How plausibility is calculated

An argument can claim a huge effect and still weigh almost nothing if nobody knows whether the effect will occur. Plausibility P is where that is accounted for, and it is not set freely: the kind of evidence available decides how high P may rise at all. This page writes out these caps, together with the rules that lead beyond them.

scoring/1.3 · as of 1 November 2026 · scale 0–10 in steps of 0.5 · hard cap 9.5

P is the probability an argument may claim

P / 10 reads as the highest probability that may be claimed for the effect. P 8.5 therefore does not mean “fairly sure” as a feeling. It means the evidence supports a statement at the level of very likely, and not a jot more.

0 is categorical

P = 0 is not a very small plausibility but the exclusion: the argument drops out of the scoring, its score is 0, and the reason is stored with it. Two roads lead there. Either it has been shown that the effect does not occur: a counter-finding, an effect that did not materialise, a refuted causal claim. Or, after a declared search, nothing carries the claim and it is unlikely in itself. The smallest number that still counts is 0.5.

10 is reserved

P 10 exists only for arithmetic identities and settled science. Anything that people first have to do cannot reach 10; behavioural responses never get it.

The hard cap is 9.5

Even the best body of studies plus precedent plus mechanism does not get past 9.5. The gap to 10 is the room left for error.

How far each kind of evidence carries

Every argument names its evidence basis, and the basis sets the cap. This is the real rigour of the procedure: an argument without evidence may not claim to be certain. It gets 5 at most, however convincing the sentence sounds. Where nothing carries the claim, it sits below that; how far below is set by the lower half of the scale.

Mechanism

open

A causal relationship that holds by itself: budget identities, legal consequences, arithmetic. There is no fixed cap. P assesses the stability condition, not the state of the research.

mechanismus · cap from evidence.mechanismus

Converging studies

≤ 9.0

Several independent studies with different data and methods point in the same direction.

konvergierende_studien · cap from evidence.p_ceilings.studie_konvergierend

Study

≤ 9.0

Generic token in the stock without a finer grade. It receives the highest value of its family. A study declared as a single study sits at 7.

studie · cap from evidence.p_ceilings.studie_konvergierend

Multiple precedents

≤ 8.5

The measure has run several times elsewhere, and the cases are comparable enough to be transferred.

mehrere_praezedenzfaelle · cap from evidence.p_ceilings.praezedenz_mehrfach

Precedent

≤ 8.5

Generic precedent token without a finer grade. It receives the highest value of its family.

praezedenz · cap from evidence.p_ceilings.praezedenz_mehrfach

Single precedent

≤ 7.0

A single case elsewhere. It also carries the transfer deduction when the argument decides the ranking.

einzelner_praezedenzfall · cap from evidence.p_ceilings.praezedenz_einzeln

Projection

≤ 6.0

A model calculation without a track record: scenarios, forecasts, extrapolations.

projektion · cap from evidence.p_ceilings.projektion

Plausibility

≤ 5.0

A reasoned expectation without evidence. It cannot claim more than “about as likely as not”.

plausibilitaet · cap from evidence.p_ceilings.plausibilitaet

Whether the evidence shows the effect at all

The evidence basis counts evidence: where it comes from and how much of it there is. Whether a piece of evidence shows the claimed effect at all is a different question, and the amount of evidence does not answer it. Three studies that observe the same co-occurrence remain three observations of a co-occurrence; none of them shows that one thing causes the other. On the caps of the basis alone they could claim P 9.0. That is more than a single clean trial with a control group, which ends at P 7.0.

The identification level ends this inversion. Every argument that claims the measure causes something names, next to the basis, the level of the evidence that carries exactly this claim. It sets a second cap, and the lower of the two applies:

p_ceiling = min(basis cap, tier cap, identification cap)

Definitional

no cap

No causal claim at all: the legal consequence of the adopted text, an accounting identity, a cost item, a purely moral statement, or a relationship fixed by physics or by contract. There is nothing to identify here; the directness rule further down applies.

definitorisch · cap from evidence.identifikation.p_ceilings

Experimental

≤ 8.5

Chance decides who is treated: a controlled study with random assignment, a field trial with a control arm, comparison units drawn by lot.

experimentell · cap from evidence.identifikation.p_ceilings

Quasi-experimental

≤ 8.0

A named external variation carries the evidence: difference-in-differences with a pre-trend shown, regression discontinuity, an instrumental variable with the instrument named, synthetic control, an event study with a lead period.

