The unit of comparison is the whole trick. Jaccard over a document's shingle set divides shared 8-grams by total 8-grams [3], so a paragraph that is word-for-word identical in two articles gets diluted to a small fraction of a long page. The scan in the dev.to writeup does not compare documents. It walks every pair of same-position paragraphs across a batch [5], which is why it surfaced pairs sitting at 1.00 in a corpus that passed uniqueness checks [7][2].
Then the 8-gram budget. A span of W words contains W - 8 + 1 8-grams, so a 20-word slot contributes 13 [18], and the author puts that at 60 to 70 percent of the paragraph's entire shingle set [11]. Run it backwards: 13 shingles at that share implies 19 to 22 shingles in the whole paragraph, about 26 to 29 words [19]. That paragraph was one long sentence plus a clause. Fix the sentence and nothing else in the block can rescue the fingerprint, which is why more variants per block moved the numbers so little [10].
The fix's arithmetic is where I would push back at review. The design calls for 3 to 4 slots per paragraph, each 12 words or fewer, each with 3 to 6 phrasings [13]. The stated space is 216 to 1,296 combinations for a four-slot paragraph [14]. Four slots of six options is 1,296. Four slots of three options is 81, not 216; 216 is 6^3, the three-slot corner at six options each [20]. Take the real floor of the stated range, three slots of three options, and you get 27 combinations serving 70 articles [21], which is the exact failure the post's own rule of thumb warns against [15].
Hashing does not allocate, either. pick() takes an MD5 of slug, slot name and reseed, modulo the option count [13]: sampling with replacement, not dealing from a deck. Seventy articles make 2,415 pairs. Divide by the combination space for the expected number of pairs landing on identical text, and you get roughly 1.9 at 1,296 combinations, 30 at 81, and 89 at 27 [22]. A space three times larger than demand is not a collision-free space. At the top of the range, a couple of shared paragraphs inside otherwise distinct articles is survivable; at 81 it is the old problem in smaller print.
What the material does not establish is the ranking half. That traffic works out to a click-through rate of 0.014% [17], with no penalty notice and no post-fix ranking series, since the text breaks off mid-sentence on the limits of per-slot independence [16]. The Jaccard 0.5 line described as the threshold where near-duplicate detection gets interested is the author's characterization, not a published search engine figure [8]. For the diagnosis to transfer, you would need rotated blocks that survive into the indexed HTML and no competing explanation for the flat positions. The detector itself is two functions over text you already have [5], so the measurement is cheap and yours. The causal story attached to it is one operator's case.
