Reconstruction workpaperJuly 2026 cyclespec recon-error-v1 (frozen 2026-07-15)

Every number, including the ones that make us look bad.

We reconstructed Polymarket's top-25 profit leaderboard from public on-chain fill data. We could only score 8 of them. This page shows the per-wallet errors, the 15 we refused with their verbatim rationales, and the pooled figure a critic would compute — so you don't have to take our word for any of it.

Scored 8 of 25 Refused 15 of 25 Classified without PnL 2 Generated 2026-07-15

The headline, with its caveat attached

Our accuracy claim is selective: when the public fill data supports a reconstruction, we publish one; when it doesn't, we refuse to score the wallet rather than guess. A selective claim is only honest as a pair — coverage and conditional accuracy — so both are stated together, always.

5.1 percentage points — the median gap between our reconstruction and Polymarket's displayed 30-day profit, across the 10 wallets meeting a structural rule applied blind to all 25. Aggregate: our reconstructions sum to 98.7% of the displayed dollars. Worst case 15.2 points (wco26). 5/10 within ±5 points, 6/10 within ±10, 10/10 within ±20. No wallet in this set flipped sign. 4 of the 10 carry open positions, which Polymarket values with its own /value endpoint rather than our fills; for 2 of them that mark dominates the total — decomposed in the correction below.

The reconstruction systematically under-reports: 8 of 10 land below the displayed figure (median signed error -2.8 points). That direction has two documented causes in our own pipeline before any market explanation is needed. First, 6 of the 10 wallets were fetched over a per-market code path that under-counts redemption and merge inflows and misses conversions entirely — a known engine defect, scheduled for repair in recon-error-v2 — and those 6 under-report more (mean signed error -4.9 points) than the 4 fetched cleanly (-2.7). Second, we debit trading fees and the displayed figure does not (§03), which depresses every reconstruction slightly. Off-chain hedges, cross-venue positions and a trader's other wallets are also invisible to on-chain fills and may widen the gap further — but that is a hypothesis, and we will not attribute to markets what we can trace to our own code. The net direction still runs the way we would choose: we would rather understate a profit than inflate one.

Correction, added 2026-07-15 (after first publication): 4 of the 10 are not pure reconstructions — and here is exactly how impure. For a wallet still holding open positions, our published total is realized cash-flow from fills plus the live value of the open book — and that open-book value comes from Polymarket's own /value endpoint, not from our fill log. For 8 of the 10 headline wallets this component is ≤10% of the total (6 of them: zero). For 2 it dominates: asparagus2012 (98% of our figure is Polymarket's valuation) and wco26 (open-book value exceeds the published total — the fills-only cash-flow is negative). For those 2, "our reconstruction agrees with the displayed profit" is substantially "Polymarket's valuation agrees with Polymarket's other valuation" — a much weaker test, and this page failed to say so when it first went up. The frozen headline above stands as published, wco26 worst case included. The fills-independent subset (open book ≤10% of total, n=8) reconstructs to a median gap of 3.6 points, aggregate 99.6%, worst case 15.0 points (filthybetz), 6/8 under. We are not swapping the headline for the better number — a frozen metric does not get to improve after the fact, even in our favour. The independence requirement is registered into recon-error-v2 (§03) and binds from the next cycle forward.

Second correction, added 2026-07-16: one of the 10 was reconstructed from a fill log our engine had already flagged as incomplete — and this page did not say so. DEEDDIT's fill history is fetched market-by-market because its full log exceeds the /activity endpoint's hard offset cap — and its largest market hits that same cap on its own: the endpoint returns exactly 3,500 rows and refuses the rest. The engine recorded this honestly (market_capped: true, visible in the published decode index the whole time); the prose on this page ignored it. What the truncation can hide is bounded by the reconciliation itself: DEEDDIT still lands at 102.3% of displayed, so the unfetched fills net to roughly −$187,597 — 2.3 points — unless errors offset, which a known-incomplete log cannot rule out. Two consequences, stated plainly. DEEDDIT is the largest wallet in the headline set (37.1% of its displayed dollars) and it flatters the aggregate: without it, the set reconstructs to 96.5% of displayed dollars (median gap 5.2 points, n=9, worst case still 15.2). The frozen headline keeps it, because frozen means frozen in both directions — the same discipline that kept wco26 in. And the fills-independent subset in the correction above also contains it; the subset that is both fills-independent and fetch-complete is n=7: median 4.9 points, aggregate 97.6%. From the next cycle, an engine-flagged truncated fetch excludes a wallet from headline membership by rule (recon-error-v2, §03).