quasi_experimentell · cap from evidence.identifikation.p_ceilings

Controlled

≤ 7.0

There is a comparison group but no external variation: a matched cohort, a controlled time series, before and after with a genuine control region.

kontrolliert · cap from evidence.identifikation.p_ceilings

Mechanistic

≤ 6.0

A behavioural chain without a measured key parameter, but with a named stability condition and checked counter-mechanisms. The cap sits below the “likely” threshold: an unmeasured response cannot claim “likely”.

mechanistisch · cap from evidence.identifikation.p_ceilings

Associational

≤ 5.5

No comparison case: a cross-sectional comparison, two time series moving together, an uncontrolled before and after. A finding, not proof of an effect, and deliberately only half a step above a mere analogy.

assoziativ · cap from evidence.identifikation.p_ceilings

Undetermined

≤ 5.0

The design was not checked or cannot be named from the source. The missing statement is the lowest rung, never the highest value of the family.

unbestimmt · cap from evidence.identifikation.p_ceilings

What every causal claim states

Four statements, made in the visible plausibility text of the argument: not only in an internal field, but where anyone can read them. An argument that does not make them sits on the lowest rung; the missing statement is not read generously.

  1. The comparison. What it was measured against: a control group, a control region, a pre-trend, random assignment, or explicitly that there is none.
  2. The design in one word from the ladder. “Methodologically sound” or “published methodology” are not statements of design and fall back to undetermined.
  3. The one thing that could distort the result. The factor most likely to produce the observed relationship even without the effect, and how the design deals with it, or that it does not.
  4. The direction check. Could the outcome have caused the measure rather than the other way round, or was there selection into the treatment? Half a sentence is enough, but it has to be there.

The level measures attribution and nothing else. Whether a finding is disputed is decided by the state of research further down; whether it comes from another country, by the transfer deduction. An undisputed consensus about an association that was never identified therefore remains robust and associational at the same time. The lower cap applies; there is no double penalty.

How far the stock carries it already

Identification level · capArgumentsAvg P at the level’s cap
Definitional · no ladder cap 75 / 31 measures 7.66 0 · 0%
Experimental · ≤ 8.5 7 / 6 measures 5.86 0 · 0%
Quasi-experimental · ≤ 8.0 72 / 27 measures 5.28 0 · 0%
Controlled · ≤ 7.0 25 / 13 measures 5.04 1 · 4%
Mechanistic · ≤ 6.0 188 / 47 measures 4.77 40 · 21%
Associational · ≤ 5.5 20 / 11 measures 3.80 2 · 10%
Undetermined · ≤ 5.0 6 / 5 measures 4.00 4 · 67%

393 of 417 arguments carry the statement. All others count as undetermined until they are revised: without a checked design, no more than 5.0 can be claimed, and 0 of them sit above that today. The last column counts the arguments whose P sits on the cap of their level: there, attribution binds, not the evidence.

In the balance, P stays what it is on the card: a discount. The balance runs every debate 20,000 times; an argument with P 7 counts at 70 percent in every run. From P 4 downwards, an argument only occurs in as many runs as P says, at P 3 in three of ten, and then at full weight. On average this is the same, but a verdict that hangs on a weakly supported argument becomes less certain. On top of that, the size of the effect varies around its card value as far as its range reaches. That is why doubt about whether an effect comes at all belongs in P and not in a range that starts at zero. The same doubt is then counted only once: either as “whether” or as “how much”, never twice.

Where P lands when nothing carries the claim

The caps say how high P may rise. Where P lands is decided by something else: what carries the claim. There are three carriers. Research: a study that measured the effect. Precedent: the measure ran elsewhere, and the outcome is observable and at least partly transferable. And a closed causal chain: a mechanism that leads there practically by itself. If one of them carries the claim, P may go up to the cap. If none does, the argument sits below 5.0. This is not because something speaks against it, but because nothing speaks for it.

5.0 is therefore not a zero point but the upper bound of an unsupported claim. “About as likely as not” is the best a claim without a study, without an example and without a closed chain can reach, and it reaches it only with a closed chain. Downwards the ladder is open to zero. A claim for which a declared search finds nothing, and which is unlikely in itself, sits at 1.0 or 0.5, or drops out of the scoring. The same applies, and more heavily, to counter-evidence: a claim that meets findings showing the effect does not or did not occur sits there as well. That is the second check, further down. Behind this stands a setting, and it is stated openly: a causal claim that nothing carries is more likely wrong than right. This is the record of evaluation research rather than scepticism. Peter Rossi’s “iron law” of 1987 states that the expected net impact of a large social programme is zero. A claim of more has to be shown.

