For those who are interested in the assessment of the German intel (BND), here is a quick explanation, and comparison with US assessment for the FBI, CIA and DoE.
BND:
➡️ probability of research accident at 50-80% using public information.
➡️ at 80%-95% when adding classified information.
These two probability ranges refer to the likelihood that a research accident has happened.
They are in line with the US ranges defined by the US Intelligence Community Directive 203 (ICD 203, Analytic Standards, pub. 2007, rev. 2015):
An analyst can choose any of the 3 ways (given by the three rows), but to avoid confusion, an analytic product should not start mixing up terms from different rows.
Note that the US range for 'likely' is 55%-80%, not 50%-80%.
dni.gov/files/documents/ICD/ICD-203.pdf
It could be a particularity of the German intel probability yardstick, or a mistake (55% -> 50%).
I have no idea what that probability yardstick is. It's not the NATO standard, for sure:
researchgate.net/publication/335771404_Variants_of_Vague_Verbiage_Intelligence_Community_Methods_for_Communicating_Probability
For full comparison, the current UK standards are below. They are a pain in the back, with their odd gaps. A good example of British logic IMHO.,
Yardstick from the UK Professional Head of Intelligence Analysis (PHIA):
assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/843803/intelligence-collection-management-dissemination-pf-31-oct-2019.pdf
All in one image:
dx.doi.org/10.1080/02684527.2017.1394252
Anyway, that's half the fun.
The likelihood/probability of an event to happen/have happened is just one part of the analytical judgment, The other part is the confidence level in that probability.
That confidence level/rating refers to the complexity of the scenario, quality of data/sources, strength of the argumentation.
In the US (image above), that's given by the JP2 (Joint Intelligence Publication 2-0, 2013):
Typically, there is standard Confidence grading using discrete levels for each of the three factors.
This can result in a 5x5x5 matrix, or whatever is used in some agencies.
For instance, in the US source descriptors are used for source quality and credibility. See UCD 206, and recent UCS 206.01.
Now, to not confuse anybody, an intel product should deliver both the likelihood and the confidence rating, but not in the same sentence.
The idea is that by using separate sentences to deliver each, one may pay more attention to the distinction between likelihood and confidence when reading the intel product.
Now, for the layman, that is actually very confusing.
So there can be exceptions for non-intel products.
For instance, in the Declassified Assessment of Oct 2021, we get;
👉🏻 FBI at likelihood: 'most likely' (i.e. 'likely' in the standards) an accident, with 'moderate' confidence.
👉🏻 4 other ICs at 'most likely' (i,.e 'likely') not an accident, with 'low' confidence.
Since then, we know that DoE is now at 'most likely' accident (again 'likely' in the standards), with 'low' confidence.
And we can say that the BND at 'very likely' accident, but with no clear confidence level reported in the German press (??). @holger_stark
By the way, if the probability/confidence framework looks ad-hoc and complex, please know that it is really a work still in progress.
It used to be much more of a mess, before Sherman Kent started cleaning up the approach in the US,
Another useful thread on this general subject:
x.com/gdemaneuf/status/1565082019742113793
Gilles Demaneuf
@gdemaneuf
Pointy Head. Opinions, analyses and views expressed are purely mine and should not in any way be characterised as representing any institution or company.