FILE 07 · THE RECORDAs of Jul 26, 2026

What we shipped. What we got wrong. What’s next.

This is the honest chronology — including the errors. MSR Decode publishes a monthly leaderboard audit and a forward-tested ledger, and when we find our own mistakes we correct them here, dated, in public, without softening the original. Nothing below is edited after the fact; corrections are appended, never erased.

Cycle July 2026 Corrections dated & append-only As of Jul 26, 2026

Shipped

What actually went live this cycle, in order.

Jul 15

The Leaderboard Scorecard — July 2026 cycle

Audited Polymarket’s top 25 wallets by 30-day profit. 6 of 25 earned a Real-edge, copyable rating; the rest were variance, structure, or unverifiable. Re-run monthly. Read the scorecard →

Jul 15

Reconstruction Workpaper published — methodology v1 frozen

Every per-wallet reconstruction error published, spec recon-error-v1 frozen the same day. Headline: median gap 5.1 points, aggregate 98.7% of displayed dollars, across the 10 wallets in the structural set — and 15 of 25 refused rather than guessed. Read the workpaper →

Jul 2026

Two case files published

The Leaderboard Autopsy — a top wallet’s celebrated “edge” decomposed, fill by fill, into redemption arbitrage a copier could never have captured. Sample Decode: myzbsq — a $249K → $1.22M World Cup bet reconstructed to 99.6% of Polymarket’s displayed profit (MSR’s cleanest case, a closed single-market wallet) and classified a one-shot, not a repeatable edge. Autopsy → · myzbsq →

Jul 2026

Evidence-confidence gate shipped

A second, independent label — Verified / Provisional / Insufficient data — grades fetch completeness separately from the six-way verdict, so a copyable stamp can no longer imply a complete fetch it doesn’t have. Of the 25: 6 Verified, 4 Provisional, 15 Insufficient data. See the rule →

Jul 19

Prospective Copyability Ledger launched

Every Real-edge verdict now gets one ledger entry: follower-accessible prices frozen at entry, outcomes published at 24h / 7d / 30d — gains and losses alike. Cohort 1 froze 6 entries on Jul 19; the d7 checkpoint published on schedule Jul 26, 2026 and d30 comes due Aug 18, 2026. Read the ledger →

Jul 24

Case file: the BNB 5-minute wallet

The viral wallet that turned $449K into $1.37M in 3.1 hours, reconstructed to the cent: net realized $917,311.53, matching the viral “~$917k” claim to within $311.53. Read the case →

Jul 25

Method report: the endpoint that invents fraud — and a retraction of our own finding

A standard integrity check told us three top-leaderboard wallets had booked $926,462.58 of profit on positions they never paid for. The check was ours and it was wrong. Polymarket’s /trades endpoint defaults to takerOnly=true, so it silently returns only part of a wallet’s fills — and in markets with no order book that hides nearly all of them. Re-fetched properly, the 150 largest of the 4,725 flagged markets reconciled 150 of 150. All three traders were ordinary. Read the retraction →

Jul 26

The d7 ledger checkpoint published on schedule

Cohort 1’s 7-day outcome came back identical to the 24-hour marks: hot2trot $1,475 (+47.5%), highnetworth $1,665 (+66.5%) on a simulated $1,000 each, the other four flat because they had nothing a follower could buy. Every copyable position had resolved within about a day of the freeze — so the copyable window on this cohort was a day, not a week, which is a limitation of our own checkpoint interval and is published as one. Read the checkpoint →

Corrections & refusals

The part that differentiates us: what we caught, what we corrected in public, and what we refused to say because the data didn’t support it.

15 of 25 wallets refused rather than guessed. “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 the workpaper — a statement about what a follower can know, not an allegation about those traders. See every rationale, verbatim →

Jul 16, 2026 — the incomplete-fetch disclosure. After publishing figures from it, MSR found one wallet whose fetch held 2.9% of its 30-day book (0.5% of its lifetime book) — while every completeness check MSR owns read green. The follow-up sweep of the cycle’s other 24 wallets found 16 consistent with their published counts, 3 more incomplete and rebuilt, and 5 inconclusive. See the method →

Jul 16, 2026 — four wallets re-stamped. CandleHammerDrums, asparagus2012, filthybetz and wco26 moved from Real edge to Insufficient evidence after closer review found their agreement with the displayed figure rested on Polymarket’s own open-book valuation, or fell outside the published ±15% band. The workpaper’s scored set now stands at 8 of 25, with 15 of 25 refused. Read the corrections →

Jul 25, 2026 — we retracted a fraud finding of our own. We had flagged three top-leaderboard wallets for $926,462.58 of profit that did not reconcile against their trades — the shape of PnL arriving from transfers rather than trading. The defect was in our method, not their trading: Polymarket’s /trades endpoint defaults to takerOnly=true and silently returns only part of a wallet’s fills, which in markets with no order book hides nearly all of them. Re-fetched with takerOnly=false, the 150 largest of the 4,725 flagged markets reconciled 150 of 150, against $1,361,899.05 of real buying the default view never showed us. All three traders were ordinary. The per-market reconciliation and the calls to reproduce it are published with the retraction. Read the retraction →

21 of 24 markets closing net-positive in the BNB 5-minute wallet, reconstructed fill by fill — not the viral “23 of 24” claim. The most charitable reading MSR could construct still tops out at 22 of 24; no reading in the dataset reaches 23. Read the correction →

4 of 6 — a different count, on purpose. On day one of the Prospective Copyability Ledger, 4 of the ledger’s first 6 Real-edge entries had zero follower-accessible positions the instant their prices were frozen — every position had already resolved or gone redeemable, or the wallet held nothing at all. This is an accessibility count about a reader acting after publication, not a re-statement of the 6 of 25 verdict count above — the two measure different questions and neither wallet re-uses the other’s figure. Read what day one shows →

What’s next

Only commitments already on the record — nothing new promised here.

d30 checkpoint — Aug 18, 2026. Cohort 1’s 30-day outcome comes due on the same frozen entries. Every copyable position in the cohort had already resolved by the d7 checkpoint, so this one is determined rather than forecast — it publishes on schedule either way. Ledger →
Monthly scorecard refresh. The top-25 audit re-runs monthly; the next cycle’s accuracy rule (recon-error-v2) is already pre-registered and frozen ahead of the results. Scorecard →
Dataset access on request. Structured historical-and-prospective verdicts, including the append-only ledger, remain under evaluation for research, risk & compliance teams — ask about intended use. For teams →

Nothing above is a new promise. Every line in this section is a date or a policy already published on the page it links to; this page adds no figure and no commitment that didn’t already exist there.