The German version is the reference text; this English version is a translation (4 November 2026).
The Global Debate Evaluation Standard
Argupedia · Whitepaper · Version 2.0
A standard for making political arguments visible, testable and comparable. It has two forms of use: simplified in a workshop and technical on the platform.
| Item | Status |
|---|---|
| Whitepaper | Version 2.0, as of 20 October 2026 |
| Technical procedure of the platform | Scoring 1.3 |
| Balance model of the platform | bilanz/2.0 |
| Publisher | Argupedia |
| Author | Julian Immanuel Ohm, M.Sc. |
Change note V2.0: The whitepaper has a new structure. It separates the simplified use (three estimates from 0 to 10 within a session) from the technical use on the platform (shared scales, Normalised Impact open at the top, net balance). The scale definitions of V1.1 now apply only to the simplified form. The chapters on the balance, on the role of Argupedia and on the Lab are new.
Summary
GDES starts with three questions for every argument: what is at stake? How large is the difference between the futures being compared? How reliable is the claim? The answers are called Value (V), Impact (I) and Plausibility (P). They stay visible on each argument separately and can be challenged separately.
For a workshop, simple estimates within the session are enough. The technical use adds documented conversions, shared evaluation scales and rules for uncertainty. Both forms disclose their assumptions. They differ in effort and in how far their comparisons reach.
Argupedia is one application of the standard, not the standard itself. The platform currently serves as a proof of concept: it shows that political proposals can be evaluated in this form throughout, and it relies mainly on AI to do so. Its evaluations are work in progress and make no claim to be final.
01 · The problem and the basic idea
Judgements form from cues of uneven worth
Anyone who evaluates a political argument processes cues. Some carry information about the matter: evidence, logical structure, documented effects. Others say more about the surroundings: the speaker’s confidence, their party, the mood of one’s own group.
Research on judgement describes a gap between the two kinds. How strongly a cue shapes a judgement often does not match what it contributes to the right answer. Cues from the surroundings are easier to access and often weigh more than their content justifies.
Three tendencies contribute to this. Fast judgement builds a coherent story from what is available and rarely notices what is missing. Deliberate reflection is effortful and often looks for reasons for the side the person already leans towards. Working memory is narrow, and disordered information quickly overloads it.
The consequence is hard to notice: a judgement built on weak cues feels from the inside like a well-founded one. Weighing value, effect and credibility happens in every decision, mostly invisibly and therefore hard to correct. The full derivation is in the problem statement.
A decision is a comparison of two futures
A measure is either implemented or not. F1 is the future with the measure. F0 is the realistic course without it, that is, the best estimate of how the situation would otherwise develop. F0 is not today’s state.
An argument claims a difference between F1 and F0 on a named dimension, a delta. “The reform cuts waiting times” means: in F1, waiting times are shorter than in F0. For every delta, GDES determines three things: how much the affected good weighs, how large the difference is and how reliable the claim is.
| Symbol | Meaning |
|---|---|
| F1 | Future with the measure |
| F0 | Realistic comparison scenario without the measure |
| Delta (Δ) | Claimed difference between F1 and F0 on one dimension |
| V | Value: weight of the affected good |
| I | Impact: size of the delta |
| P | Plausibility: reliability of the claim |
The formula is a modelling choice
GDES combines the three quantities in a simple formula:
Argument score = V × I × P ÷ 10
The comparison of two futures explains why differences are considered. It does not determine that effects must be multiplied and then added up. The formula is a reasoned modelling decision: it represents the expected significance of an effect under the chosen assumptions.
The decision has preconditions. Adding up several arguments only makes sense if their effects do not overlap. Some limits cannot be offset against benefits and therefore stand beside the calculation (chapter 05).
02 · GDES simplified in a workshop
Three estimates within one session
In the simplified form, a group assigns V, I and P each from 0 to 10 and notes a reason for every number. The group agrees its scales itself, and they apply only to this session. The result is an ordered overview of how the group weighs the question.
The numbers are not globally comparable quantities and not measured probabilities. Their use lies in the breakdown: disagreement becomes visible at a specific assumption instead of disappearing in the overall judgement.
The following steps are a suggestion for getting started. Duration and format have not yet been tested in practice.
- Set the comparison. The group names the decision, the realistic alternative and the time period.
- Collect consequences. Each card names one concrete effect, the group affected and the reason or evidence. Pro and con are collected separately.