The ladder does not cap, it places: every band has a lower and an upper bound, and it never lifts P above the cap. The smaller of the two values applies. This is no side issue: 33 percent of the evaluated arguments sit below 5.0.

P = min(band value, p_ceiling)

Nothing carries the claim

Unevidenced and unlikely

P 0.0 – 1.0

The search is declared and found nothing (no study, no example, no analogy), and the claim is also unlikely in itself: the chain would have to run through links for which there is no reason. 1.0 if a remainder stays conceivable; 0.5 if only the sentence itself is left; 0 if nothing, not even an analogy, speaks for the claim. Then the argument drops out of the scoring, with score 0 and the reason stored. It is the same exclusion as for a refuted causal claim.

≈ 0–10% probability of occurrence · unbelegt_unwahrscheinlich · from evidence.befundleiter.baender

No carrier, but conceivable

P 1.5 – 2.5

Neither study nor precedent nor closed causal chain, but there could be something behind it: the direction is conceivable and a mechanism imaginable, it just cannot be shown yet. The upper value applies if the direction at least follows from a partly transferable analogy (another field, another instrument). The lower value applies if, on top of that, an obvious counter-mechanism is unchecked. Whoever can no longer see even the conceivable belongs one band lower.

≈ 15–25% probability of occurrence · ohne_traeger · from evidence.befundleiter.baender

Chain open

P 3.0 – 3.5

The causal chain is named, but at least one link that carries the quantity is unchecked or a counter-mechanism is unanswered, and no evidence and no precedent supports it. The upper value with one open link, the lower with several.

≈ 30–35% probability of occurrence · kette_offen · from evidence.befundleiter.baender

Chain closed, unevidenced

P 4.0 – 5.0

The chain is closed: every link named, the condition under which it holds stated, counter-mechanisms checked and ruled out. Only the evidence is missing. The upper value, 5.0, requires an additional weak support, such as expert consensus or a partly transferable analogy; otherwise 4.0 to 4.5. A claim without a study and without an example gets no higher.

≈ 40–50% probability of occurrence · kette_geschlossen · from evidence.befundleiter.baender

Something speaks against it

The second check weighs more heavily than the first. It is not only that nobody has shown the effect, but that someone has shown it does not or did not occur. A claim that meets such a finding sits lower, down to zero, and the finding decides how much lower, not a feeling. If both apply, nothing carries the claim and something speaks against it, the lower band binds, and only that one is stored.

Counter-finding

P 0.0 – 1.0

It has been shown that the effect does not occur: the case closest in context shows a finding in the opposite direction, or it can be verified that the chain cannot start at all (missing competence, conflicting law, an expired deadline, a physical state). 1.0 for one such finding; 0.5 if several independent counter-findings exist and no supporting finding of the same proximity; 0 if the counter-findings hit the causal claim itself. Then the argument drops out of the scoring.

≈ 0–10% probability of occurrence · gegenbefund · from evidence.befundleiter.baender

Effect did not materialise

P 1.5 – 2.0

The measure or its close twin has already run, in the same jurisdiction or with the same target group, and the claimed effect did not materialise. The difference from the band below: a mere absence, not a positive finding against it. The lower value applies with two null findings or when the implementing body itself contradicts the claim.

≈ 15–20% probability of occurrence · ausgebliebener_effekt · from evidence.befundleiter.baender

Partial aspect supported

P 2.5 – 3.0

A supporting and a counter-finding of the same proximity both exist, and the supporting finding demonstrably carries only a named part: a subgroup, one link of the chain, one of several channels. If this part carries the quantity, the upper value applies; if the counter-finding sits on the higher identification level, the lower one.

≈ 25–30% probability of occurrence · teilaspekt · from evidence.befundleiter.baender

Load-bearing link disputed

P 3.5 – 4.0

The direction is supported, but precisely the link that carries most of the assumed quantity has a finding against it that it does not rule out. If the counter-finding is less well identified, the upper value applies; if it is equally strong or a second one is added, the lower.

≈ 35–40% probability of occurrence · tragendes_glied · from evidence.befundleiter.baender

Coverage gap

P 4.5

A named link that carries the quantity is uncovered, with neither supporting nor counter-finding, while the rest of the claim is carried. The gap is named in the visible plausibility text.

