How RN1's Bot Works, Read From 911,860 Fills
Polymarket · Polygon · 10 Aug – 10 Nov 2026 · 92 days
(This describes how one wallet's on-chain fills and signed orders were used to reconstruct the logic behind them. It is a reconstruction, not RN1's code, and it is not trading advice.)
"RN1 Doesn't Need to Know Who Wins" finished on a list of things I could not see, starting with the orders nobody hit and the ones RN1 cancelled, and it closed with a line of small print saying the four strategies had also been rebuilt as pseudo code. This is that piece. The first one asked what RN1 earns, and the answer was a toll of about a cent and a half per pair, collected on a quote one cent wide from people who want in on a game right now. This one asks what kind of program would leave exactly the trail RN1 left, and how far 911,860 fills can carry you before the trail runs out.
Somewhere in those fills are 14,818 that hit one of RN1's resting bids in the same block as its bid on the other side of the same market, in 5,236 blocks across 2,679 markets. Nobody outside the operator has seen the code, but two bids filling in one block means two bids were resting at once, and that small fact is a line of the program recovered from outside. Most of the reconstruction works like that, one footprint at a time, and most of it is less certain.
What a Fill Will Tell You
A fill is a receipt. Each OrderFilled log on Polymarket's exchange records that somebody's order met one of RN1's, which side was resting1, the outcome, the shares, the price, the fee and the block it landed in. The receipt is thin, but the transaction that carries it holds the whole order RN1 signed: the size it posted, the time it signed to the millisecond, a salt2 and the signing key3. An order's price is fixed, so a fill at a new price is a new order, and cancels and orders nobody hit stay unseen. This part is preliminary: I have read the signed orders in two samples, 415 across the window and all 1,117 from 74.8 minutes of one evening, 16 September. Reverse engineering here means writing the simplest program that would produce these receipts, then marking every place where a different program would too.
The salt is a footprint too: it fits only Polymarket's own Python client, though a custom client that copied the rule would look the same.
Split the fills by whether RN1's order was resting or crossing, and by whether the game had started on Gamma's4 schedule, and you get the four strategies from the first piece, quoting and taking, each before and during games.
Where the receipts fit more than one build, the teardown lists the builds and puts a percentage on each, a judgement of how strongly the evidence favours it and not a fitted model, adding to 100 within a point. Three points are measured: how RN1 gets out of its positions, how wide it quotes, and whether what it holds moves the quote. The other fifteen are weighed.
The Skeleton Under All Four
# Shared by all four strategies. for M in gamma.markets(tag = "sports", outcomes = 2): U inferred watch(M) on fill(M, outcome, shares, price, fee): held[M][outcome] += shares cost[M][outcome] += shares * price + fee # no sell path: 0 sells in 911,860 fills on resolved(M): X measured pass # Polymarket's auto-redeemer pays out about 12 seconds later if cash_is_short(): C inferred pairs = [(M, min(held[M][A], held[M][B])) for M in open_markets] for chunk in chunks(sort(pairs, by = size, largest_first), 10): merge_all(chunk) # one transaction per 10 markets, 1 pUSD back per pair
The market list is read from breadth and depth. RN1 has fills in 87,642 markets across 110 series and 18 market types, and half of those markets got 3 fills or fewer and cost under 257 pUSD5, which is the wide, shallow pattern a rule leaves. Inferred U: every two-outcome sports market Gamma lists, by rule (65%); sports markets that pass a volume or reward filter (25%); a hand-kept list of leagues (10%).
The sell path is the easiest decision in the teardown, because there is nothing to decide. 62,255 of 62,328 redemptions came from Polymarket's auto-redeemer6, and RN1 redeemed for itself at one moment only, which comes up below. Measured X, share of the pUSD it got back: held to resolution and auto-redeemed, 98.65%; merged before resolution, 1.35%; redeemed by its own call, 0.00%; sold, 0.00%.
The merge7 sweep is where the loop comes most directly from the chain: 434 merges went out in 45 transactions, 433 of them in games still being played. The two big sweeps ran in the two hours with the most pUSD tied up of all 2,202, at 1.88M and 1.87M, and the wallet's pUSD balance shows what fired them: they began 102 and 78 seconds after its cash fell under 200K.
pUSD tied up in markets not yet resolved
Inferred C: cash runs short and the operator runs a script by hand (60%); cash runs short and a floor in the bot fires, one that was not there on 8 September (40%). On that day cash sat under 200K for 41 minutes beside 885,864 pUSD of pairs RN1 could have merged, and it merged nothing. In the first piece I leaned towards a person, and the first big sweep is why: it opens with a 109 pUSD test, then the only redemptions RN1 made for itself in the window, and the rest follows 213 seconds after the test. The second has no test, which fits a script or a floor, and twice merges went out with cash to spare, which no floor would order.