- Agree the scales. For each of the three questions, the group records examples. For I, it decides which large effect the 10 stands for.
- Estimate individually. Each person assigns V, I and P and notes a short reason for each. Open questions are marked instead of forcing a number.
- Compare reasons. The group discusses the differences. Disagreement may remain. Double-counted effects and legal or ethical limits are named.
- Present the weighing. The group can calculate V × I × P ÷ 10 and add up pro and con separately. The totals are an overview for discussion.
- Record the next learning step. The group notes which assumption carries the result and which new information would change the estimate.
An argument card has these fields: consequence, pro or con, comparison and time period, group affected, evidence or assumption, V, I and P with a reason each, open question.
Example: a 130 km/h speed limit on motorways
The numbers in this example are made up and only show the procedure. The group compares a speed limit of 130 km/h with keeping today’s rule over ten years. As the 10 for I, it agrees the largest effect named in this debate.
| Card | Side | V | I | P | Points |
|---|---|---|---|---|---|
| Fewer road deaths and serious injuries | Pro | 10 | 6 | 8 | 48 |
| Less CO₂ from transport | Pro | 7 | 4 | 7 | 20 |
| Longer journey times | Con | 6 | 7 | 9 | 38 |
For journey time, two participants assign V 4 and V 8. The difference lies in the question of whether lost time is a loss of comfort or lost lifetime. This question is the real point of dispute in the session, and the card makes it nameable.
03 · Applying GDES technically
Shared scales replace the session’s agreement
The technical form answers the same three questions. It replaces the session’s scales with rules that are the same for all debates. The purpose is comparability across topics. The price is a much higher effort.
| Question | Simplified use | Technical use |
|---|---|---|
| Comparison | Set the decision, alternative and time period together | Document comparison scenarios, scope, time period and groups affected |
| V | Estimate and justify the importance of the good | Placement in a shared value register, 0 to 10 |
| I | Estimate from 0 to 10 with agreed examples | Derive the natural quantity and convert it into Normalised Impact, open at the top |
| P | Estimate reliability with reasons | Check the kind of evidence and causal attribution, 0 to 10 with caps |
| Result | Overview of how the group weighs the question | Expected contributions and a net balance with uncertainty |
| Reach | Within the session | Through shared scales, under suitable conditions for comparison |
In the technical form, V and P stay limited to 0 to 10. I has no upper limit, and so the argument score has none either.
From claim to number
The derivation follows a fixed order:
- Determine the natural quantity. The effect is derived in its own unit: tonnes, years of life, hours, euros. Time period and frequency are part of it.
- Convert. Fixed, published conversion rules bring the quantity onto the shared scale. The result is the Normalised Impact, a unitless number.
- Place V. The value register states which class the affected good belongs to.
- Justify P. The check asks how reliable the claimed connection is.
- Document the range. Every effect size carries the range within which, according to the sources, it may lie.
Two kinds of uncertainty stay separate. Doubt about the size of an effect belongs in its range. Doubt about the connection itself belongs in P. This way no uncertainty is deducted twice.
Rare harms count in the quantity with their rate of occurrence. A harm that occurs in one year in ten enters I at one tenth. P then only describes how reliable this calculation is.
For P, the technical form checks two things separately: the kind of evidence and the causal attribution. Several studies that observe the same co-occurrence do not yet show that the measure causes the effect. The limits for each kind of evidence are on the page Plausibility.
Normative settings do not sit only in V
V makes the weight of the affected good visible. The conversion also contains normative settings, that is, assumptions the analysis makes rather than derives: how much a year of life counts against a tonne of CO₂, how much more the same amount weighs on a small income than on a large one, and from what point many barely noticeable effects are damped.
A shared scale is therefore a convention and not value-neutral. The technical form distinguishes settings from empirically derived quantities and publishes both. The settings are in the Baseline and in the value catalogue. They are changed centrally and never for a single debate.
The term Baseline thus has a fixed meaning: it refers to the shared conversion rules. The comparison scenario of a single evaluation is called F0.
Example: the same card in the technical form
The numbers come from the Argupedia evaluation of the 130 km/h speed limit, version 7, as of 23 September 2026. Later versions may differ.
The card “Fewer road deaths and serious injuries” starts with the quantity. Avoided deaths, serious and minor injuries add up to around 3,900 years of life gained per year. The range of the sources runs from about half to 1.6 times this value.
