Levandor

crypto-shortterm — "Certain Winner" Competitor Refuted (wowitsamazing / 0xd02b)

17 min read

A 38-day lifetime harvest — now bulletproof-validated by a pre-registered 3-leg validation (clean-room re-derivation + adversarial statistics + on-chain reconciliation — all three legs closed, matrix complete) — refutes the claim that Polymarket wallet 0xd02b6d910a38479c3125308fc4737a46509cd6df (wowitsamazing / pseudonym Novel-Plow-Toot) is the market’s sole skill-verified winner. He is a net loser of $84–90k whose win rate equals the price he pays in every price×τ cell (any BTC-5m edge >~1.8¢ excluded at 100% label coverage); his loss is exactly the Polymarket taker fee (1.72% effective after rebates). The adversarial leg sharpened the claim: no FEE-CLEARING edge — not “zero skill”. Not a threat but a live validation of our own net-edge gate on ~190k real fills.

Summary

Our 24h competitor tape study flagged this wallet as the only statistically-credible winner (z=4.29, “242-0”, +84,112 / −89,882** — original cutoff / cleanroom same-cutoff / full; cash components reproduce exactly). At 100% CLOB-truth label coverage (70,493 fills, size-weighted, day-block bootstrap): overall win − price = +0.22¢, 95% CI [−1.48, +1.77]¢ — any aggregate BTC-5m edge larger than ~1.8¢/share is excluded. What sinks him is the Polymarket taker fee 0.07·x·(1−x)·size (400 roll / $8 fills. Canonical repo doc: docs/superpowers/winner-dissection-2026-07-03.md.

The most valuable takeaway — this wallet validates our net-edge gate

The fee that bleeds him dry is the exact 0.07·eff·(1−eff) term in our validated live net-edge gate (walkforward.rs trade_decision_slipped → the EMIT_MIN_EDGE / observe::net_buy_edge gate). A high-volume taker with no fee-clearing edge who crosses the spread loses precisely that fee — confirmed on ~190k real fills (per-fill ratio ≈1.0, corr 0.9915). A $4.3M/38-day natural experiment, run by someone else, proving fee-unaware flow is −EV and that subtracting 0.07·eff·(1−eff) before firing is the right discipline. The validation’s book-calibration audit independently re-validated EMIT_MAX_ASK 0.85 and the edge-vs-ask rule from the book side too. See Why this matters — live validation of our net-edge gate.

Corrects / supersedes the 2026-07-02/03 competitor thread

This overturns the empirical premise behind several LOG entries and the oracle-state work:

  • “the only statistically-credible winner (z=4.29, +12.9k was one variance-positive day; the “242-0” was a capped-tape measurement artifact (mechanism named: activity-dependent tape truncation, §C8); the “79.3%” was a redeem-proxy overcount (two-sided hedging always redeems).
  • The “sole Bonferroni survivor” methodology (per-wallet skill z on capped tapes) is retracted entirely — capped /trades?market= tapes are unusable for per-wallet PnL in both directions (maker-invisibility + recency slice). Archetype/band-level cell economics of the 24h study survive with its §0 caveats. Erratum prepended to docs/superpowers/competitor-analysis-2026-07-02.md.
  • “ORACLE-STATE TIER DISCOVERED — the competitor’s 100% record solved” (see crypto-shortterm-algo-accuracy-audit-2026-07-02) — the near-certain record it set out to explain does not exist; his real per-cell win rate just equals the price. Chainlink Data Streams hypothesis falsified three independent ways (see below). Corroborates the same-day oracle-state Phase-0 kill AND the refit study’s H4 print-conditioning harm — print-state is now closed from three independent directions.

The prior claim (24h tape)

The competitor-analysis session (6-agent workflow, 379,732 trades / 12,715 wallets, 24h BTC 5m+15m tape) named this wallet the sole wallet clearing the selection-adjusted skill bar: z=4.29, +$12.9k / 24h, apparent 242-0 on BTC 5m/15m. That single record motivated the head-to-head (joined on condition_id) and the oracle-state-tier spec (docs/superpowers/specs/2026-07-03-oracle-state-tier-design.md), on the theory that he conditions on settlement-oracle print state to win near-certainly.

