Shelfware Labs
SW-03·Technical teardownDeep dive

I Tried to Copy Every Winner — Technical Teardown

The receipts: the beat-the-line z-scores, the churn, the full candidate ranking. Every wallet sits behind a stable alias (W-01…), consistent across every table, so you can track a specific one all the way through — without a single real handle or address appearing.

TL;DR
  • I ranked every leaderboard candidate by beat-the-line z — did it win more than the prices it paid implied — not by dollars won.
  • z has two false-positive modes: market-makers (enormous z, ≈zero profit) and beats-the-line-but-loses (positive z, negative ROI). A copyable edge needs all three of z ≳ 1.8, positive ROI, and liquid markets.
  • Exactly one wallet cleared all three — and 93% of its edge was a single World Cup. Copyable while the Cup ran; gone the week after.
  • The mega-whales ($2–5M "profit") scored z ≤ 0 — size × variance, no edge. Leaderboard rank barely persists period-to-period: corr ≈ −0.11.
  • Net accounts worth copying forward: 0.
on the aliases

Every account below is a stable pseudonym — W-01 through W-10 for individual wallets, Fund-X for the staking source. The mapping is fixed across all tables, so W-08 is the same wallet everywhere you see it. No real handle, display name, or 0x… address appears anywhere on this page. The metrics are the real ones.

1 · The right ruler: beat-the-line z, not dollars

Dollars won measure bankroll and variance as much as skill. The honest ruler is whether a wallet won more often than the prices it paid said it should. For a set of bets at entry prices pi (each price is the market's implied probability), the expected number of winners is Σ pi, and:

z = ( actual_wins − Σ pi ) / √( Σ pi(1 − pi) )

A wallet that merely rides favorites, or lands one giant lucky bet, gets no credit — z is significance-weighted and immune to a single whale ticket. Positive z = it beat the closing prices; higher = more certainly. But z has two ways of lying, so it's necessary, not sufficient.

2 · Why z alone is a trap — the two false positives

the two wallets that look elite by z and are un-copyable
Aliasbeat-line zwin ratenet P&Lwhat it actually is
W-0812.485%+$85A market-maker: quotes both sides ~a cent apart and skims the spread thousands of times. Wins almost every "bet" (avg price 0.56) → colossal z. Total lifetime profit: eighty-five dollars.
W-093.836%−$5,328Beats the line and still loses. Genuinely buys longshots below fair (wins 36% of bets priced ~30%) → positive z — but sizing and the spread paid on entry eat the per-bet edge. Real skill, negative money.
Bar to be realz ≳ 1.8 · positive ROI · liquidW-08 fails ROI; W-09 fails ROI. Neither is copyable.

This is the whole reason dollars and raw z both mislead. You need positive z (real edge over the line), positive ROI (the edge survives fees and sizing), and liquidity (you can actually take the price). Almost nothing clears all three.

3 · The candidate scorecard

The wallets that looked most like genuine, independent sharps, ranked and dissected. (An early per-wallet P&L pass — reconstructed off the raw trade tape — never reconciled against the authoritative payout feed and was replaced by it. So this table leans on behavior, sample size, and liquidity, and quotes hard dollar figures only where the reconciled numbers exist — rows W-10, W-08, W-09.)

genuine-looking directional candidates
Aliasarchetypewin%bets / spanthe tell that classified itcopyable?
W-01moneyline sharp57%167 / 65d99% of its edge still there at +5 min — real, takeable✅→⚠ 65-day sample, unproven
W-02totals sharp58%417 / 114dtextbook-sharp on the surface (76% <3h pre-game, 78% solo) — but luck-o-meter 1.4, and stripping its one opening streak drops it to 0.6: the record is the streak✅→✖ a hot streak
W-03volume grinder55%1,237 / 49d~69 bets/day, $75M turnover — thin edge, enormous sample✅→⚠ bot-scale only
W-04hot streak50%173 / 25d100% World Cup, only 19% independent picks⚠ one hot run
W-05schedule-bettor55%552 / 11donly 11% of bets are game-timed — bets a clock, not late info⚠ 11-day, unproven
W-06fade40%989 / 488dlongest history, but wins only 34% of its disagreements with peers✖ contrarian-losing
W-07raw-best65%best pure numbers (86% independent) — but trades illiquid elections✖ un-followable
W-10the closest one61%59z 2.40, +30% ROI (+$1.36M) — but see §5✅→✖ transient