The conversion leads to a Normalised Impact of 1.56 on the scale of this page. V is 10, because life and health are the highest class of the register. P is 8.5, because a field trial with comparison stretches of road supports the effect. The card thus counts 13 points.
The 6 from the workshop and the 1.56 on the platform measure the same effect on different scales. The first applies within the session. The second can be compared with any other evaluation on the platform once the scale of the respective page is taken into account.
04 · From arguments to the balance
Three levels answer three questions
The argument score says how much a single argument weighs. The totals of the pro and the con side show how much weight lies on each side. The balance says whether the measure comes out ahead overall, how much it achieves and how much of its effect is positive.
Totals and balance answer different questions. A strong pro side does not yet say how reliable a positive net effect is. The technical form therefore derives its category from the net effect: pro minus con.
The balance includes the uncertainty
The platform runs every debate 20,000 times. In each run, every effect varies around its card value as far as its derivation is uncertain; the range of an argument covers eight in ten runs. Plausibility acts as a fixed discount: an argument with P 7 counts at 70 percent in every run. On average, the calculation therefore gives exactly the sum of the cards. The balance states three numbers:
- Probability: the share of runs in which pro is larger than con. This determines the category, from “very likely better” (from 90 percent) through “likely better” (from 70 percent) and “unclear” to “very likely worse” (up to 10 percent).
- Effect: pro minus con in Normalised Impact, comparable across all debates and graded in steps of ten, from “very small” (below 0.05) to “very large” (above 50).
- Share positive: pro divided by pro plus con.
A measure that comes out ahead on average is never called “worse”. “Unclear” means that the evidence does not separate pro and con, not that both sides are equally large. The probability says only how certain the direction is, nothing about the size: a small measure with almost no downsides can be “very likely better” and still achieve little.
In the example, the 130 km/h speed limit comes out ahead in 95 percent of runs and is therefore “very likely better”. Its effect is medium (+1.7), and 66 percent of its effect is positive.
The share applies under the assumptions of the model. It is not measured proof that a political decision is right. Money that only changes hands varies on both sides together and counts in effect and share only with the difference the transfer makes. The rules are under How we evaluate.
A shared scale is not yet a comparison
The shared scale creates the precondition for setting two proposals side by side. A fair comparison requires more: the same starting situation, the same time period and the same value scale. Two proposals can each come out ahead of their own starting situation without that settling their ranking.
Overlapping measures cannot be added together. If two proposals act on the same quantity, their combined effect is smaller than the sum of the individual ones. The platform shows numbers for each page on its own scale. A comparison between pages requires that this scale is taken into account.
05 · Limits and quality assurance
Three states of an argument
An argument is evaluated, excluded with reasons, or included and still open. An effect is excluded if it has been refuted or is already counted elsewhere. An argument stays open if a named piece of information needed for the evaluation is missing. A missing evaluation is not an established zero effect.
What the standard does not do
- Evaluations remain judgements. V, I and P are reasoned estimates. Two careful people can evaluate differently on the same evidence. The standard moves the dispute to a visible place.
- F0 is itself a forecast. A wrongly set comparison scenario distorts every delta. Disclosure makes this assumption open to negotiation and does not remove the problem.
- Numbers look more precise than they are. Ranges and expandable derivations counter this. The risk remains.
- Missing arguments stay invisible. A balance applies to the arguments included. An overlooked argument changes it.
- Not every limit can be offset. A high V does not adequately represent legal and ethical constraints. They stand as a separate finding beside the balance.
- Not everything political is a delta. The standard captures questions of identity, dignity or procedural fairness only in part.
- The application decides. Applied badly, the standard gives existing biases the appearance of objectivity. Only disclosure and the ability to challenge help against this.
Review and revision
Every evaluation on the platform carries a review log: who reviewed it, when and with what result. If a number or the set of arguments changes, a new version is created, and the old one is kept. Changes to the text or to the derivation are made to the existing version and are logged.
Every evaluation comes with the revision question: which new information would change the estimate? A testable presentation invites debate. It does not guarantee that the debate takes place.
06 · Argupedia as a proof of concept
The platform is an application, not the standard
GDES can be used independently of Argupedia. Argupedia applies the technical form to political proposals and publishes every evaluation with its derivation. The current aim is a proof of concept: to show that such evaluations can be produced throughout and what can be learned from them.
