> but if you consider each person an imperfect detector of information, then the more people reporting a thing, the more accurate the picture becomes thanks to sensor fusion/the Kalman filter:
Not necessarily. Adding a bunch of extra imperfect sensors is not the same as adding imperfect sensors with imperfect interpretations. The storage medium that interpretation is stored in is also imperfect. An accounting of that interpretation will change non-linearly depending on how many times it's described and how recently from the event it's described.
Adding more and more of these recountings don't necessarily mean you've got better data.
Yes it does. If you simply treat that drift as additional "error", you're still fundamentally dealing with an imperfect sensor with an error rate. And mathematically, that makes it all still fall under the Kalman filter metrics. (To be fair, you'd still need to quantify the reported thing somehow.)
Which should make intuitive sense. If 1 person claims they got raped by someone, and that is the only evidence against them, you might be skeptical. If 50 people claim they got raped by someone, and that is the only evidence against them, a reasonable position IMHO would be that the chances that this person is not a rapist are vanishingly small, even if (on average) 5 to 10% of the reports are non-truthful and motivated by perverse incentives (such as avoiding shame of promiscuity).
At some point we need to figure out how to deal with eyewitness evidence without complete dismissal, because complete dismissal of it is wrong (it's literally gaslighting- telling people that their reported experiences are irrelevant/invalid, which is the same as asserting that they didn't happen IMHO). Plenty of experiences (both positive and negative) happen that only have the retelling as the evidence, and ignoring all retellings omits a possibly large amount of evidential source of truth.
Not necessarily. Adding a bunch of extra imperfect sensors is not the same as adding imperfect sensors with imperfect interpretations. The storage medium that interpretation is stored in is also imperfect. An accounting of that interpretation will change non-linearly depending on how many times it's described and how recently from the event it's described.
Adding more and more of these recountings don't necessarily mean you've got better data.