winning-wallet-finder_github/FINDINGS.md

Findings — what works and what doesn't on Polymarket

A research log of an honest attempt to find a systematic, automatable edge on Polymarket using public data. The short version: almost nothing works — the market is efficient — and the one thing that does isn't a money-printer, it's a detection signal.

The goal

Find a repeatable way to make money on Polymarket: identify "smart money" wallets, copy them, farm rewards, or arbitrage — anything systematic and automatable from public data.

Scorecard

Approach	Verdict	Why
Copy high-win-rate wallets	❌ dead	Win rate was an illusion (see below). True rates ~50%. Flat-size copying backtested −48% over a week.
Rank by leaderboard / PnL	❌ dead	Raw PnL is variance; top wallets win ~50% and profit via sizing/timing you can't copy.
LP reward farming	❌ dead	The fat "thin-book" APRs are illusory — Polymarket refunds unearned pool to the sponsor when liquidity is low. Real yield is modest and adverse-selection-dominated.
Binary YES+NO arbitrage	❌ dead	Efficient — min observed sum was 1.001 (the spread). Closed instantly by the engine.
Multi-outcome logical arb	❌ dead	True partitions priced efficiently (min sum 0.999). Apparent "arbs" were non-exclusive market groupings.
Cross-venue arb (Polymarket↔Kalshi)	❌ dead	Venues agree to ~1¢; locking both legs costs >$1 after fees. Real gaps last ~seconds and are taken by bots.
Insider / sharp detection	✅ works	Statistical improbability (z-score of wins vs. odds) is a real, hard-to-fake edge signal. See insider.py.
In-play surge momentum (tape)	❌ dead	The +$24-46/fill was resolution-timing survivorship (bias round 3, below). Chain truth: −$6/fill over 1,108 forward fills — pre-registered kill met 2026-07-22, three days before the funding decision. Wallet-identity null stands.
Sub-5¢ surge longshots	❌ dead	0-for-38 under chain truth; the exploratory scan's 4 "winners" were the same resolution bias.
Crypto oracle fair-value taking	⏳ forward window	Chain resolution grew the sample 30× and flipped the read: noisy-around-zero, slight positive tilt at E≥0.07 (+$1.8-2.3/fill, n≈470-860 fwd). E0.04 tier killed; stricter tiers accumulating — #17.

The big technical findings

1. Win rate on Polymarket is survivorship-biased — badly. The platform only redeems winning shares; losing shares are worth $0 and sit unredeemed in /positions at curPrice 0 forever, never entering /closed-positions. Measuring win rate over /closed-positions alone counts almost only winners. We saw a wallet read 90.6% that was truly 48.3%. A correct win rate must union both endpoints. Lesson: a high reported win rate is a red flag for a measurement bug, not a sharp.

2. Win rate ≠ profit, and PnL ≠ reliability. A wallet winning 54% made millions; the all-time #1 wallet (43% win) was −$3.8M over 90 days. Profit comes from sizing and entry prices, not hit rate.

3. The market is efficient. Six systematic public-data edges, all closed or illusory. There is no turnkey retail edge sitting in public data in 2026 — durable edge requires speed/infrastructure (arb bots), private information, or getting paid to provide liquidity (modest, adverse-selection-dominated).

4. The one real signal: statistical improbability (z-score). Each bet entered at price p has an odds-implied win probability p. A wallet winning far more than Σp is beating the market's own pricing — measured as a z-score and one-sided p-value. This is the rigorous version of the edge metric the whole project was chasing. It distinguishes:

• Sharps — high z, normal entry timing (skill over many bets).
• Insiders — high z + late (pre-resolution) entry + fresh wallet.

Plus funding-cluster linking (à la Bubblemaps / the 2026 60 Minutes investigation): trace each wallet's USDC funders on Polygon and link wallets that share a personal funding hub — judged by the funder's own outbound degree so shared exchanges don't false-link everyone. (See insider.py.)

Practical conclusion

• Don't fund copy-trading, LP farming, or arb based on this work — we tested them and they don't clear.
• Do use insider.py's z-score as a rigorous "who actually has edge" filter, far better than leaderboard or win rate.
• A genuine money-making edge has to come from you — a niche you understand better than the market — with tooling built around it, not from a public-data scanner.
• Legal note: detecting suspected insider trading is fine; trading on material nonpublic information is illegal, and blindly following a suspected insider is not a safe strategy.

Insider detection — what the z-score signal actually found

Building insider.py and sweeping markets (hunt.py, huntwide.py) surfaced genuinely improbable wallets. Out of ~289 scored:

• DREAMBIG. (z=8.9, p≈2e-19) and qcp14 (z=5.3) on the Iran ceasefire market — 45–77% of wins entered <24h before resolution. Textbook insider fingerprint, on exactly the theme the 60 Minutes story covered.
• Famecesgoal (z=9.6) won only 14.5% of bets — but bet longshots and hit +98 above what the odds implied. The clearest "beats the prices it pays" case.