Five properties of the application
Every evaluation of a debate rests on values, and every justification ends somewhere at a setting; this is the Münchhausen trilemma (Albert 1968), and no method escapes it. One evaluation therefore differs from another not in whether it makes settings, but in how visible and how stable those settings are. Argupedia commits to five properties:
- One standard for every debate. Every debate is evaluated by the same procedure; every argument receives a weight from value, impact and plausibility, with reasons and sources.
- Values in the open instead of hidden. Which value counts how much, and with which anchor tonnes, years of life and hours are brought onto one scale, is set out in the value catalogue and in the Baseline, with derivation, range and change history.
- A scale that does not shift. The same values and anchors apply to all debates. If an anchor changes, it changes centrally, with a date and for the whole stock, never for a single case (changes to the method).
- Your own values. Anyone who weights differently can take the published values and anchors, put in their own and recompute a debate to see which verdicts flip under a different weighting and which hold.
- Evaluations that learn. Every evaluation goes through review loops of AI and humans, carries versions and a review log; the whole stock is available for download under Data (CC BY-NC 4.0).
This application does not answer the question of who evaluates the evaluators with a higher authority, which would itself have to weigh with values, but with disclosure: everything is explicit, nothing implicit, and in the end the reader makes the judgement. The word for this is not objectivity but consistency (background).
The evaluations are produced mainly with AI
An AI produces the first evaluation: it researches sources, derives every quantity and documents the steps. Automated check rules control the calculation and the evidence. The reason for working this way is the effort: a proof of concept across many topics could not currently be achieved with human work alone.
The evaluations are revised continuously, including with newer models each time. Every page names the reviewing model and the date. Human review is added where it takes place and is shown separately. Pages without human review are marked as such.
The evaluations are meant to be useful as contributions to a debate. They are not mature in every respect and make no claim to be final. Objections to individual assumptions are welcome, and every correction can be traced.
Participation is possible and currently not the focus
The platform contains functions for comments and for users’ own estimates. Building it into a community that carries the work is currently not the focus. Human feedback is nonetheless taken in continuously and considered in the next review round.
AI supports judgement and does not replace it
Research on automation describes a known risk: whoever only monitors gradually loses the ability to intervene. Something similar applies to AI in political discourse, because it produces fluent arguments in large numbers. Breaking an argument down into V, I and P keeps every assumption testable for people. The decision stays with people and their institutions.
07 · Outlook on the Lab
In the Lab (German), Argupedia tests which further tools can grow out of openly evaluated arguments. The prototypes show directions of development. Their methods and results have not yet been checked sufficiently and are not part of the standard.
| Direction of development | Prototypes | Possible use |
|---|---|---|
| Examining debates | Debate vs. analysis (German) | Show which load-bearing reasons appear in a debate, are backed by evidence or are missing |
| Exploring robustness | Simulating debates (German) | Show under which assumptions and weightings a balance stays stable and where it flips |
| Comparing alternatives | Evaluating the optimum (German), Comparing proposals (German) | Find more favourable variants within the shared framework and show which effects carry the difference |
If individual tools prove themselves, they can gradually become part of the regular platform. A gap between a debate and the analysis does not by itself show that the debate is poor. An optimum applies only to the variants and assumptions examined.
08 · Sources and version history
Research, derivation and open questions
The table separates what a study shows from the design decision GDES derives from it. Whether GDES itself improves judgements has not yet been studied. This effect is an open research question.
| Source | Finding | Derivation for GDES |
|---|---|---|
| Brunswik (1952); Hammond et al. (1975) | The weight of a cue in a judgement differs from its contribution to the right answer | Make reasons visible one by one |
| Kahneman (2003, 2011) | Fast judgement builds a coherent story from what is available and passes over what is missing | List missing arguments and open questions explicitly |
| Evans & Stanovich (2013) | Automatic and controlled processing can both help and both err | Do not leave the check to the first impression |
| Petty & Cacioppo (1986) | Messages are processed through content or through surface features | Separate content and evidence from the sender |
| Sweller (1988) | Working memory is narrow, and disordered information overloads it | One effect per card, fixed structure |
| Hart et al. (2009); Taber & Lodge (2006) | On average, people prefer information that fits their attitude | Build pro and con by the same rule |
| Wilson & Brekke (1994) | Unwanted influences on a judgement are hard to recognise from the inside | Record assumptions outside one’s head |
| Lu, Yuan & McLeod (2012); Lorenz et al. (2011) | Groups pass over unshared information, and social influence narrows estimates | Estimate individually first, then compare |
| Mellers et al. (2014) | Training, teamwork and feedback improve forecasts | Record estimates and check them later |
| Fishkin et al. (2021) | Structured deliberation with balanced information changes judgements | A shared, balanced basis |
| Bainbridge (1983); Endsley & Kiris (1995) | Merely monitoring automated systems weakens the ability to intervene | People check individual assumptions themselves |
References
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
- Brunswik, E. (1952). The Conceptual Framework of Psychology. University of Chicago Press.