≈ 45% probability of occurrence · deckungsluecke · from evidence.befundleiter.baender

Refuted

P 0.0

Not the quantity is disputed but the causal claim itself, and it does not hold. Then the argument drops out of the scoring: score 0, reason stored, out of every total. Two bands lead to the same zero: “Counter-finding”, when it has been shown that the effect does not occur, and “Unevidenced and unlikely”, when nothing carries it.

no band, edge of the ladder · from evidence.befundleiter.baender

Four questions before a number falls below 5.0

  1. The axis question. Is it in doubt whether the effect occurs, or only how large it is? Pure doubt about the size of a quantity that the analysis derived itself is not a case for the ladder: it belongs in the derivation of the size, and P stays high. A figure from someone else is different, such as a party’s promise or a quoted demand. There P may correct the claimed size, and the argument says that it does.
  2. The expected-value question. Is it about a rare event with large damage: an accident, a crisis, a disaster? Then the probability belongs in the size of the argument, not in its certainty. The size is the expected value: rate of occurrence times damage per event, per year, with the rate from a named source and its uncertainty in the derivation. P then says what it says everywhere else: how reliable this calculation is, how well the rate is supported, how well the damage is quantified, whether the transfer to the case at hand holds. The reason for the separation matters. The caps on this page measure evidence. Applied to an accident frequency, they would cap a physical rate at 50 percent, and a 0.5 grid that ends at 5 percent cannot represent a frequency of one in ten thousand per year anyway; the size axis carries it easily. Both at once is not allowed: if the rate is in the size, P does not fall again because the event is rare. And whoever wants to weight disasters more heavily than their expected value does so through the value or a written-out premium, never through a P held high. Risk aversion is a question of values, not of evidence.
  3. The carrier question. Which of the three carriers stands behind the claim: a study, a precedent, a closed chain? None: then the evidence ladder places it as no carrier, chain open or chain closed. At least one: then P may go up to the cap, unless the next question lowers it. A precedent only counts at a comparable dose: an eighth of the amount in a neighbouring country is no example of the full measure.
  4. The finding question. Is there a finding that speaks against the effect: a source that measured, observed or legally established an opposite or absent result? Then the finding ladder places it: counter-finding, effect did not materialise, partial aspect, load-bearing link, coverage gap. A counter-finding only counts if it measures the same thing as the claim.

Four locks on top of that prevent the same fact from costing twice. The chosen number appears once as a statement about the world in the visible plausibility text (“this happens in roughly one in three comparable cases”, “nothing carries this beyond the chain”); “weak” and “rather low” are moods, not claims. Thin evidence counts once: the cap of 5.0 for the basis plausibility and the evidence band answer the same question, and the band is the placement below it, not a second deduction. If the counter-finding already stands as a separate argument of the other side on the same sheet, it is counted there and does not additionally lower this side’s band. And the ladder is applied per measure, never only to the inconvenient side.

What can raise the number again

A low placement is a state, not a verdict. An argument may carry which proof would raise it and how far: an ongoing evaluation, a precedent that is about to report, a key parameter that could be measured. This changes nothing about today’s score. It says where the number might move, and it is shown on the argument card so that nobody takes the low number as final.

How far the stock carries it already

Band · rangeArgumentsAvg P Share below 5.0
No carrier, but conceivable · P 1.5 – 2.5 3 / 3 measures 2.00 2%
Chain open · P 3.0 – 3.5 54 / 27 measures 3.33 39%
Chain closed, unevidenced · P 4.0 – 5.0 47 / 28 measures 4.41 34%
Counter-finding · P 0.0 – 1.0 2 / 2 measures 1.00 1%
Effect did not materialise · P 1.5 – 2.0 6 / 6 measures 2.00 4%
Partial aspect supported · P 2.5 – 3.0 8 / 6 measures 2.88 6%
Load-bearing link disputed · P 3.5 – 4.0 14 / 6 measures 3.93 10%
Coverage gap · P 4.5 8 / 5 measures 4.50 6%

142 of the 137 arguments below 5.0 name the band that puts them there. The rest sit low without saying whether nothing carries them or something speaks against them; they are brought up to date topic by topic.

What the scale is anchored to

The numbers are not chosen freely. They lie on the probability bands that other disciplines have worked with for decades: the likelihood bands of the IPCC and the probability levels of ICD-203 (“likely” 55–80%, “very likely” 80–95%, “almost certain” 95–99%).