Two refinements proved essential:

• Trade count separates insiders from bots. bjprolo scored z=37 — but on 306,873 lifetime trades. That's a market-maker grinding a tiny systematic edge, not information. Real edge wallets show concentrated z over 1–3k trades.
• Funding-cluster linking (Alchemy, the Bubblemaps "who-funded-whom" step) works only with a personal-hub filter: a shared exchange (everyone uses Coinbase) is not a shared operator. Judge a funder by its own outbound degree.

The copy-trade verdict — in-sample vs out-of-sample

The decisive test: does copying z-selected wallets make money?

• In-sample (copyback.py): copy the edge wallets from May 30, z-weighted, reinvest 100%. Result: +545% in 15 days. Looks incredible — and it's circular (the wallets were selected for winning over that very window). 86% of it came from one wallet; the highest-z pick contributed $23.
• Out-of-sample (oos.py): select wallets using only data through Apr 30, then copy forward May 30→now. Result: +168% — but entirely from one longshot lottery wallet (1.5% pre-period win rate hitting again). The two strongest pre-period signals made $0 forward. Forward hit rate was 27%. That's variance, not edge that persists.

Conclusion: even the one real signal (z-score), when tested for whether you can profit by copying it, fails out-of-sample — joining every other strategy. The detector is valuable for finding anomalous wallets; copying them is not a proven, fundable edge. The live watcher (webhook_receiver.py) exists to gather real forward (out-of-sample) data on these wallets — observe before you size up.

Practical conclusion 2

• Don't fund a copy strategy — both the +545% and +168% are variance/concentration, not repeatable edge.
• Do use the detector to find statistically anomalous wallets and watch them live; judge persistence forward with your own eyes.
• A durable trading edge has to come from you (a niche you know), with this tooling built around your judgment.

The skilled-3% scan, and a clean out-of-sample loss (June 2026)

External validation arrived: an LBS/Yale study (Gomez-Cram, Guo, Kung, Jensen, Apr 2026; SSRN 5910522) over 1.72M accounts found only ~3.14% of traders are genuinely skilled — measured by randomizing each trader's bet directions 10k× (a Monte-Carlo z-score) and requiring out-of-sample persistence. That is exactly this project's z-score + oos.py method, independently confirmed.

Built live/ to operationalize it at scale: enumerate recent liquid markets → cache every candidate's resolved bets locally (~26k wallets / 12.5M bets, so re-scoring at any cutoff is seconds) → a 5-gate funnel (n≥15, z>0, BH-FDR, split-half OOS, MM/bot cap). It surfaced 107 "validated" wallets.

The decisive test. Copying the high-win-rate "favorite-rider" cohort, $1000, no execution lag, June 1→now:

• selected through the test window (look-ahead): 99% win rate, +23.6%.
• selected on pre-June-1 data only (honest): 68% win rate, −7.4% (−19% on the settled portion).

The +23.6% was selection bias. Done cleanly, the favorites lose — a textbook reproduction of the paper's "~60% of lucky winners become losers out-of-sample," now on our own live data. Lesson reinforced: high win rate is the most misleading signal on the platform; favorite-riders are uncopyable. The underdog/value archetype (beats longshot prices) is the only one left worth testing.

Train/test wallet selection, and the capital wall (June 2026)

Built live/strategy.py (train on bets resolved before May 30, validate June 1+) and live/followability.py (entry-time + lead-time + cadence filter). Selecting on copy-ROI + z + monthly consistency + diversification (not win rate) gave 150 wallets; 59/100 stayed profitable forward (p=0.044), and filtering to followable markets lifted it to 49/77 (p=0.011), +23.4% pooled out-of- sample. So a real, persistent, copyable edge does exist — unlike favorites.

Then the reality check (live/pnl_basket.py, live/pnl_focused.py): a $1,000 copier with missed-trade accounting (capital tied in open positions).

• Broad 10-wallet basket: the wallets fire 1,210 markets in June; $1,000 can follow only ~2–13% of them. At realistic stakes it loses (−$384 to −$800); the gains sit in the trades you couldn't afford ($14k–$153k "missed"). Capital, not edge, is the binding constraint.
• Focused + conviction: copy only 1–2 top wallets and only their larger-stake (≥$200) bets → trade count drops to ~30–40, $1,000 affords them all, and it clears: +91% to +247% across stakes, stable, no blowup.

Lesson: a small-bankroll copier cannot follow a skilled wallet's whole feed — the edge is only capturable by concentrating on few wallets' high-conviction bets. The live tracker (jaxperro.com/trading) now runs exactly that config.