Third correction, added 2026-07-16: a fetch-completeness defect was found on a wallet outside this set, and we cannot prove the wallets inside it are unaffected. Our fetch for swisstony (rank 2, outside this headline set at 434 markets) returned 15,410 fills across 434 markets. A recursively-subdivided re-fetch of the same window returns 539,113 fills across 15,122 markets — we had seen 2.9% of its book. Every completeness signal we own read green while it did. market_capped is false for that wallet, because the flag detects a per-market offset cap, and this truncation is at market enumeration — which the flag cannot see. That matters for this page beyond the one wallet: the retroactive check recorded in §03 (iv) — discovered markets meet or exceed the public profile count for all 10 headline wallets — is the only evidence this set is unaffected, and it is not proof, because it compares our count against another figure Polymarket serves us rather than against the chain. We have not re-fetched the other 24 wallets in this cycle, and we do not claim their fetches are complete. If the defect proves general, reconPct, the verdicts that cite it, and the figures frozen above are all affected. The cause is not established here — the engine is not in this repo — so it is logged rather than explained: U2 in the standing unreconciled-items table.

Why this set and not the 8 we scored. Our verdict assignment uses reconstruction agreement as an input — every "Insufficient evidence" verdict below cites it as the reason. So "the wallets we scored reconstruct well" is partly true by construction, and we don't lead with it. The headline above instead uses a purely structural rule (nMarkets <= 50 AND nFills <= 10000) applied to all 25 regardless of verdict. It excludes 1 wallet we did score — alwayslatetotheparty (812 markets) — which is the evidence it was not reverse-fitted to flatter us. For completeness, the 8 scored wallets have a median gap of 3.6 points and an aggregate of 99.4%. As first published 2026-07-15 this read 12 wallets, 8.0 points, 96.8%. The 2026-07-16 re-stamp (above) moved CandleHammerDrums and asparagus2012 and filthybetz and wco26 out of the scored set, because this set is defined by verdict. The figure therefore describes a different population — it is not a better reconstruction, and nothing was re-measured. We print both because a published number does not get to improve quietly, and this one moved in our favour. The frozen headline is unaffected: it is chosen without reference to verdicts, and the build refuses to publish if it moves.

The number our critics will compute

If you pool all 25 wallets — mixing the ones we scored with the ones we explicitly refused — you get this. We are publishing it ourselves, because you would find it anyway and it deserves an explanation rather than a discovery.

Pooled across all 25, including refusals

30.1 pts

median absolute error · worst case 931.3 pts · 8 sign flips

The 10 we actually claim

5.1 pts

median absolute error · worst case 15.2 pts · 0 sign flips

Both numbers are real. They describe different things. The pooled figure averages our successes together with cases we told you we could not measure — it is a statistic about a population we never claimed to describe. The honest use of the pooled number is as evidence that the refusal rule is doing real work: 7 of the 15 refused wallets produce reconstructions that flip sign entirely, and the refusal mechanism caught every one of them before publication.

The distinction that matters, and we will not blur it: of the 15 refusals, 7 genuinely contradict the public data — the fills we can see cannot produce the displayed profit. The remainder are outside our engine's verified range: very large multi-market books where merges, redemptions and per-market API limits defeat cashflow reconstruction. The first is a finding about the wallet. The second is a limit of the method. Both get the same label — "Insufficient evidence" — and neither is an accusation.

How error is defined

Errors are measured against Polymarket's displayed 30-day profit — the number shown to anyone deciding whether to copy a wallet. That is the target because it is the claim under test. It is not treated as adjudicated truth.