- Endsley, M. R. & Kiris, E. O. (1995). The out-of-the-loop performance problem and level of control in automation. Human Factors, 37(2), 381–394. https://doi.org/10.1518/001872095779064555
- Evans, J. St. B. T. & Stanovich, K. E. (2013). Dual-process theories of higher cognition: Advancing the debate. Perspectives on Psychological Science, 8(3), 223–241. https://doi.org/10.1177/1745691612460685
- Fishkin, J., Siu, A., Diamond, L. & Bradburn, N. (2021). Is deliberation an antidote to extreme partisan polarization? Reflections on “America in One Room”. American Political Science Review, 115(4), 1464–1481. https://doi.org/10.1017/S0003055421000642
- Hammond, K. R., Stewart, T. R., Brehmer, B. & Steinmann, D. O. (1975). Social judgment theory. In M. F. Kaplan & S. Schwartz (Eds.), Human Judgment and Decision Processes (pp. 271–312). Academic Press. https://doi.org/10.1016/B978-0-12-397250-7.50016-7
- Hart, W., Albarracín, D., Eagly, A. H., Brechan, I., Lindberg, M. J. & Merrill, L. (2009). Feeling validated versus being correct: A meta-analysis of selective exposure to information. Psychological Bulletin, 135(4), 555–588. https://doi.org/10.1037/a0015701
- Kahneman, D. (2003). A perspective on judgment and choice: Mapping bounded rationality. American Psychologist, 58(9), 697–720. https://doi.org/10.1037/0003-066X.58.9.697
- Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
- Lorenz, J., Rauhut, H., Schweitzer, F. & Helbing, D. (2011). How social influence can undermine the wisdom of crowd effect. PNAS, 108(22), 9020–9025. https://doi.org/10.1073/pnas.1008636108
- Lu, L., Yuan, Y. C. & McLeod, P. L. (2012). Twenty-five years of hidden profiles in group decision making: A meta-analysis. Personality and Social Psychology Review, 16(1), 54–75. https://doi.org/10.1177/1088868311417243
- Mellers, B., Ungar, L., Baron, J. et al. (2014). Psychological strategies for winning a geopolitical forecasting tournament. Psychological Science, 25(5), 1106–1115. https://doi.org/10.1177/0956797614524255
- Petty, R. E. & Cacioppo, J. T. (1986). The elaboration likelihood model of persuasion. Advances in Experimental Social Psychology, 19, 123–205. https://doi.org/10.1016/S0065-2601(08)60214-2
- Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285. https://doi.org/10.1207/s15516709cog1202_4
- Taber, C. S. & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs. American Journal of Political Science, 50(3), 755–769. https://doi.org/10.1111/j.1540-5907.2006.00214.x
- Wilson, T. D. & Brekke, N. (1994). Mental contamination and mental correction: Unwanted influences on judgments and evaluations. Psychological Bulletin, 116(1), 117–142. https://doi.org/10.1037/0033-2909.116.1.117
Method of the platform: How we evaluate, Baseline, value catalogue, Plausibility.
Version history
- V2.0, addendum (20 October 2026). Section “Five properties of the application” added to chapter 06: standard, open values, constant scale, own values, evaluations that learn, with the Münchhausen trilemma as the justification.
- V2.0 (October 2026). New structure. Simplified and technical use separated. Balance as a chapter of its own. Normative settings outside V named. Role of Argupedia as a proof of concept and outlook on the Lab added. Sources with full details.
- V1.1 (3 July 2026). End points of the scale defined: 0 as exclusion, 10 as the maximum of the respective anchor. These scale definitions describe the simplified form.
- V1.0 (June 2026). First version, archived as a PDF (German).
The standard is designed as a living document. Feedback, objections and counterexamples are welcome.