Lets look at two mind experiments. Experiment 1. We generate pair of entangle photons in opposite polarization states. At Alice's wing we have polarizer with polarization axis oriented vertically and detector that registers "clicks". At Bob's wing we have the same setup - polarizer with polarization axis oriented vertically and detector that registers "clicks". And we have coincidence counter attached to Alice's and Bob's detectors. As we register only photons with the same polarizations we should have perfect negative correlation e.g. no coincidences at all. Experiment 2. We take the same setup as in experiment 1. but slightly modify it. We add two polarizers to Bob's wing after first polarizer - one rotated by 45° relative to first polarizer other rotated additional 45° relative to previous polarizer so that we have rotated polarization axis by 90° from that of first polarizer. Now according to Malus law we have reduced intensity of light (count of photons) to 1/4 that of unmodified setup. But as we register only photons with opposite polarization states we should have perfect positive correlation e.g. maximum coincidences. Are expected results correct from perspective of QM? As I see we have exactly the same result if we replace all polarizers with polarization beam splitters and dump one unnecessary output. Where I see paradox. All photons that are registered in Bob's detector in second experiment should be absorbed by first polarizer (dumped in case of PBS) as demonstrated by first experiment. If all above is correct then I think this clearly refutes fair sampling assumption in EPR experiments e.g. photons that are registered in Bob's detector in second experiment are undetected (and are "not entangled") in first experiment. Does it seems correct?