// per wallet i, where D = displayed profit, R = our reconstruction reconPct = 100 · R / D APE = 100 · |R − D| / |D| // absolute error, in points — the headline metric SPE = 100 · (R − D) / |D| // signed error — bias diagnostic (negative = we under-report) // failure class — a sign flip is not a large error, it is a different phenomenon SIGNFLIP if R · D ≤ 0 → excluded from every aggregate; counted and shown OK otherwise → enters the distributions above

We report the median, not the mean. With a heavy-tailed error mechanism, one blown reconstruction moves a mean by tens of points while leaving the median untouched — a mean would let a single outlier either flatter or condemn the whole set. Counts within ±5, ±10 and ±20 points are published alongside, because a band count is the least gameable form of an accuracy claim.

Basis mismatch, direction: against us. Our reconstruction debits trading fees paid (it sums the usdcSize of each fill); Polymarket's displayed profit is computed before fees (price × size). Every reconstruction on this page is therefore pushed slightly below the target by definition — it is one of the documented causes of the under-report bias in §01, not evidence about any wallet. Measured like-for-like on the flagship wallet (myzbsq, fee-excluded on both sides, computed 2026-07-15), agreement moves from 99.3% to 99.98% on the cycle. We leave every headline figure on the harsher basis rather than restate it upward after the fact; fee-basis matching is part of recon-error-v2 below.

Pre-registration, stated honestly. The structural rule (nMarkets <= 50 AND nFills <= 10000) was chosen on 2026-07-15 with this cycle's data already in view. For this cycle it is therefore post-hoc, and we say so rather than implying otherwise. It is frozen as of that date and will be applied unchanged to every future cycle before we look at the results. The forward test — not this page — is the falsifiable claim.

Amendment — recon-error-v2, registered 2026-07-15, before any next-cycle data exists. Effective next cycle: (i) headline membership additionally requires the open-book /value component to be ≤10% of the reconstructed total — wallets above it are still published, labelled value-passthrough, and claimed as nothing; (ii) error is measured fee-basis-matched against the displayed figure; (iii) the per-market fetch defect described in §01 is repaired, gated on re-validating every wallet in this cycle's table.

(iv) — registered 2026-07-16, still before any next-cycle data exists. A wallet whose fill fetch the engine flags as truncated (market_capped) is excluded from headline membership — published, labelled fetch-truncated, claimed as nothing. And because that flag fires only when an individual market's fetch hits the cap — it is a floor on truncation, not a certificate of completeness — market discovery is separately validated against the wallet's public profile market count, with any shortfall disclosed per wallet. (Checked retroactively for this cycle on 2026-07-16: discovered markets meet or exceed the public profile count for all 10 headline wallets.) Nothing on this page was recomputed under v2. This cycle's numbers stay as frozen under v1, including where v2 would make them look better.

Every wallet, every error

All 25 wallets in the July 2026 cycle. Nothing is withheld or trimmed. Sign-flipped reconstructions are shown with their raw error and excluded from aggregates.

