Method

Questions about the method

Every score on Argupedia follows one fixed, open procedure: Scoring version 1.3 of the Global Debate Evaluation Standard. The essay on the start page explains the framework; this page answers the questions it raises most often and lists, with dates, every change to the method since the move to a global scale. The reference pages with all constants (Baseline, value catalogue, plausibility) are in German; the whole stock is available as data.

Frequently asked

Questions about the procedure

Who judges the judges?

Nobody, and that is the point. Every evaluation of a debate needs values, and every justification stops somewhere at a setting; another reviewer only moves the question one step up (the Münchhausen trilemma). Argupedia does not answer with a higher authority but with full disclosure: every value, every conversion anchor, every source and every calculation is published and applies to every debate alike. Readers can check every number, reject every setting and recompute the debate with their own values. The verdict belongs to the reader, not to the editors.

The Münchhausen trilemma →Value catalogue (German) →Baseline (German) →

Does the same yardstick apply to every debate?

Yes. The same value catalogue, the same conversion anchors and the same evidence rules apply to every debate, from speed limits to pensions. An anchor is never adjusted for a single case; when it changes, it changes centrally, with date and reasoning, and the whole stock is brought up to date. That prevents what easily happens with implicit values: the yardstick drifts unnoticed until the wanted result appears.

Changes to the method →

Can an evaluation be recomputed with my own values?

Yes. Under the argument overview of every German debate a link leads to the Werte-Shift. There the values can be weighted from half to double, individually or as a political preset, and the page shows for every scored debate which verdicts flip under that weighting and which hold. The tool is in German; the English debates are not yet part of it.

Werte-Shift (German) →

What is AI and what is human?

The first evaluation is produced by an AI following exactly this procedure; after that every evaluation passes through review loops of AI and humans. Every page shows at its foot which model reviewed it last, how often AI and how often a human did, and evaluations without any human pass carry a visible mark. The loops work in both directions: every objection that improves a number also improves the rules, the tools and the data on which the next evaluation is built.

As of 31 October 2026: 1,855 logged reviews across 223 records, 1,855 of them with AI and 118 with human sign-off; 74 of the 223 records have at least one human pass. Models involved: Claude Opus 5.5 (589), Claude Fable 5.1 (411), GPT-6 (409), Claude Opus 5 (320), Claude Fable 5 (90), Claude Opus 4.8 (36).

AI with humans in the loop →

Do evaluations change?

Yes, and every change stays visible. When a number changes or an argument is added, a new version is created; older versions remain readable. The review log of every page records when it was reviewed and with what result; the Changes page lists every new version, changed verdict and update of the stock in chronological order. Changes to the procedure itself are listed with dates on this page.

Changes to the evaluations →Changes to the method →

Why does every text sound the same?

Because tone must not be an argument. Every visible text follows the same style rule: short declarative sentences, numbers at the centre, no rhetorical questions, no metaphors from war or sport, no grading of sources. No text favours a side through language, and two debates can be laid side by side.

What distinguishes Argupedia from a fact check or a voting advice tool?

The question. A fact check tests whether a claim is true; a voting advice tool such as the German Wahl-O-Mat shows which party shares a position; Argupedia evaluates what consequences a measure has and how they weigh against each other. Political debates consist mostly of forecasts, which no fact check can test, and of trade-offs, which no measure of proximity answers. Argupedia starts there: with fixed evidence rules for plausibility and published values for weight.

Argupedia, fact checks and voting advice tools →

May the data be used?

Yes, for research and private purposes. The whole stock (areas, measures, evaluations with every argument, value catalogue, baseline, scoring constants) is available as one JSON file under the licence CC BY-NC 4.0: attribution required, no commercial use without agreement.

Data and licence →

Log

Changes to the method

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. This list records every change since the move to the global scale. What changed in an individual evaluation is in its review log at the foot of the page.

  1. Balance 2.2: three questions, three numbers

    The balance of every debate has since answered three questions separately: how likely the measure comes out ahead (share of draws with a positive net effect), how large the effect is (net effect in Normalised Impact) and what share of the effect is positive. Plausibility acts as a fixed discount; the uncertainty about size scatters log-normally around the card value. Verdicts have since read “very likely better” to “very likely worse”; “balanced” is no longer a verdict. Versions and review →

  2. Occurrence and size separated

    Whether an effect occurs and how large it turns out are two questions. If the occurrence follows from the measure itself (definitional), the argument carries P 9.5; the uncertainty about size sits in the impact range, not in plausibility. Plausibility (German) →

  3. Anchors for votes and life satisfaction

    A vote counts by level (federal 1,000, state 800, municipal 650 euro-equivalent per year), a year of life satisfaction 5,700. Plus eight booking rules, among them on public-order goods, job loss, family separation and the relation of money and time. Baseline (German) →

  4. Three missing anchors, no time discounting

    Deprivation of liberty (40,000 per year), lifetime income (1.5 million) and one standard deviation of competence (352,500) are added as anchors. The flat discounting of future effects is dropped: only evidenced persistence shortens the horizon. Baseline (German) →

  5. Housing in three tiers

    A dwelling is two goods: comfort and preference at the market price (benefit and cost cancel) and basic need under social law (3,700 at the end of the moving chain, V 6). Homelessness avoided carries its own anchor (12,700, V 8). Sixteen debates were re-versioned on this basis. Baseline (German) →

  6. Money flows as pairs

    Every financing, every transfer and every interest burden is booked with both legs: payer and recipient, each leg with the weight of its own side. Open financing paths run as a flat public euro. Baseline (German) →

  7. Balance from simulation instead of ratio

    The verdict of a debate has since come from 20,000 simulated futures (share of draws with a positive net effect), no longer from the ratio of the pro to the con total. Twelve debates changed category as a result. Versions and review →

  8. Calculation path at every argument

    The derivation of the impact stands as a chain of numbers under the impact text: quantity, rate, weight, result, every step with its source. The stock was brought up to date debate by debate. Baseline (German) →

  9. Evidence ladder and finding ladder

    A claimed effect without evidence sits well below P 5. The evidence ladder orders what a source shows; the finding ladder, how robust the finding is. Together they set the ceiling of plausibility. Plausibility (German) →

  10. Baseline: eleven anchors with ranges

    The conversion anchors (life year, tonne of CO₂, hour, euro and others) get their own page with derivation, range and change history. CO₂ has since been valued at abatement cost. Baseline (German) →

  11. Causality rule: the identification rung caps P

    A claimed effect must be causally identified, not merely correlated. Correlation carries at most P 5.5, a clean design up to 8.5, definitional relations above that. Plausibility (German) →

  12. Value catalogue as a required field

    Every argument names which value dimension its value weight prices. The catalogue holds the dimensions with class, anchor and relations; the value weight must match the class. Value catalogue (German) →

  13. Scoring v1.3: one global scale

    Impact is carried as Normalised Impact on a global scale, no longer calibrated per debate. Conversion anchors apply to every debate alike; distribution and salience weights are added. Every evaluation is thereby comparable with every other. The framework →

  14. Scale 0 to 10 on a half-step grid

    Value and plausibility sit on 0 to 10 in steps of 0.5. Zero is categorical: an argument with a zero axis is excluded and carries the reason. The framework →

  15. Reasoning per factor

    Every argument reasons value, impact and plausibility in three separate texts with source references. Before, one joint reasoning per argument stood there. The framework →