Perfect recall, 42.9% false positives: what stacking five detectors actually buys

Noisy-OR earns its keep when detectors are independent witnesses to one event, because then several hedged yeses compose into one confident yes ^[12]. Five readings of the same figure-dense sentence are not several witnesses. A digit-heavy line nudges the rules tier. A section header nudges the repetition sketch, a formulaic passage nudges compressibility, and none of them is alarmed on its own, but the rule multiplies survival probabilities and turns five scores of 0.3 into 0.83 ^[13]. The clean corpus makes this unavoidable: economic prose thick with figures, percentages, variable codes and markdown tables, which the author calls the most corruption-looking legitimate text you could ask for ^[11].

Raising the bar does not rescue it. To spare that benign pile-up you need a threshold above 0.832 ^[18]. But with one detector active and the other four silent, noisy-OR returns exactly the active score, so a single saturated detector now has to clear 0.832 by itself ^[18]. At that operating point the fusion contributes nothing a maximum over the five scores would not, and the saturated-single case is precisely what the author says real corruption produces: one or two pinned signals, not five mild ones ^[14].

The reason this trap is easy to walk into is that the architecture prices detectors at zero. One pass over the character stream, one state object updated per character, detectors as stateless reads at checkpoints, so a sixth detector adds no passes and no memory growth ^[6]. The guard runs at about 50,000 characters per second on a single CPU core with no GPU ^[9]. The one corrupt generation the post measures, a 13,200-character table echo, took 269 seconds to finish ^[5], which is a generation rate near 49 characters per second ^[15] and puts the detector roughly a thousand to one ahead of the thing it is watching ^[16]. Detection also worked on the first try: every configuration tested reached a true positive rate of 1.000 ^[10]. Nothing in the compute budget or the recall numbers tells you to stop adding detectors. The fusion rule is where each addition is charged, and it charges against healthy output.

Which reframes what the five tiers are. They are not five votes on whether a stream is corrupt, they are five distinct failure signatures, and the rolling SimHash exists because fluent same-script regurgitation sits about 0.5 normalized Hamming from the clean-prefix fingerprint while leaving character statistics untouched ^[8]. A rule that sums weak evidence spends that specificity to buy nothing, since coverage was already complete before fusion entered the picture ^[10].