The repeatable profile: conviction bets + a timing gate (the best result)

Refining the above: instead of all bets, score wallets on their HIGH-CONVICTION bets only — the top 20% of each wallet's own stake sizes (per-wallet p80). This replaced the original flat >= $200 cutoff (2026-06-22): p80 reproduces flat-$200's win-rate lift across the sharps while adapting to each wallet's scale — a whale's $200 bet isn't conviction, a minnow's is. The top wallets win 70-80% of their big bets on genuinely-uncertain (~0.4-0.6 priced) markets — real edge, not favorite-riding — and it persists. live/conviction_scan.py (train pre-June, validate June) under p80 finds 218 wallets matching the profile; forward, 62/83 stayed profitable (p≈0), +16.0% pooled. A reproducible class, not a fluke. (The original flat-$200 run found 69 wallets, 25/37 forward, +11.7%.)

Then the decisive copyability filter, live/validate_timing.py: a near-100% win rate is only useful if we can mirror it. The tell is entry->resolution lead time on winning conviction bets — this is a copyability heuristic, NOT proof of inside information (a short lead can be a genuine insider or just someone good at fast-resolving markets; we can't tell, and for copy purposes it doesn't matter). Of the 218, the gate drops the "last-minute" wallets (median lead <24h — you can't get the trade in before resolution), then a 30-day-active filter, leaving ~31 validated copyable sharps (watch_sharps.json) with multi-day leads. The standout 0x60ec1744… held 80% win over 1,017 forward conviction bets; even the suspiciously-perfect 0x72e1… (99/100% win) enters ~7 days early — a real forecaster, clearly not last-minute. These 50 are surfaced live on jaxperro.com/trading.

Lesson: score conviction bets, not all bets; require avg entry ~0.4-0.6 (edge, not favorites); and gate on lead time to drop last-minute (un-mirrorable) wallets. That funnel produces a copyable, forward-validated set — the strongest evidence in this project that followable skill exists.

Copy P&L: position win% ≠ copyability (the scalper trap, 2026-06-23)

The biggest caveat on the whole sharps table: a high conviction win% does not mean you can profit copying the wallet. The win%/record are computed from curPrice ≥ 0.5 on resolved positions — a position snapshot. For a high-frequency scalper that massively over-counts: he buys ~$0.50, sells seconds later for ~+$1, and the snapshot records a "win" even though he never held to resolution. ArbTraderRookie shows ~100% conviction win (398-2) yet a flat-$50 copy of his conviction bets, held to authoritative clob resolution (winner by token_id), nets −$790 (held 0-19) — two independent replays agree (live portfolio −$687 ≈ standalone clob −$790).

So validate_timing.display_stats now also computes copy_pnl — what a flat-$50 copier actually realizes since June 1: replay their conviction entries, mirror their exits, settle held bets at clob resolution. This is surfaced as the Copy P&L column on the dashboard (default sort). The verdict it delivers: most "sharps" lose when copied. Of the ~31, only a handful are copy-positive — Kruto2027 +$1,184 and fortuneking +$430 (true hold-to-resolution betters); names that looked great on win% (iohihoo 88.7% → −$749, ArbTrader 99.5% → −$790) are scalpers that bleed. The live tracker now follows fortuneking + Kruto2027 — the two copy-positive wallets — at $50/trade.

Lesson: judge a copy target by Copy P&L (a trade-replay with real resolution), never by position win%. Conviction must be measured at the position level (a wallet's total stake in a market), not per individual buy — a scalper splits one position across many small buys, so a per-trade threshold copies far more (and worse) bets than intended.

Capital recycling & the $1k book (2026-06-23)

The $1,000 paper book (live/portfolio.py → portfolio.json, rendered at jaxperro.com/trading) surfaced two things:

• "Saturation" was mostly a measurement artifact. The old browser replay froze capital in positions whose resolution date the data-api didn't return, so it skipped bets it could afford (340 phantom misses on a 4-wallet book). Computing the book off the cache — which stores each bet's resolution time (res_t) — frees cash at the true resolution moment: misses dropped to ~0 and the book recycled ~23× over the window. With real money this isn't even a problem (cash returns on redemption); it was purely the paper sim mis-measuring.
• More wallets help only up to the bankroll's slot count. A combo backtest over the copy-positive holders showed returns rise with basket size until peak concurrent demand exceeds ~$1k ÷ $50 = 20 slots, after which a high-volume wallet just crowds out the others. So: pick wallets that fit the bankroll, favor fast-resolving markets (capital velocity > bet size on $1k), and don't diversify past what you can fund. Two well-chosen holders beat four that overflow.