Top-25 leaderboard — displayed vs. reconstructed, July 2026 cycle
#	Wallet	Verdict	Displayed	Reconstructed	Recon %	Error (pts)	Markets	Fills	Note
1	DEEDDIT 0x09b4…7320	Real edge	$8,052,185	$8,239,782	102.3%	2.3	17	8,310	in structural set · fill log truncated in its largest market — see §01
2	swisstony 0x204f…5e14	Insufficient evidence	$6,830,769	$1,494,391	21.9%	78.1	434	15,410	—
3	maz26 0x6754…e3bc	Insufficient evidence	$3,244,219	−$7,150,894	-220.4%	320.4	135	38,386	sign flip — excluded
4	CandleHammerDrums 0x7c1e…bab3	Insufficient evidence	$2,613,512	$2,271,712	86.9%	13.1	207	14,695	—
5	Sparkling8899 0x2e25…d49e	Insufficient evidence	$2,421,465	−$1,703,555	-70.4%	170.4	29	22,682	sign flip — excluded
6	Jsram 0x8372…5223	Insufficient evidence	$2,369,907	−$533,695	-22.5%	122.5	159	18,350	sign flip — excluded
7	0x2c335066FE58fe9237c3 0x2c33…0563	Market-making	$2,368,906	−$19,692,313	-831.3%	931.3	1,607	333,368	sign flip — excluded
8	ramadamaramadam 0x5b4e…0232	Insufficient evidence	$2,229,687	$1,401,755	62.9%	37.1	67	11,588	—
9	asparagus2012 0x476e…396d	Insufficient evidence	$2,148,287	$2,036,726	94.8%	5.2	16	3,408	in structural set · open book = 98% of total via /value — see §01
10	muchobliged 0x095f…90b2	Real edge	$2,029,201	$1,788,380	88.1%	11.9	20	6,296	in structural set
11	weatherman12 0xd1c5…cbd0	Real edge	$1,822,553	$1,819,117	99.8%	0.2	41	9,898	in structural set
12	hot2trot 0xb61b…8a06	Real edge	$1,787,887	$1,981,652	110.8%	10.8	47	5,929	in structural set
13	palegrit 0xf5fa…62df	Real edge	$1,692,747	$1,691,816	99.9%	0.1	24	1,207	in structural set
14	Allezpapa 0xe549…2d54	Insufficient evidence	$1,666,640	$2,409,121	144.5%	44.5	141	47,271	—
15	myzbsq 0x9468…5d4b	One-shot	$1,459,042	$1,448,170	99.3%	0.7	5	1,003	in structural set
16	ndb1 0xfea3…d227	Artifact	$1,178,380	$1,533,079	130.1%	30.1	1,302	46,458	—
17	0x076daa87 0x076d…8d4c	Insufficient evidence	$1,114,095	−$968,342	-86.9%	186.9	2,960	10,739	sign flip — excluded
18	cnyek 0x6501…1220	Insufficient evidence	$1,107,781	$1,815,006	163.8%	63.8	67	1,876	—
19	alwayslatetotheparty 0xb687…aee5	One-shot	$1,106,963	$916,529	82.8%	17.2	812	270,446	—
20	highnetworth 0xbb5f…2e21	Real edge	$1,064,758	$1,012,363	95.1%	4.9	16	988	in structural set
21	0xC41D736bDed9ED1acCD6 0xc41d…74fd	Insufficient evidence	$1,023,925	−$1,150,007	-112.3%	212.3	221	22,400	sign flip — excluded
22	pada 0x714a…ab2f	Insufficient evidence	$1,012,076	−$143,724	-14.2%	114.2	91	6,093	sign flip — excluded
23	FootballFan98 0xc31d…4ebe	Insufficient evidence	$972,897	−$619,521	-63.7%	163.7	137	21,028	sign flip — excluded
24	filthybetz 0xc6a6…9458	Insufficient evidence	$960,759	$816,222	84.96%	15.0	42	8,067	in structural set
25	wco26 0x3f6e…18d9	Insufficient evidence	$676,248	$573,378	84.8%	15.2	11	7,348	in structural set · open book = 223% of total via /value — see §01

Reading the flagship. myzbsq's error is 0.7 points — the 3rd-best of the 8 scored wallets. It is our cleanest case, not our typical one, because it is a small, self-contained, book — 5 markets in this cycle, and a single self-contained market at the July-4 snapshot the case study decodes. Any claim we make from myzbsq is a claim about a best case, and we label it that way.

Why you may see 99.6% elsewhere on this site. Two different measurement windows, both real. The flagship case study decodes a July-4 snapshot — one self-contained World Cup market, $1,223,224 displayed, reconstructed to 99.6%. This table uses the July 2026 30-day cycle: $1,459,042 displayed across 5 markets, reconstructed to 99.3%. Same wallet, different window. Only the cycle figure enters the distribution above.

The 15 we refused, and why

These are the wallets we could not score. Each rationale is reproduced verbatim from the decode run — not written for this page — except the 4 rows marked Re-stamped, whose verdicts were corrected after publication; those sentences were written for the correction and say so. None of these is an allegation of wrongdoing. "We cannot verify this from public data" is the ceiling of what we claim, and it applies to us as much as to anyone else copying them.

