence identifies that effect: through random assignment, through a named external variation with a control group, through a comparison group without such variation, through a causal chain with a named condition, or not at all, because only co-occurrence was observed. Each of these levels sets its own cap, and the lower of the two applies: three studies observing the same co-occurrence do not get above 5.5; a single clean experiment with a control group may reach 8.5. Four statements belong in the visible text: what the comparison was, which design, which single confounder could produce the association even without the effect, and whether the direction is right. An argument that does not make them stands on the lowest level.
No evidence, no certainty. 5 means: about as likely as not. An argument without evidence can never claim more, however convincing the sentence sounds. How far below that it stands depends on two things: what carries the claim (research, a precedent, a closed causal chain) and whether anything was found against it: a finding pointing the other way, an effect that failed to appear in a predecessor, a disputed link in the chain. How low is decided by the finding, not by a feeling; the chosen value appears in the text as a statement about the world, not as a mood. Mere doubt about the size of an effect, on the other hand, never belongs in plausibility, but in the derivation of the size.
What happens in the conversion
The Normalised Impact is a unitless number, but it does not come out of nowhere: behind it is a conversion that is the same for all debates and whose constants are listed in full in the Baseline. Three of its rules are worth knowing in order to read the numbers.
An effect is not its price tag. Where an effect comes as a flow of money (taxes, transfers, subsidies), the amount is only the carrier, never yet the value. Three questions decide what counts: Whose money? €100 weighs more for a low-income household than for a top earner; the weights for this are fixed (currently, for example, ×2.5 for the lowest income fifth, ×0.4 for the top tenth) and are not reinvented for each debate. What happens in the transaction? Voluntary, informed exchange creates value; pure redistribution only shifts it, and where voluntariness or information is impaired, as with addiction, the payment itself is a harm. Where does it go? Public spending is traced to its final use: “ten billion for health” is not yet a health benefit; what is evaluated is what demonstrably arrives in the end.
Trifles do not add up to catastrophes. Millions of barely noticeable mini-effects are damped before they add up arithmetically to an existential weight. Noticeable and existential effects count in full. The damping factor appears per stream in the argument's derivation where it is recorded in machine-readable form; the Baseline shows how far that already reaches in the stock.
Red lines are not offset. Where a measure only works by inflicting existential harm on selected people without consent, way out or compensation, this is not weighed against benefits but reported as a separate finding next to the score. Everyday statistical risks are expressly not included; they are ordinary contra arguments.
What the procedure deliberately does not do
No blanket verdict. The assessment is always conditional: it says under which weighting of goals a measure is worthwhile, not what you should vote for.
No false precision. A score is a reasoned estimate, not a measurement. The reasoning names what the number depends on and how far apart the sources are; doubt about size is stated there, not hidden in the plausibility. If the evidence cannot separate pro and contra, the system says exactly that, instead of declaring a false winner.
No hidden assumptions. For tonnes, life years and euros to be combined into one number at all, the conversion needs anchors, and every single one of them is publicly documented: the reference value of a healthy life year (€40,000) as well as the income weights. These anchors are tools of calculation, not a statement about what a life is worth. They are changed only centrally, never in an individual case and never with an eye to an ongoing debate.
GDES by hand and on the platform
The platform · from source to page
Where the numbers come from and where they end up
- Sources Studies, official statistics, precedents from other countries.
- Evaluation One file per proposal: arguments, V · I · P, calculation, source list.
- Review AI first pass, human sign-off — every pass is logged.
- Checks Formulas, ceilings and source references are recomputed by machine before anything is published.
- Database Read-only in operation; rebuilt from the files at any time.
- Pages & analyses Proposal pageArea overviewElection comparisonRecently updated
If something changes, a new version is created — the old one stays readable.
The procedure is based on the Global Debate Evaluation Standard (GDES). In its basic form, GDES is deliberately simple: three questions, three values from 0 to 10, calibrated to the debate at hand, usable with pen and paper. This is exactly how we recommend getting started if you want to approach a topic yourself: the basic form structures your own judgement and makes your own assumptions visible, without any register or correction factors.
On Argupedia we evaluate with machine support, and deliberately accept more complexity for it: a global scale instead of calibration per debate, the conversion into Normalised Impact, distribution and salience weights, evidence and causality caps. The reason is a single one: comparability. An evaluation calibrated per debate is coherent in itself, but two debates cannot be laid side by side that way. The platform variant pays with complexity for every evaluation being comparable with every other, across topics, areas and elections. Both forms are the same procedure at two levels: by hand to get closer to a topic, with machine support for the database. The whitepaper describes both, with a guide for workshops and the same example in both forms.
Versions, AI evaluation and review
Every evaluation carries its scoring version, and every anchor of the Baseline keeps its change history. If an anchor changes, this happens centrally and with an entry in that history; an older evaluation stays traceable because its reasoning states which value it was calculated with, and it is brought up to date at the next review. Versions are never mixed silently. Rules added within a version, such as the causality rule and the ladders for the lower half of the scale in summer 2026, apply immediately to every new and every revised evaluation; the stock is brought up to date topic by topic. As long as an argument does not yet carry the new statement, its value is unchecked against the new rule; the reference pages count live how many such arguments there are and how many of them claim more than the rule would allow.
Today an AI produces the initial evaluation using exactly this procedure: it researches sources, derives every quantity and documents every step. Argupedia is currently a proof of concept: the evaluations are continuously revised, including with newer models, and make no claim to finality. Human input is welcome at any time. That is the advantage of scores broken into parts: you can disagree precisely. Not “the evaluation is wrong”, but “the effect of argument 3 is set too high; here is the better source.” It is from exactly such objections that the system is meant to improve over time.
For that reason, every measure and every area shows who last looked at it: which AI model, on which day, and what came out of it. Review runs in rounds: a model looks for errors, gaps and new developments and corrects them, often without any change, sometimes with a new source or a new version. An AI review is a round run by a model alone; AI review, approved by a human means that a member of the editorial team looked through the errors and changes found and approved them, with their name in the log. The counter shows how many rounds and corrections an evaluation has been through and how often a human approved it; where approval is still missing, it says exactly that. More on this under Editorial team and review.
- Reviews
- 2,585
- with AI
- 2,585
- with a human
- 295
- Records with a human pass
- 107 of 261
Live from the review log of all measures, areas and elections, using the latest version of each measure. Models involved: Claude Opus 5.5 (1,305), Claude Fable 5.1 (411), GPT-6 (409), Claude Opus 5 (320), Claude Fable 5 (90), Claude Opus 4.8 (36), Claude Fable 5.2 (14). Most recent review on 18 November 2026.
Records whose current figures no human has yet approved additionally carry the yellow marker AI evaluation · not yet approved by a human at the top of the page. At present this applies mainly to the proposals from the German state elections of 2026: they are fully sourced and built using the same procedure as every federal measure, but a second pair of eyes is still outstanding, and figures and assessments may still change. What counts are the figures since the last re-evaluation: the marker disappears only once a human has reviewed and approved exactly that version. An earlier approval does not cover figures set by a later AI re-evaluation; the marker then reads AI re-evaluation · not yet approved by a human. Text-only passes after an approval leave the approval in place.
Changes to the method and frequently asked questions
The yardstick moves only centrally: a rule or an anchor changes for every debate at once, with date and reasoning, never for a single case. Every such change since the move to the global scale is listed with its date under changes to the method. What changed in individual evaluations, new versions, changed verdicts and updates, is listed chronologically under Changes and on each page in the review log at its foot. Who judges the judges, whether the same yardstick applies everywhere, what is AI and what is human: the frequently asked questions answer these.