Reading a Bid From the Fills That Hit It
on game_start(M): S fills jump at the start quoting[M] = true while quoting[M]: B both bids live at once bid[A] = tick_below(fair_price(M, A)) F how often fills win bid[B] = 0.99 - bid[A] E sum of the two bids size = base_size(league(M)) Z sizes on signed orders # held[M] moves neither bid nor size I quote width by holdings for outcome in (A, B): keep_ladder(M, outcome, bid[outcome], size) L bids signed together wait(book_change or fill) R age of orders that filled on game_end(M): T last fill in each market cancel_all(M) quoting[M] = false
The start is read from timing. Maker fills jump from 3,671 in the 5 minutes before the start to 24,728 in the 5 minutes after.
Fills by minutes since the game started
Inferred S: at the listed start time, or when the market turns live (75%); bids rest earlier but far from the price until the start (15%); on the first in-game event from a score feed (10%).
Both sides are read from coincidence. Beside the same-block fills, a second test finds 21,440 fills made while an order on the other outcome was provably live, though in only 4,232 markets. Inferred B: bids rest on both outcomes at once wherever it quotes (92%); on both in busy markets, one at a time elsewhere (8%).
Fair price is read from calibration, meaning how often fills at a given price go on to win, and on that test RN1's maker fills sit within 1.1¢ a share of fair in all ten 10-cent bands.
Profit per share bought, by price band
A second signal comes from who beats it, since it loses 504K pUSD to its 50 busiest takers and makes 1.4M from the rest, so somebody's price is ahead of its own. Inferred F: the market's own book, mid or last trade, perhaps smoothed (55%); an outside feed, sportsbook odds or a live win model (30%); a blend, an outside price before the game and the book during it (15%). My own lean is the book: a model for each of 110 series is a lot to build, and a quote one tick8 wide is the inside of the book while both bids rest, so it needs no outside price to explain where it sits.
The margin is measured by adding the two bids. Measured E, share of the 5,236 blocks where both bids filled: a sum of exactly 0.99, 71.77%; wider, 22.84%; between 0.99 and 1, 4.87%; 1 or above, 0.52%. On the second test it is 79.27%. A pair pays 1 pUSD, so two bids that sum to 0.99 make a quote one cent wide, the tightest a one-cent grid allows. The measure favours busy, tight markets, since both bids must fill, and it cannot say whether RN1 sets the best price or joins it.
Size is read from the signed orders. All 315 resting orders sampled across the window were posted in whole shares, only 7.3% in a whole pUSD amount. On the sampled evening the size followed the league, about 1,500, 3,000, 5,000 or 7,500 shares, the same on both outcomes in 95 of 99 markets and from one order to the next. Inferred Z: a fixed share size per league (80%); a size that follows the book's depth or volume, which held still that evening (20%). A fill takes a median 3.3% of the order it hits, so fill sizes say little about posted ones.
A market maker that wanted a flat book would buy the outcome it holds less of, and RN1 does that on only 42.87% of the buys that change its balance.
Which outcome each buy added to
One resting order filled by several takers counts once per fill, which pushes the overweight share up.
But a falling price hits the same bid again and again, so the cleaner test of a lean is the quote itself. Measured I, share of bid pairs one tick wide, by the gap between its holdings: under 500 shares, 71.86%; 500 to 2,000, 75.00%; 2,000 to 10,000, 72.14%; 10,000 or more, 71.18%. Posted size was the same on both outcomes on the sampled evening, so what RN1 holds moves neither price nor size, though a bot that pulled the heavier bid outright would escape this test.
Levels are read from orders signed together. On the sampled evening RN1 signed 20 pairs of bids one cent apart and 6 sets of three on consecutive cents in the same millisecond, the same size on each. Across the window 88.59% of the gaps between its levels filled in one block are one cent. Inferred L: two or three bids on consecutive cents, topped up as they fill (70%); one bid most of the time, a second in busy markets (20%); a deeper ladder whose lower levels never fill (10%). A taker's order filled 2 or more of RN1's orders in only 2.63% of cases, which says little once a fill is known to take 3.3% of an order: the taker is spent on the top bid.
Speed is read from the age of filled orders. Resting orders that filled were a median 11.8 seconds old in the sample across the window, and the same in-game on the sampled evening. A scan for a fixed timer finds none, and after one sharp price move RN1 signed 26 orders in 3.1 seconds. Inferred R: on each book change or fill, within seconds (80%); on a timer too uneven for the scan to catch (20%). Pre-game orders that filled were older, and how long RN1 leaves an order it then cancels is unseen.
The stop is read from the last fill in each market. In busy leagues it comes at the final whistle, a median 109 minutes after the listed start in the Premier League, and RN1 keeps buying at the edges, 7.5% of maker shares at 0.90 or more and 10.5% under 0.10.
Shares bought at each price
Inferred T: when the game ends or the market closes (75%); at a price bound nearer 0 or 1 than its fills reach (15%); it pulls or widens late (10%). In 44.21% of markets the last maker fill was at 0.90 or more, or 0.10 or less, and in a quiet market a last fill may only mean the takers went home.