Insufficient evidence — verbatim rationales
Wallet	Recon %	Class	Rationale (verbatim)
swisstony	21.9%	—	Our fetch for this wallet saw 15,410 of at least 539,113 fills in the window — 2.9% of its book — so nothing below is evidence about the wallet, including the 21.9%. Re-measured 2026-07-16; unreconciled (UNRECONCILED.md U1). We do not certify it either way.
maz26	-220.4%	sign flip	Our own fill log for this book is truncated, so rebuilding it lands at −220% of displayed — a figure about our fetch as much as about the wallet. MM-leaning on structure; the gap is unreconciled and we attribute it to neither.
CandleHammerDrums	86.9%	re-stamped	Re-stamped 2026-07-16 (was Real edge): the wallet's own fills rebuild 82.47% of displayed profit — 17.53 points from par, outside the ±15% band the scorecard publishes as the condition for the stamp. An open-book mark worth 5% of the total narrows the gap to 86.92%, but that mark is Polymarket's valuation, not our reconstruction. We do not certify an edge on the number under test.
Sparkling8899	-70.4%	sign flip	Directional-looking on structure, but our own fill log here is truncated — the −70% rebuild is not evidence about this wallet either way. The gap is unreconciled.
Jsram	-22.5%	sign flip	Broad sports activity; PnL does not reconcile.
ramadamaramadam	62.9%	—	63% reconstruction on a fill log our own engine flags truncated — the shortfall may be ours, and no dominant tell separates it. The gap is unreconciled.
asparagus2012	94.8%	re-stamped	Re-stamped 2026-07-16 (was Real edge): our own fills reconstruct 2% of displayed profit, so the 94.81% agreement is mostly Polymarket's own mark on positions the wallet still holds (98% of the total), not our rebuild of its trades. We do not certify an edge on the number under test.
Allezpapa	144.5%	—	Over-reconstructs to 145% on a fill log our own engine flags truncated — an overshoot a partial log can produce on its own. The gap is unreconciled.
0x076daa87	-86.9%	sign flip	2,960 markets, spray pattern; PnL won't reconcile.
cnyek	163.8%	—	Reconstruction overshoots to 164% — cannot verify.
0xC41D736bDed9ED1acCD6	-112.3%	sign flip	Reconstruction fails; directional shape, unverified.
pada	-14.2%	sign flip	Some late-price buys but no dominant tell; won't reconcile.
FootballFan98	-63.7%	sign flip	Reconstruction fails; no decisive structural signal.
filthybetz	84.96%	re-stamped	Re-stamped 2026-07-16 (was Real edge): reconstructs to 84.96% of displayed profit — 15.04 points from par, outside the ±15% band the scorecard publishes as the condition for the stamp. The figure is built from the wallet's own fills; it simply does not land inside our promise.
wco26	84.8%	re-stamped	Re-stamped 2026-07-16 (was Real edge): our own fills reconstruct -104% of displayed profit, so the 84.79% agreement is mostly Polymarket's own mark on positions the wallet still holds (223% of the total), not our rebuild of its trades. The rebuilt total is itself 84.79% — 15.21 points from par, outside the same band. We do not certify an edge on the number under test.

That 15 of 25 of the most-copied wallets on the leaderboard cannot be verified from public data is, in our view, the most important number on this page. It is not a statement about those traders. It is a statement about what a person copying them can possibly know.

What we do not claim

Not "we reconstruct wallets to 99% accuracy." 3 of the 10 do reconstruct above 99% — the best is 99.9% (palegrit) — but the median across the set we claim is 94.9%, and the pooled all-25 figure is far worse (§02).
Not an average across all 25. Pooling wallets we scored with wallets we refused produces a number that describes no real population. The pooled median is 30.1 points and we publish it in §02 rather than an average we'd have to defend.
Not "we audited 25 wallets successfully." We scored 8. We refused 15. We classified 2 by structure without a usable PnL reconstruction.
No classification-accuracy percentage. There is no ground truth for "was this really skill?" Our verdicts derive from published structural signatures — two-sided fill share, post-resolution buying, market concentration — and those inputs are on every wallet's page so the reasoning can be checked, not just the conclusion.
No claim that Polymarket's numbers are wrong. We measure agreement with the displayed figure. A disagreement is a reason to look closer, not a finding of misconduct.
No generalisation beyond this leaderboard. Every figure here is scoped to the top-25 by profit, July 2026 cycle. Error behaviour on other wallets is unknown and unclaimed.