The holder blind spot: two data bugs that hid the best copy targets (2026-07-04)

A clean re-run of the May→June train/test on a trusted subset of the cache overturned two earlier verdicts. Two mechanisms were poisoning the data:

1. The res_t = ts fallback. When the data-api omits endDate on a closed position, insider.resolved_bets stores the wallet's sell time as res_t and won = curPrice >= 0.5 at pull time — so a scalper's sold-at-profit position looks like a resolved win at a fake resolution time. ArbTraderRookie's 1,997 cached rows were 100% this. Fix: live/trust.py — only trust rows whose res_t matches the market's modal res_t across ≥2 wallets (endDate rows agree; sell-time rows scatter), market over, wallet pulled after resolution, resolved not False. 13.5M of 19.2M rows pass.
2. The held-leg window bug. validate_timing's Jun-1→now replay only counted held bets entered and resolved inside the window; a ~7-day-lead holder always showed held 0-0, ~20 unresolved and failed held_n>=8. And before the 2026-07-02 winner=False settle fix, those unresolved held bets were booked as losses — which is exactly where the "scalper trap" numbers for iohihoo (−$749) and ArbTrader (−$790) came from. Those two verdicts were bug artifacts, not scalper traps. (ArbTrader still deserved rejection pre-fix — his cache stats were res_t=ts poison — but his real trade record was a ~160h-lead holder.)

What the clean test found (select on ≤May trusted rows only, validate on June, fees+slip): population baseline −1.4%/bet; the existing profile +8.7% pooled; adding a whole-book z gate (z_all > 2) roughly doubles it at every tier; a practical top-basket (also gated on med conviction stake ≥ $50 and holder/borderline lead verdicts) went +80% pooled, 7/7 wallets profitable. A capital-aware $1k replay of the 8-wallet pre-June basket did +504% in June (118 bets, 97W-21L, fees+slip, 53 missed for cash) — with the three informed holders a combined 62-0 and two basket members losing money (toosmart 4-12), so the selection is good, not magic.

Where the edge lives: the top holders (Stavenson, whale 0x73afc816… with $20–120k clips, iohihoo; ArbTraderRookie until 2026-07-03) bet low-tier tennis (ITF/qualifiers/Wimbledon doubles) and tier-3 esports (CCT CS, Dota 2 EPL) at ~0.5 entries, win 95–100%, enter ~160h before resolution, and hold. That's informed money — plausibly match-fixing-adjacent — which is copyable precisely because of the long lead. The regime risk is real and demonstrated: ArbTraderRookie was wiped from every data-api endpoint mid-analysis on 2026-07-03. Treat every month of this edge as possibly its last; re-select weekly; never size as if the 100% win rates are permanent.

Lesson: selection metrics are only as honest as the rows they read. Gate on trusted rows, judge held edges on windows longer than the wallet's lead time, and add z_all — skill must show in the whole book, not just the big bets.

Making P&L equal reality — the survivorship correction, finished (2026-07-08)

The sharps table's All-Time P&L had been a won × entry × size reconstruction, and decomposing it against each wallet's Polymarket profile (lb-api /profit, the PM P&L column) exposed it diverging by up to 10× — and flipping signs. Four distinct bugs, each earned by decomposing an outlier:

1. A 2,000-row pull cap (max_pages=40) truncated high-volume wallets — ewww1's 4,088 positions ($409k) showed as 740 ($40k).
2. Both-sides positions double-dropped — one-per-market dedup kept the winning leg and silently dropped the paired losing leg (suraxy: +$35k of phantom profit).
3. initialValue = 0 on big longshot winners mis-sized the reconstruction.
4. Corrupt near-epoch res_t rows polluted the sums.

The fix that killed all four at once: stop reconstructing, and sum Polymarket's own realizedPnl per closed position over the wallet's full history (cache.closed_exits, incremental). This is the wallet's realized track record — what a copier mirroring their buy/sell/hold actually banks — and it needs no size/entry/timestamp, so both-sides, iv=0, and bad res_t all become moot, and it sums to PM by construction.

Then the deeper one — the founding survivorship bias, live inside PM itself. A residual gap remained: PM /profit is itself survivorship-biased. Bets that lost, went to $0, and were never redeemed sit in /positions at curPrice 0 — real losses, but PM under-counts them unevenly (it subtracts Coteykens' $52k of abandoned losers to land at $14k = our number, but does not subtract oliman2's $161k, leaving PM at $112k against a true $20k). So _open_split now folds those decided-but-unredeemed positions into the realized total, leaving only genuinely in-flight positions in a new Open P&L column. The result: where our All-Time reads below PM, PM is the biased number and ours is the truth. oliman2 $181k → $20k, JuiceFarm $380k → $32k — wallets that looked elite on redeemed-only P&L are mediocre once you count the bets they walked away from. Full-list check: 27 of 31 sharps match PM within a few percent, 4 are honestly-lower, and — the correctness signal — zero over-count.