  • P 9.5 hard cap: nothing above it except identities
  • P 9.0 lower bound of “very likely”
  • P 8.5 “very likely” (ICD-203, 80–95%)
  • P 7.0 middle of “likely” (55–80%)
  • P 6.0 below the “likely” threshold
  • P 5.0 about as likely as not

Where Argupedia deliberately departs

  • Mechanisms rank at the top here, not the bottom. In medical evidence hierarchies the study is the gold standard and the argument from a causal relationship is the weakest link. Politics has identities that no study can beat: a deleted budget line saves exactly that amount.
  • Numbers instead of words. The bounds are set as figures, not as verbal categories. The criticism of vague probability words goes back to Sherman Kent’s essay of 1964 and ends in fixed percentage bands in ICD-203.
  • The hard cap sits below “virtually certain”. The top level of the IPCC reaches 99–100 percent; Argupedia’s cap ends below that, at 9.5. Only what is no longer a forecast at all gets past it: an identity or the direct legal consequence of the text.

What leads above the cap, and what below

The cap of the basis is the starting point, not the last word. Three rules raise it, three bind it. Each has to be named in the argument; none applies silently.

+0.5 … +1.0

Combination

At least two independent bases: a different type, a different data source, the same direction, a compatible order of magnitude. The starting point is the highest single cap.

Single study + precedent → at most 8.0 · hard cap 9.5

→ P 10

Directness

If the effect is the direct legal consequence of the adopted text (a budget line, a tax rate, a cancelled benefit), P 10 is permitted for this item; with a named enforcement risk, 9.5. What follows from it is a separate argument with its own plausibility.

↑ level

Upgrade

A projection with a published, externally checked track record counts as a study. A forecaster who has demonstrably been right for years is not penalised for calculating into the future.

−1.0 … −2.0

Transfer deduction

Evidence from another country, another time, another scale. If a single precedent decides the ranking, the expected deviation has to be quantified. External validity is around 50 percent at the median.

required

Counter-mechanism

An argument from a mechanism names the condition under which it holds and the counter-mechanisms that were checked. Without this statement the basis does not carry.

cap

State of research

If the finding is contested or refuted, that caps P as well: promising ≤ 7.0 · contested ≤ 5.0 · refuted ≤ 2.0.

What the stock actually claims

The cap table says what would be permitted. These numbers come live from the database and say what is used, above all how often the cap binds rather than the evidence. They count the whole English-language stock in its latest version.

Two caps, not one. Next to the evidence basis stands the rule cap from the table above: what this kind of evidence carries in general. The last column, by contrast, counts the arguments that sit on their own cap: the value left for this one argument after additions and deductions, as shown on its card. It can lie below the rule cap, on it, or above it with an addition, and for a mechanism there is no rule cap to hold it against. The two numbers therefore cannot be converted into each other.

Evidence basis · rule capArgumentsAvg P at their own cap
Mechanism · no rule cap 135 / 44 measures 5.69 38 · 28%
Projection · ≤ 6.0 100 / 36 measures 5.41 39 · 39%
Study · ≤ 9.0 61 / 25 measures 5.02 6 · 10%
Plausibility · ≤ 5.0 46 / 25 measures 3.51 22 · 48%
Precedent · ≤ 8.5 39 / 20 measures 5.15 7 · 18%
Converging studies · ≤ 9.0 23 / 14 measures 6.00 0 · 0%
Multiple precedents · ≤ 8.5 9 / 6 measures 5.89 4 · 44%
Single precedent · ≤ 7.0 4 / 3 measures 4.00 0 · 0%

How P is distributed across the stock

0
2
4
6
8
10

Every argument names an evidence basis, since the schema lets none through without one: here that is all 417. The telling column is the last one: 116 arguments sit on their own cap. A high rate is not a good sign. It means that there the rule sets the limit, not the evidence.

Ask the bot

The bot explains the method and the debate on this page. It answers only from Argupedia's content and does no research of its own.

Before you ask

  • The answers are written by an AI model (Claude Haiku by Anthropic). They may be wrong or incomplete. The content of the pages is authoritative.
  • The bot gives no legal, financial, medical or voting advice.
  • What you type is sent to Anthropic in the USA. Do not enter names, contact details or other personal information.

The terms of use (section 9a) and the privacy notice for the chat apply.