How to attack this

The fair criticisms, pre-empted — because you should not have to find them yourself.

"You only score the ones you can reconstruct, so of course you look accurate." Correct, and it is the design. That is why coverage (8/25) is printed next to every accuracy figure, why the refusal rule is published, and why the headline uses a verdict-blind structural rule.
"Your thresholds were fitted after seeing the data." For this cycle, yes — stated plainly in §03. Frozen going forward; the next cycle is the real test. The rule excludes one of the 8 wallets we score after the 2026-07-16 re-stamps (alwayslatetotheparty) — and excluded two of the 12 as first scored — so it was not tuned to bless our own verdicts.
"n = 10 is small." It is. That is why counts sit beside every percentage and why we claim no significance beyond them. The remedy is accumulation: every future cycle publishes this same table, including the bad ones.
"Your ground truth is Polymarket's own unaudited number." Correct — and it is the right target, because the product question is whether the displayed claim survives contact with the fill log.
"Your #1 wallet's fill log is incomplete — your own index says so." Correct. market_capped: true has been in the published decode index since this page went up, and until 2026-07-16 the prose ignored it. Disclosed and bounded in §01 — the reconciliation caps the omission at about 2.3 points unless errors offset — and excluded from headline membership by rule from the next cycle (recon-error-v2, §03).
"Part of your 'reconstruction' is Polymarket's own number too." For 2 of the 10 headline wallets, yes — the open book is marked by Polymarket's /value endpoint and it dominates their totals, which makes the test circular for those 2. Disclosed and decomposed in §01, with the fills-independent subset (n=8: median 3.6 points, aggregate 99.6%) printed beside the frozen headline. From the next cycle, wallets like these are excluded from the claim by rule (recon-error-v2, §03).
"Top-25 by profit is an extreme sample." Correct. All claims are scoped to it.
"You can't see everything." True and load-bearing: we see on-chain fills only. Off-chain hedges, cross-venue positions and a trader's other wallets are invisible to us, and sophisticated traders deliberately split books across wallets. This is exactly why large multi-market books get refused instead of scored.

Reproduce it yourself: every wallet address, market condition ID and fill timestamp behind these numbers is public Polygon data. If you find an error in any reconstruction here, we will correct it on this page and say what changed.

Disclosures

Verdict inputs. Reconstruction agreement with the displayed figure is one input to verdict assignment. The scored-set accuracy figure is therefore a self-consistency statistic; our headline claim uses the verdict-independent structural set instead.
Positions. We hold no position in any wallet, market or outcome discussed here, and we take no payment from any wallet owner, copy-trading tool or platform.
Business model. MSR Decode sells wallet-analysis reports and a monitoring subscription. We sell no signals, no picks and no copy-trading product. Nothing on this page is for sale and no verdict is purchasable.
Subjects. We analyse pseudonymous wallets, not identified people, and we do not allege illegality. Classification is opinion, formed from the published facts and labelled as opinion.
Corrections. Errors found in this workpaper are fixed here with a dated note. Method changes are versioned (recon-error-v1 (frozen 2026-07-15)).

Data. Realized cash-flow is reconstructed from the public Polymarket /activity fill log on Polygon: sells + redemptions + merges + rewards + rebates − buys − splits, inclusive of fees paid. For wallets still holding open positions, the open book is valued by Polymarket's own /value endpoint and added — so a published total is fills-derived cash-flow plus Polymarket's mark on whatever remains open, and it is only as independent of the platform as that second term is small. Open-book share is flagged per wallet in §04 wherever it exceeds 10%, and decomposed in §01. Where a single market's fill fetch hits the /activity offset cap, the engine flags the wallet (market_capped) and the flag is disclosed here. It is a floor, not a certificate: it fires only on per-market truncation, and market discovery itself relies on the same capped endpoints — so an unflagged wallet is one whose known markets were fetched completely, not one guaranteed to have no unknown markets. Reconciled against the platform's displayed profit. Cycle: July 2026. Generated 2026-07-15. Analysis, not investment advice.