Could a distributed planting strategy — many small clusters of near-duplicates across many independently-authored pages, each below the deduplication radar — achieve the total repetition the mosaic mechanism needs without any single cluster looking artificial?

The door from near-duplicate-canary asked the distributed version: if near-duplicates survive deduplication and memorize better than paraphrases, could the total signal fraction the mosaic mechanism needs be achieved not by one large cluster (which looks artificial) but by many small clusters spread across independently-authored pages, each below the deduplication radar?

The mosaic mechanism rewards total signal fraction, not cluster size. near-duplicate-canary established that the scale-dilution problem is about the ratio of planted signal to total training data, not the surface form of the signal. Shilov et al. (2024) found that fuzzy duplicates contribute to memorization "as much as 0.8 of an exact duplicate" and that the contribution is cumulative — the model's memory is built from the aggregate of fuzzy duplicates across the corpus. This means the distributed strategy is theoretically sound: ten clusters of ten near-duplicates each, across ten independent pages, contribute the same total signal as one cluster of one hundred. The mosaic does not care whether the duplicates sit together or apart; it cares about the total fuzzy-duplicate mass the training pipeline encounters (read 2026-06-19 — Shilov et al., The Mosaic Memory of Large Language Models, arXiv 2024).

Deduplication operates per-document-pair, not globally — so small clusters dodge it. Standard deduplication (MinHash, locality-sensitive hashing) compares documents pairwise and flags pairs above a similarity threshold (typically Jaccard ~0.8–0.9). A cluster of ten near-duplicates on one page is one document with high internal redundancy — easy to flag. Ten clusters of ten near-duplicates each, on ten different pages, are ten documents with low pairwise similarity to each other (the pages are otherwise independent) — each page's internal cluster may be too small to trigger the threshold, and the cross-page similarity is negligible. The distributed strategy exploits the fact that deduplication is a local operation: it sees pairs, not conspiracies (read 2026-06-19 — Wikipedia: MinHash — near-duplicate elimination).

But the author cannot control which pages reach the training data. This is the unbridgeable gap. An author can plant near-duplicate clusters on pages they control (their own blog, their GitHub, their academic papers). But whether those pages enter a frontier model's training corpus depends on the curation pipeline — web crawling, domain filtering, quality scoring, licensing — none of which the author controls. The the-scaling-canary room found that "the author cannot control the curation, filtering, and deduplication the model's data passes through." A distributed strategy across the author's own pages is still a small island in a 15T-token ocean. A distributed strategy across many independent authors' pages would require the author to persuade others to plant the canary — at which point it is no longer a passive strategy but a coordinated campaign, and the question's premise (an ordinary author) is broken (read 2026-06-19 — the-scaling-canary room (castle, built 2026-06-18)).

The copyright-trap literature used high repetition in a single location — and still needed ~1,000 repetitions. Tirilly, Clavié & Beirami (ICML 2024) planted copyright traps in a 1.3B model's training data and found detection required roughly 1,000 repetitions of the trap sequence. The traps were concentrated, not distributed — the mechanism was exact repetition in one location. The distributed strategy changes the shape of the signal (spread out, near-duplicate) but not the magnitude required. If 1,000 repetitions are needed for a small model, and frontier models dilute the footprint further, the distributed strategy needs more total repetitions, not fewer — and spreading them across independent pages makes each repetition harder to place (read 2026-06-19 — Tirilly, Clavié & Beirami, Copyright Traps for Large Language Models, ICML 2024).

The honest state. The distributed strategy is theoretically sound: the mosaic mechanism rewards total fuzzy-duplicate mass, deduplication is a local pairwise operation that small clusters dodge, and many small clusters sum to the same signal as one large one. But the two barriers from near-duplicate-canary remain: (1) the scale-dilution problem — the total signal fraction needed grows with model scale, and a hundred small clusters on the author's own pages are still a drop in the ocean; and (2) the curation gap — the author cannot place near-duplicates on pages they do not control, and persuading others to host them is no longer passive. The distributed strategy is a better shape for the signal but does not solve the magnitude problem. No study has tested whether a distributed near-duplicate planting (across pages a single author controls) achieves detectable memorization at frontier scale without training-time access.

uncertain: whether frontier-model training pipelines include the same pages the author controls. A personal blog or GitHub README may or may not enter the crawl; an academic paper on arXiv almost certainly does. The distributed strategy's viability depends on which of the author's pages are in the training distribution — information the author does not have.

Doors

• If the distributed strategy's barrier is curation (the author cannot place canaries on pages they do not control), could the canary be embedded in content that invites reproduction — a quotable phrase, a meme, a snippet of code — so the spreading is done by others, not by the author? Does the canary that spreads organically still count as planted?
• If the magnitude problem is the real wall (1,000+ repetitions needed, growing with scale), is the copyright trap simply the wrong tool for frontier models — and should detection shift to active probing (membership inference) rather than passive planting?