The claim that a dealer level holds is testable. Doing it honestly is harder than it sounds, and the difficulties are where most published claims quietly fall apart.
You need the level as it was, not as it is now
This is the first and largest obstacle. To test whether a level computed at ten in the morning held, you need what the map said at ten in the morning — not what it says now, and not what it would say if you recomputed it today from an end-of-day chain.
Recomputing historically from settled data is the standard shortcut and it is a lookahead bias. The end-of-day chain includes positioning that did not exist at ten, and the resulting level is one no trader could have seen.
Only a stored series of what was actually published at the time avoids this.
Define held before you look
Held is not a fact, it is a definition. Price touched and reversed within some tolerance, for some minimum distance, within some window. Every one of those is a choice, and choosing them after seeing the data is how you get a result that means nothing.
Write the definition down first. Then apply it to every case, including the ones you would rather not count.
Every case, not the memorable ones
Selection is the killer. The days a level worked are vivid and the days price sailed through without pausing are forgettable. A test that starts from cases you remember has already reached its conclusion.
Report the base rate
The honest output is a frequency with a sample size, compared against what price does in general. A level that holds slightly more often than chance is a small real effect. One tested on nine cases is a story, whatever the frequency says.