Lesson: a wallet's redeemed P&L flatters "sell your winners, abandon your losers." The honest record counts the abandoned losers; the profile number doesn't. Rank on realized-including-abandoned, and read the open book as a separate risk.

Choosing the month's follow set from corrected data (Set D, 2026-07-08)

With P&L finally honest, the follow set was rebuilt by simulation. Ranking on the signals that actually predict forward copy profit — 2-month Copy P&L (validated outside the backtest window), 30-day conviction form, copyable lead (no sub-hour snipers), a clean open book, and moderate bet size (a $1k book can't follow a $3k-clip wallet) — then backtesting candidate portfolios:

set	equity (30d, $1k)	W–L	missed
old live set	$15,359	250–79	0
5 high-lead big-bettors	$3,661	44–18	144 (capital-capped)
Set D (6 moderate-bet)	$27,799	279–75	0

Set D = LSB1, imwalkinghere, Kruto2027, 42021, 0xbadaf319, BikesAreTheBikes — the sweet spot where return, a 79% decided win rate, full capture (0 missed), and diversification all peak. Two rules fell out: moderate-bet wallets beat big-bettors on a small book (the whales get capital-capped, 144 missed), and imwalkinghere + LSB1 are irreplaceable (dropping them halves the return). The backtest is in-sample (a ceiling), but every Set D wallet also clears the out-of-window Copy P&L signal — that's what separates it from curve-fitting. The live July book remains the only out-of-sample truth.

Superseded the same day by Set E, after the alignment audit below found the replay itself was still dropping and mislabeling bets.

Aligning the three books — backtest ↔ bot ↔ Polymarket (2026-07-08)

The live bot showed Kruto2027 mirror-sells the backtest didn't have. Pulling that thread found the replay was silently losing real bets three ways:

1. Stale entry maps — the daily freshen reset the bets and exits cursors but never pulled_entries (14-day TTL), so any market a wallet first entered since the last entries pull had no first_buy and the replay dropped its bets entirely (if not et: continue) — not won, not lost, not open. Gone. The recovered bets included hidden LOSSES — the stale backtest was flattering.
2. res_t can't detect in-play sells — it's endDate metadata (game-day midnight, sometimes future), so the exit < res_t − 300 test never fired on in-play markets and every pre-resolution sell booked as held-to-resolution. Fixed with the price test: a redeem prints exactly the payout; a mid print is a sell (booked at the wallet's actual exit price — which also self-corrects poisoned won flags).
3. Resolved round trips vanished — the round-trip path skipped resolved-on-chain markets assuming "the cache row will cover it", but rows with bogus forward res_t never qualify. Now redeem-closes book at chain truth and sell-closes mirror the sell.

Plus one parity fix: Set wallets now replay on the bot's pinned floors (copybot.paper.json), not a recomputed p80 that drifts a few percent and takes different bets.

Proof of alignment: matching every settled bet in the bot's real book against the backtest row-by-row — 7/9 agree exactly, 0 absent (was 0/9 agree before the fixes). The 2 disagreements are genuine execution divergence (the sharp sold on a fast-resolving market, the bot couldn't catch the exit and rode to resolution) — each book correctly records what happened to it, and that divergence class is permanent.

The honest price: the 30d Set D replay fell $29.1k → $17.4k as the flattering artifacts (phantom held-to-$1.00 winners, missing salvage exits, hidden losses) came out.

Set E (deployed 2026-07-08): with the replay finally honest, all 35 sharps were re-ranked by individual 30d copy replay, and combined sets tested in a shared book (one position per market, shared cash):

set	equity (30d, $1k)	note
top-4 only	$12,777	pruning alone loses carry
Set D (control)	$17,362
Set E (7)	$24,378 (+2338%)	every member positive
Set E + lma0o0o0o	$24,437	carries a −$1,775 wallet — rejected

Set E = LSB1, imwalkinghere, Kruto2027, 0xbadaf319 + gkmgkldfmg, AIcAIc, 1kto1m. Dropped: 42021 (+16% on 22 bets), BikesAreTheBikes (+12%). Rejected on the audit evidence: oliman2 (true lifetime ~$19k, not PM's $112k; +21% to copy with 22 stuck-open) and leegunner (elite lifetime +$274k but negative to copy — 7.6-day holds kill compounding). Newcomer caveat: gkmgkldfmg (z=2.05) and 1kto1m (z=2.4) sit near the selection gate floor — their seats are earned on a strong month, not deep statistical edge; AIcAIc's held-win is only 42% (his profit is sell-timing, the most lag-fragile edge class). They are the demotion watch-list, in that order.