The Same Loop, Left Running Early
# The quoting loop, running before the start in some markets. if now < game_start(M) and early_ok(M): P inferred quote(M) # What a careful build would add before the start, and no receipt shows: G inferred # a wider quote, a smaller size, # a pull when team news or a line move lands.
Why some markets get quoted early is read from which they are. Only 5,558 markets had a pre-game maker fill, tennis is 30% of the pre-game maker spend, and 8,974 of the 21,962 pre-game maker fills came an hour or more early, too early for a clock error in football. Inferred P: on purpose, in a chosen set of leagues (45%); by accident, a wrong start time letting the in-game loop run early (35%); left-over or test orders (20%).
The guard is inferred from absence: a careful build would quote wider before kickoff, yet pre-game bids are one tick wide in 92.86% of the 84 same-block pairs that show them. Inferred G: no guard, the same quote and size as in-game (70%); a guard the receipts cannot show, such as a pull on team news, that is too weak (20%); a fine guard, and the loss the first piece showed is bad luck (10%). The loss sits in fills an hour or more before the start, and the daily payments do not buy it back. Those that follow pre-game maker fills come to 38,275 pUSD against 282,723 lost, and adding every payment tied to no strategy leaves it 92,763 down.
Footprints of an Order That Never Rests
on book_change(M) while live(M): for outcome in (A, B): ask = best_ask(M, outcome) if should_take(M, outcome, ask): W inferred n = take_size(M, outcome, ask) K inferred send(M, outcome, limit = ask, size = n, kind = FILL_NOW_OR_CANCEL) O inferred # fee = rate * n * ask * (1 - ask), rate 0.05 in most markets; # a daily payment returns half of it the next day before game_start(M): ref = outside_price(M, outcome) Q inferred ask = best_ask(M, outcome) if ref - ask > fee(ask) + threshold: send(M, outcome, limit = ask, size = n, kind = FILL_NOW_OR_CANCEL)
The take trigger is read from how the takes fared. Taker fills beat their price in 8 of 10 price bands after a fee that averages 1.63% of the pUSD paid, 2.5% at a price of 0.50, half of which a daily payment returns the next day. Where RN1 mostly took, pairs cost more than par, 1.009 and 1.029, so taking is not pair arbitrage, and 79.8% of taker fills came in a market where one of its bids had already filled, so it takes where it quotes. Inferred W: its fair price beats the ask by more than the fee, whatever it holds (50%); its holdings, buying the lighter outcome when that makes cheap pairs (25%); an event, a score or news it sees before the book moves (25%). A take on a model price and one on a fresh score look the same on chain.
Take size is read from the 108 taking orders of the sampled evening, which fall into two kinds. 58 were signed 5 seconds or less before they filled, a median 670 shares, 40 of them filled in full. The other 50 were older and far larger, a median 29,212 shares, of which a median 0.13% filled. Inferred K: round share sizes set by rule, small for the quick orders and tens of thousands for the rest (65%); the quick orders sized to what is on offer, rounded to 10 shares (35%).
Order type is read from the same orders, 82 of which filled exactly at their limit and none worse. The quick kind behaves like a fill-now order9. The older kind does not: five were signed 32.6 minutes before an 18:45 kickoff and filled at 19:47, and no field on chain holds an order's type. Inferred O: fill-now orders throughout, the older ones signed early and sent late (45%); fill-now for the quick ones and a resting order for the large, matched as the taking side (35%); ordinary limit orders throughout (20%). What the older orders are is open.
The pre-game take is read from breadth, with takes spread over 94 series and 1,984 markets, the breadth an odds feed gives. Inferred Q: an outside line, sportsbook odds with the margin taken out (50%); the in-game take rule, running early or on a wrong start time (30%); its own pre-game model (20%). It rests on 4,374 fills.
Where the Footprints Stop
Everything above rests on the orders somebody hit. The teardown lists what that leaves out.
The book itself cannot be had after the fact, because a cancelled order and an order nobody hit were never written to the chain. What can still be had is the signed orders for the whole window, which are already on chain and which turned size, levels and speed from guesses about fills into readings of what RN1 posted, so far only in samples. Of the missing inputs they matter most, and after them I would want every wallet's fills for a few busy hours, to settle whether RN1 sets the best price or joins it. Until then the code above is the smallest program that explains the receipts, which makes it a floor on what RN1 runs.
How this was rebuilt: every OrderFilled log on Polymarket's exchange where RN1 is maker or taker, 10 Aug to 10 Nov 2026 UTC, Polygon blocks 91,743,315 to 97,042,228; the wallet's pUSD transfers; and signed orders read from the calldata of fill transactions, in two samples, so figures that rest on them are preliminary. The percentages are a judgement of how strongly the evidence favours each build, not a fitted model. Figures are rounded. First piece: "RN1 Doesn't Need to Know Who Wins".