4 · The behavioral microscope — how each was classified

Four tells did most of the sorting, straight off the public trade tape. Read the columns as: what a real directional bettor looks like, versus a market-making bot, versus a wallet that just bets a schedule.

the four tells (and what each archetype scores)
Signaldirectional bettormarket-maker / botschedule-bettor
turns per market (trades ÷ distinct markets)1–530–400low
sleep gap (quietest 6h as % of activity)a real dead zone (human sleeps)none — trades 24/7varies
near-game % (bets placed <3h pre-start)60–100%~11%
solo % (independent picks vs the crowd)high (61–86%)low (~19%)

So W-02 reads as a human sharp (~1 turn/market, sleeps, 76% near-game, 78% solo); W-08 reads as a bot (30–400 turns, no sleep gap); W-05 reads as a schedule-bettor (11% game-timed); W-04 as a crowd-follower on a hot run (19% solo, 100% one tournament).

5 · The closest thing to a real edge — and its ceiling

W-10 was the only wallet to clear all three gates: beat-line z 2.40, +30% ROI (+$1.36M), and its bets sat in liquid markets (55 of 59 in books ≥ $100k) you could actually copy. As close to a real, take-it-yourself edge as the whole board got — though a z of 2.40 over one short tournament sits in the gray zone between genuine skill and an unusually clean hot streak; the score you'd fully trust is nearer 3 (see §1).

Then I split its profit by event. ~93% of it came from the 2026 World Cup (z 2.4 over 41 Cup bets). Its tennis bets were negative (z −0.3); club soccer was flat. Every month it looked "consistent" was simply another month of the World Cup. The edge was real and copyable — while the tournament ran. The week the Cup ended, the markets it was sharp in stopped existing, and so did the edge. A real signal with an expiry date, not a repeatable engine.

6 · The mega-whales: dollars ≠ skill

The accounts with the biggest raw profit ($2–5M) — the ones a dollar-ranked leaderboard puts on top — scored z ≤ 0. They bet favorites and volume at size, and finished ahead on variance plus the same World Cup. No edge over the line at all; just a fat bankroll meeting a good run. And it doesn't stick: comparing leaderboard rank in one window to the next, the correlation is ≈ −0.11 — essentially a fresh coin-flip each period.

7 · The syndicate — a footnote, with the graph

One structural curiosity, because it's too good to omit. Tracing on-chain funding, the wallets that looked like independent stars were staked from a single source:

the funding topology (on-chain)
Noderolefans out to
Fund-Xsingle funder~45 betting wallets
of those ~45surfaced on the leaderboard~8
the stablemixedhuman sharps + market-making bots + no-edge wallets

In-and-out amounts reconcile to the penny, so the P&L is the fund's, and the "diverse field of self-made talent" is one desk wearing eight name tags. It changes nothing about copyability — you can still mirror any single wallet you like; it just means the apparent crowd of geniuses was, in part, one backer. (It earns its own writeup further down the shelf.)

The bar, and who cleared it
📐 the testbeat-line z ≳ 1.8 · positive ROI · liquid markets — all three.
🔬 cleared itOne wallet (W-10). Then 93% of its edge turned out to be a single tournament — copyable during, gone after.
reproduce it
  1. Pull each wallet's full history from the reconciled P&L feed + the activity tape (never the raw /trades endpoint — it omits redemptions and won't reconcile; that's what threw off the early numbers).
  2. Compute beat-line z = (wins − Σp) / √Σp(1−p) over resolved bets at entry price p.
  3. Compute the four tells: turns/market (trades ÷ distinct markets), sleep gap (hour-of-day histogram), near-game % (<3h pre-start), solo % (independent vs. crowd).
  4. Gate on z ≳ 1.8 AND ROI > 0 AND market volume ≥ $100k. Then split any survivor's profit by event — that's where the World-Cup concentration shows up.
  5. Trace funding on-chain (USDC transfers, chainid 137) to see whether your "independent" survivors share a source.

Prefer the story to the spreadsheet? The plain-English version is the same audit without the tables.

SHELVEDcause: luck, bots & fading edges
"One wallet cleared the bar. Ninety-three percent of it was one World Cup."