The refund harvesters — a new sharp archetype (2026-07-08)

Splitting SOLD out of the record columns (same W/L/R/S taxonomy as the bot and backtest; win% is now held-outcomes only) exposed something the sign-based tally had been calling "wins": two of the highest-z wallets in the list are refund-harvesting machines. The signature is exits at exactly $0.50 to float precision — only refund redemptions print there — at enormous scale:

• 0xb0E43B… (z=20.2, "94.4% all-time" under the old tally): 797 exact-0.5 redeems = $148k of his $218k lifetime P&L. True held record: 76W–31L.
• ArbTraderRookie (z=29, "99%"): 1,150 of the same.

The strategy: buy ITF tennis totals just under 50¢, harvest the chronic ITF cancellation/retirement rate (50/50 refunds pay $0.50/share). The edge is real, clever, and structurally uncopyable — it clears 1–2¢/share and a taker copy pays ~0.75¢ fee each way plus slippage into it (honest replay: +6% and −5% respectively). This closes the loop on the project's oldest lesson: win rate lies, z finds real skill, and only the fee-and-lag replay decides whether the skill transfers to a follower.

Also fixed in the same pass: _open_split now classifies decided-unredeemed positions by the data-api's redeemable flag (exact, set at on-chain resolution for winners and losers alike) instead of price-pinning — verified by reproducing PM's per-position books to the dollar on the two biggest All-Time-vs-PM divergences.

The calibration experiment (started 2026-07-08)

Everything above makes the accounting honest. It does not make the +2338% forecast honest: Set E is an in-sample maximum (ranked and assembled on the same 30-day window it's scored on — winner's curse applies), and the replay's fill model (their price +0.5%, always filled) ignores adverse selection — the market moves fastest on exactly the bets where the sharp knew something, and thin ITF books won't always fill a FAK copy at size. The one piece of measured ground truth — the old paper book's +$229.79 (~23%) over two buggy weeks vs. four-figure replay percentages for the same era — says the live-to-model discount is large.

So the paper book was reset to a fresh $1,000 on 2026-07-08 (old book archived in git history + archive/copybot_fills.pre-reset-2026-07-08.jsonl) running Set E with every fix live from day one. The measured ratio between this book and the published backtest over the coming weeks is the number that sizes real money — not the replay percentage. Bank-size note for that decision: the replay compounds faster on smaller banks (--bank 500 → +2728% vs $1k's +2103%) because 4%-of-equity stakes hit the never-bigger-than- their-bet ceilings later — percentage returns from small books are the most optimistic view, discount accordingly.

The tape era opens: first RTDS findings (2026-07-20)

CORRECTED 2026-07-22: the Study-A/B EV figures in this section were inflated by resolution-timing survivorship (see the next section). The structural findings — sharp screen, identity-null, crater physics — survive; the surge P&L numbers do not.

Three days of the recorder's firehose in live/rtds.duckdb (13.8M fills, 160k wallets, every fill including the losers) killed the survivorship problem at the source and produced four findings in one day — details, code, and frozen parameters in research/ (silo'd from the bots; see its README):

1. The tape sharp screen works, and its resolution proxy is exact. Terminal-VWAP convergence (≥0.97/≤0.03, 2h quiet, sibling veto) agreed with on-chain CTF payout vectors on 742 of 742 validated bets. First run (live/tape_sharps.py): 2,360 wallets with ≥8 resolved held bets, 25 copyable candidates at z 4.0–5.5 (discrete conviction entries, sports/ esports in-play profile — the Set-E archetype), cleanly separated from an uncopyable algo-flow tier (z 9–12, 10–37 fills/bet, $M volumes: the crater-sweepers). Benchmark sanity: benched sharps scored positive, benched disappointments negative.

2. Wallet identity might not matter (the null that redirects the program). Study A (#16): a $300/60s net-flow surge into an in-play sports/esports market at 0.10–0.90 simulates to ~+$24/$100-fill after fees at our measured execution — but 10 activity-matched random wallet sets produce the SAME EV (+23.85 pooled vs +23.68 informed). The herd's lean is the signal; who leans adds nothing so far. Hypothesis revised at freeze: surge momentum primary, identity lift secondary. Verdict comes only from the forward ledger (research nightly, pre-registered thresholds).

3. The oracle edge is real on paper and mostly untakeable — worse at size. Study B (#17): fair value from the venue's own settlement tick feed flags ~9k mispricings/21h, but 86% die as FAK craters (crypto makers requote <4s), and the fillable remainder shows winner's-curse inversion — demanding a 10¢ edge loses $31/fill. Nothing froze (no cell hit 30 fills); candidate v2 uses are inverted: copy filter / maker-side.

4. Craters refill on a clock, and it's niche-shaped. 775k crater prints: crypto refills within 4s 94% of the time, esports 83% by 10s, sports needs ~25s, geo/politics tails run minutes. Shipped straight into the bots as per-niche fak_retry waits (the flat 10s was calibrated to nothing).

