Endorsements are the cheapest evidence there is. They cost the endorser nothing, they are usually returned in kind, and nobody has ever been asked to prove one.
So any system that adds them up rewards the person with the widest network, not the person with the skill. That is not a flaw in the people; it is a flaw in addition.
What the model actually does with ten vouches
One vouch from one person scores 48. Ten vouches from ten different people score 60: more, because ten independent people agreeing is worth something.
But it cannot go higher than 60, ever, no matter how many more arrive. That is the ceiling for vouching as a kind of evidence, and it is set by what a vouch can prove, not by how many there are.
Compare one live challenge: a timed problem the person solves there and then, marked by a system that cannot be charmed. Passed convincingly it scores 100. One event outranks ten endorsements, because one event is harder to fake than ten favours.
The same source doesn't count twice
Three vouches from three colleagues at the same company score 49. Three from three different companies score 56. Colleagues at one company mostly confirm that the first one said it again; the model keeps the strongest and steeply discounts the rest.
Different kinds of evidence corroborate
A modest challenge result alone scores 70. One vouch alone scores 48. Together they score 93, higher than either, because two independent kinds of evidence agreeing is hard to manufacture.
That is the model in one sentence: cap each kind, discount repeats from one source, reward agreement across kinds.
Every number in this article comes from the code that scores a real applicant. If the model changes, a test fails until this article says what the model does.