Loaded example August 6, 2026. We scored “Security keys stop 100% of phishing.” at 43 out of 100, Half True. A council of five models, scoring the same claim from the same sources, put it at 43. Every slider below is set to the component score we published. Disagree with one and you will see exactly how much it was carrying.

The claim

Source qualityPeer reviewed, preprint, self-published report, vendor material, or press release.

3/ 20

Peer reviewedPress release

Sample and methodSample size, how subjects were selected, and whether the underlying data is available.

5/ 20

Large, random, data publishedTiny or self-selected

IndependenceDoes the publisher sell the solution its own finding recommends.

8/ 20

No commercial interestPublisher sells the fix

ReplicationHas anyone independently reproduced it with a comparable method.

10/ 20

Independently reproducedNever reproduced

DriftHow far the popular use of the claim sits from what the source actually says.

17/ 20

Used exactly as measuredUsed far outside its scope

50Hype Index

Half True

A line you can quote

How to answer each one

Source quality

Ask what the artifact actually is. A peer-reviewed paper, a regulatory filing, and a vendor deck are three different objects and only one was reviewed by people with an incentive to find problems. Score low for peer review, high for marketing material. If nobody in the room can name the source at all, score it 20.

Sample and method

Look for the sample size and how subjects were selected. A few dozen volunteers is not a population. Self-selected survey respondents are not a random sample. Score low when the sample is large, randomly drawn, and the underlying data is available for inspection.

Independence

Ask whether the organization publishing the finding sells the remedy the finding recommends. This is not an accusation of dishonesty, it is a disclosure that should already be attached to the number. Score low for a neutral party, high when the publisher benefits directly from the conclusion.

Replication

Ask whether anyone independent has reproduced it with a comparable method. A single unreplicated result is a starting point, not a conclusion. Score low when several independent teams found the same thing, high when it has never been tested twice.

Drift

This one catches the most claims. Compare the population and conditions the study measured against the population and conditions the claim is now being applied to. A sound finding used far outside its scope still leads to a bad decision. Score high when the leap is large.

What the number means

Underrated0 to 19

The evidence is stronger than the attention it gets.

Holds Up20 to 39

The claim survives contact with its own source.

Half True40 to 59

Real finding, wrong scope. Correct in part, misapplied in practice.

Overstated60 to 79

Sound evidence, carried far past what it actually measured.

Unsupported80 to 100

No source that survives inspection.

Bands and anchors are rubric v1.0, effective July 26, 2026, the same version every edition is scored against. This is a structured judgment rather than a measurement. Editions are scored by a council of five models working from the fetched text of the sources, and within a single run those five land as much as thirty points apart on the same claim. If your score lands a few points from ours, that is the tool working. Send the disagreement to [email protected] with the component you think is wrong. See it applied on the record.

This is how a published score is built.