Execution realism note for everything above: the research simulator is fitted on the live bot's own 29 labeled attempts (fills + FAK misses) and carries a measured −2¢/fill optimism bias — every pre-registered pass threshold sits at least 2× above it.

Survivorship bias, round three: resolution timing kills Study A (2026-07-22)

The forward ledger said the surge signal earned +$35–46/fill. The truth was −$6/fill. The gap was a third, subtler survivorship mechanism — and the way it was caught is the real finding.

The mechanism. The ledger graded fills with the tape's proxy-resolution and skipped "pending" markets. But tape resolution is win-biased: when our side wins, the losing sibling goes quiet and the market resolves into the ledger within hours. When our side loses, the winning sibling keeps printing at 99¢ until close — the sibling-veto keeps the market "alive" and the loss hides in the ignored pending bucket. Wins scored same-day; losses waited. Jul-21 audit: tape-resolved fills hit 81% (+$46/fill); the 329 "pending" fills, chain-resolved on the spot, hit 26% (−$49.61/fill). Combined: 53%, ≈ breakeven-negative.

How it was caught: an independent instrument refused to agree. The surge paper harness (wwf-surgebot — real-time, $5 stakes, graded nightly against CTF payout vectors from day one) read 57.5% while the ledger read 73–81%. Decomposition ruled out selection (same-trigger cohorts hit alike) and entry prices (~1¢ effect); what remained was the scorer itself. The same pattern as rounds one (unredeemed losers) and two (res_t poison): a too-good number met a measurement that couldn't be sweet-talked.

The fix (research/forward.py payouts_for(), d34a4c5): chain-truth overlay is now mandatory for every arm — tape proxy first, CTF payout vectors for the remainder, refunds as scratches. Full recompute:

arm	corrected verdict
Study A surge (#16)	KILL met: −$6.03/fill over 1,108 forward fills; even the fit day negative. Identity-null unchanged (controls deflated equally).
sub-5¢ longshots	0-for-38 — the scan's "winners" were the same bias.
Study B oracle (#17)	Sample ×30 under chain resolution (crypto sprints were veto-stuck). Now noisy-zero with a positive tilt at E≥0.07; the 4¢ tier killed, stricter tiers accumulating.

What the pre-registration bought. The kill line was written 2026-07-20, before any forward data; when the corrected numbers crossed it there was nothing to argue about and no money at risk — the $100 deployment plan (#19) was three days from its funding gate. Total cost of the false edge: $0 real, ~2 days of compute, one paper book that ended at $99-and-change.

Standing rules extracted: (1) any scorer that can say "pending" must prove pending is outcome-neutral, or chain-resolve it — on Polymarket it never is neutral, because market liveness itself encodes the outcome; (2) every study needs one instrument that doesn't share the scorer's assumptions (the paper harness earns its keep even when — especially when — it disagrees); (3) at 20–50× payoffs, no small sample means anything: the sub-5¢ "edge" survived a 31-fill scan and died at 38.

The measurement split + the data moat (2026-07-22, same night)

Round three's kill landed while a parallel line of work was rebuilding the instruments; the two converged into an architecture change and one more corrected verdict.

One instrument can't sample and rehearse at once. The v1 surge book halted at its pre-registered −50% line the same evening. First diagnosis blamed the cash gate for adverse selection — the ~2% subsample it could afford read −$9/fill vs "+$41/fill" for the full signal. Round three dissolved that comparator (the +$41 was the resolved-only bias; chain truth −$6/fill, right next to the subsample's −$9): the book wasn't sampling badly, the signal was dead. The structural lesson survives the autopsy either way: a physically cash-gated harness measures a bankroll's shadow, not a signal. Both harnesses now run measurement-mode — every trigger attempted at the ledger's stake, and bankroll specs replayed OFFLINE over the recorded attempt stream (surge_book_replay.py), where gating can never touch the sample it's judged on.

No scalp was hiding in the corpse. The markout-exit study (chain-true; its res_tok v0 self-corrected within hours — same round-three mechanism, caught by the same law): 1,146 forward fills, hold −$5.92/fill, and every exit horizon negative too (best −$4.18 at +30m). The cohort split is the microstructure finding: winners drift monotonically to +$44 held; the losses the old scorer hid bleed from minute one (−$6@60s → −$44@2h). Surge moments are symmetric information events, priced in immediately — after fees and worst-print entry there is nothing for a taker at ANY horizon. Whatever earns money at those moments is on the other side of the book (maker study: staged, tape-sim first).

Two more scorer artifacts, disclosed on #17 before they could bite: the tape scorer reads sprint strikes from the FUTURE window-open tick for pre-window prints (live gates ts ≥ t₀; 404 such prints suppressed in an 8.5-min shakedown), and unlabeled non-sprint tokens default to the yes-side, turning No-side dust prints into fake ~+0.99-edge events (the harness reads labels from the trade payload itself — same rule, better input).

