The whales are real. A minute late, the edge is gone.

Piece 01 took one wallet apart. This one takes the whole table. The finder reads every taker fill on every temperature market Polymarket listed in a fourteen-day window, groups them by wallet, matches buys to later sales first-in first-out, settles whatever is still held against the venue's own resolution, and scores each wallet on the buys that have actually finished. A wallet ranks if it bought at least twenty times and paid at least five hundred dollars. Nine thousand eight hundred and ninety-eight wallets traded. One thousand six hundred and ninety rank.

They are not lucky

Pooled over all 1,690, the ranked wallets make 0.72c a share across 902,898 buys and 21.7 million shares. The 8,207 wallets too small to rank lose 1.38c a share. That gap is the whole market in one line: the money in these contracts moves from a large number of small accounts to a small number of large ones, and it is not close.

The top of the table is concentrated. The twenty-five wallets the run publishes paid $1,309,036 and took $178,024 gross, 13.6% on money deployed over a fortnight. But "the whales" are not one thing. Sorting them by how they actually behave — when they buy relative to the market's own day, how long they hold, which side and which part of the ladder — four distinct animals fall out, and they have almost nothing in common.

per share, gross, on the buys the mirror could evaluate; profiles are assigned from the fills alone

The fastest wallet in the table, Dual-Aunt, holds for a median of five seconds. It put $149,423 through 759 markets in 47 cities, 99% of it on the No side, 87% of it on the market's own day, at a median entry of 2.9c. That is exactly the machine piece 01 pulled apart, run at scale. The largest, Bland-Dynamite, made 21,432 buys across 2,729 markets in 51 cities and cleared $2,488 — 1.66c a share, 2.6% on money deployed. Volume and profit are different sports.

Now copy them

This is the question everyone actually asks, so we measured it rather than argued about it. For every buy a ranked wallet made, take the moment it printed on Polymarket's public feed — the feed anyone can watch — wait a fixed lag, then lift the next ask on the same outcome at whatever price is there, up to a six-hour limit. Hold it to settlement. No order is placed anywhere; this is a paper mirror over prints that already happened.

The copier gets filled. 93.8% of the buys find an ask within six hours at a sixty-second lag, 92.5% at five minutes, and for most profiles the median slippage is zero — the ask has not moved. Filling is not the problem.

Pooled over all 1,690 ranked wallets: they make +0.72c a share. The copier makes −0.07c. The edge is not in which contracts they pick. It is in when they get there.

Stretch the lag to five minutes and it gets worse, not better: the copier goes to −0.11c. And the pattern across profiles is the tell. The faster the wallet, the less copyable it is. The observation bots make 1.62c a share of their own and hand a copier −0.19c, because a strategy whose median hold is five seconds cannot be followed sixty seconds later — by then the stale bid it swept is gone, and the copier is buying from the wallet it is trying to imitate. They are the only profile where slippage bites: 0.61c at a minute, 0.80c at five.

The one that survives

One profile does survive copying. The 66 longshot buyers make 1.78c a share and still hand a copier +1.43c. That is real, and it is also the least useful finding in the piece, because of what the strategy is.

The clearest example is the number two wallet in the table, Bewitched-Supernatural. It paid $1,504 and took $21,426 — up 1,425% — on 137 buys at a median entry price of 1.4c, over four active days, every one of them on the market's own day, all of it settled. That is not an edge you can lean on. It is a wallet that bought a great many lottery tickets in a fortnight and had some of them come in. Fourteen days cannot tell that apart from skill, and neither can we. It is on the table because the arithmetic puts it there, with the caveat attached.

Does the table hold still?

The finder has now run three times, on windows a day apart. Twenty-one of the top twenty-five appear in all three, four dropped out and four are new. The top four hold ranks 1, 2, 3 and 4 in every run. The mean wallet moves 2.6 places between the first run and the last, and sixteen of the twenty-one move two places or fewer.

That is much weaker evidence than it looks. These are fourteen-day windows two days apart, so they share twelve of their fourteen days. Of course the same wallets rank — it is very nearly the same data scored twice. The honest reading is the opposite one: two days of turnover was enough to move one wallet seventeen places, from fifth to twenty-second. If a leaderboard built on a fortnight can be reshuffled that far by swapping two days, a leaderboard is not a thing to trade off. Testing whether any of this persists needs windows that do not overlap at all, and that is a month of waiting, not an afternoon of querying.

What it means for the desk

Three things, none of them a recommendation.

• The public-feed copy trade does not work. Not because the whales are fake, but because a settled-fill feed tells you what someone did after the price already moved. It is the opposite of a mempool: there is no pending intent to front-run, only a receipt.
• The edge that does exist is a latency edge. The profitable wallets are not forecasting better than the market. They are reading a public weather observation and acting on it before the price admits what it means. That is a clock problem, and the desk is now measuring exactly that on Kalshi — how long it takes a bracket to reprice once a station reading has killed it.
• US persons cannot trade Polymarket. Everything above is measurement of someone else's venue, not a route to a position. On Kalshi, where a US person can trade, there are no wallet identities to rank at all — which is why the desk measures the tape and the book there instead.