The data moat, hardened (DATA LAW, research/README). Everything accrues Mac-independent on Fly volumes (recorder: 25GB at 4%, ~3+ weeks of offline headroom; harness streams: attempts with top-5 asks + top-3 bids + latency, markout book re-reads at +60/300/1800s per fill, durable settle logs immune to state trims; daily volume snapshots ×5). The Mac is only the grader/archiver, and its two offline failure modes are closed: forward.py now backfills any tape-covered day the ledger has never seen, and settle logs mean trims can never rotate evidence away. meta_snap.py snapshots every active market nightly (~12k: end dates, tags, token→outcome maps) — τ becomes knowable AT TRIGGER for every tape trigger, and the label-gap artifact class dies at the source. Tests are additive, never subtractive: raw streams are append-only, analysis reads only, and a study that needs different data records a NEW stream.

The fill model is the next scorer (2026-07-23)

The first full chain grades of both measurement harnesses landed together and told one story from two directions.

Study A, confirmed dead twice. The A2 arm's 1,344 chain-graded settles read −$7.54/fill — within $2 of the corrected ledger, at real execution, on the full trigger stream. Meanwhile the offline replay of the v1 $100/5% bankroll spec over the same attempts stream sat at +26% — its cash gate picked 70 of 2,692 attempts and got lucky (29W/17L). A book trajectory at n=46 carries no information a measurement at n=1,344 doesn't overrule; that asymmetry is why bankroll specs lost verdict power in the A2 redesign, filed here as the standing example.

Study B, the divergence that matters. The oracle harness's first chain grade: E≥0.07 −$8.51/fill (n=235) · E≥0.10 −$5.90 (n=127) · ~40% hit — while the forward ledger holds those tiers slightly positive. Both now score through payouts_for, so this is not round three again — it is the same LESSON one layer down: after the resolution model, the fill model is the next place a sim quietly manufactures edge. The ledger's 6.7s-lag/3s-hold worst-print fills include entries no real order could have gotten (the book repriced) and price the rest against prints instead of the adversely-selected residue a live taker actually receives. The capturability instrument existed precisely to veto this; it has. Taker arm: evidence-dead (formal tier bars keep accruing, per pre-registration).

Five tandem tests, one day, all on already-flowing data (scripts in research/, each rerunnable): the copy edge is REAL but concentrated (esports ~all of it; the live book's drag sits exactly there — the #14 tension quantified); resting-bid copy entries beat taker FAKs +$17.45 vs +$12.86 per $100 signal at 93% 60s trade-through (→ #20, pre-registered, dark-deployed); mirrored sharp exits are bankroll ops, not signal — no wallet's exits added value, 0xbadaf319's were 99% wrong (→ #21, same); the orders_matched stream holds a 673-wallet maker-sharp species (+$5.9M pooled, ~33 expected by chance, 86% invisible to every taker screen) — not copyable by taking, the inventory-lean signal is the follow-on; and the sibling-sum scan showed print-substrate "arbs" are resolution-dust artifacts — standing-book data or nothing. The through-line of the week, now written as one sentence: every taker edge died at the requote wall, and every instrument that measured the wall found the makers standing on top of it.

Repo layout

• insider.py — the detector: z-score/p-value, timing/freshness/sizing signals, Alchemy funding-cluster ring detection.
• hunt.py / huntwide.py — market sweeps that surface edge wallets.
• copyback.py / oos.py — in-sample and out-of-sample copy-trade backtests.
• webhook_receiver.py — push-based live trade watcher (Alchemy → Discord).
• smart_money.py — data foundation + dashboard (true-win-rate scanner).
• live/ — current system: cache-backed finder + copy-positive-holder sharps selection (conviction_scan.py + validate_timing.py → watch_sharps.json, ranked by Copy P&L) + $1k paper book (portfolio.py → portfolio.json) + daily refresh. The dashboard (jaxperro repo) renders those two JSON feeds. See live/README.md. Copy execution (copybot.py, sync_floors.py) is a separate, in-progress system — this finder is selection + tracking only.
• research/ — the tape-era edge factory (SILO'd from the bots): read-only RTDS loaders, execution sim calibrated on the live ledger, pre-registered studies (#16 surge momentum, #17 oracle fair value), nightly forward ledger. Verdicts come from research/forward_ledger.jsonl only. See research/README.md.
• wide/ — bulk subgraph→DuckDB scanner (survivorship-bias-free, all wallets); public subgraph frozen at Jan 2026, so historical-only. See wide/README.md.
• archive/ — the strategies that didn't work, kept for reference. See archive/README.md.