The refutation (38-day lifetime harvest)

PnL — every method says breakeven-to-losing

MethodLifetime PnL
True USDC cash flow (orig cutoff / cleanroom same-cutoff / full)−87,755 / −$89,882
Nominal price × size (pre-fee)+$3,431 (breakeven)
Polymarket leaderboard (profit=all)−$7,590
Polymarket live MTM (user-pnl-api, noisy)−7.6k
Taker fee paid (Σ 0.07·x·(1−x)·size)**−89.4k; per-fill ratio ≈1.0, corr 0.9915)
Rebate income (maker 14,028)+$14,406 (~16% of fees → 1.72% effective fee)
Fee eracutover ~2026-06-13: pre = no fee, −4.1k nominal − $87.5k fee
Losing days22 of 35 active (daily swings −19.9k)

The true-USDC numbers are authoritative (clean-room reproduced REDEEM/SELL/SPLIT/MERGE/rebate components exactly; BUY within +0.08% — its deeper paging found ~289 extra fills, so the loss is slightly deeper than first measured). Nominal and the Polymarket figures flatter him by hiding the taker fee — see Gotcha 2 — usdcSize includes the Polymarket taker fee (not a bid-ask spread). LP/volume-rewards rationalization dead: he is ~100% taker (22 maker events lifetime; note ~32% of his fills still execute as maker — resting ladder limits).

Win rate == price paid, in every cell (100% label coverage)

At 100% CLOB-truth label coverage (4,904 BTC-5m markets; the 1,621 pre-Jun-8 labels backfilled — data/btc5m_winners_backfill_pre_jun8.csv), size-weighted, day-block bootstrap:

cell (px, τ-to-close)n fillswin%px%win−px95% CI (¢)
<0.60, ≤150s16,76910.810.6+0.25[−0.83, +1.07]
0.60–0.75, ≤150s4,80167.968.8−0.94[−4.02, +1.99]
0.75–0.85, ≤150s5,07680.979.2+1.73[−1.13, +4.19]
0.85–0.95, ≤150s1,44589.088.5+0.48[−7.70, +6.48]
<0.60, 150–300s12,95649.148.2+0.90[−3.05, +3.57]
0.60–0.75, 150–300s19,09068.467.8+0.57[−2.41, +3.63]
0.75–0.85, 150–300s9,13078.078.5−0.52[−3.41, +2.32]
0.85–0.95, 150–300s1,06282.687.5−4.87[−19.92, +2.81]
overall70,49355.555.2+0.22[−1.48, +1.77]
  • Any aggregate BTC-5m edge larger than ~1.8¢/share is excluded (day-clustered). No cell shows a positive exploitable edge. Pre-fee PnL ≈ 1.7M+ of BTC-5m buys.
  • The 1,012 losses of the resolved 5m markets (651 on-chain-confirmed, 0 redeem-lag false-positives) are the fair-frequency losing tail, not tail-risk or adverse selection. Fill-level true win rate ~54%; market-level dominant-side rate 58.1%.

The redeem proxy overstates the directional rate

The apparent “79.3%” (from the lifetime anatomy) came from redeem-presence, which counts any market that produced a redeemable token as a win. He bought both sides in 36.5% of BTC-5m markets, and a hedged market always redeems → the proxy overcounts wins. The 24h “242-0” was a different artifact (capped tape — see §242-0). Both rules live in Redeem rows + the win-proxy caveat.

Bulletproof validation (3-leg, 2026-07-03)

Same-day, pre-registered design: (1) clean-room re-derivation (independent harvest + independent code, no shared lines), (2) adversarial statistics (attack every headline number), (3) on-chain reconciliation (Polygon as the final non-Polymarket source — ALL PASS). Every number in this note survived at least two independent derivations sharing no code.

#FindingCleanroomAdversarialStatus
F1Fee identity 0.07·x·(1−x) + cutover ~Jun 13exact componentscorr 0.9915; era split; alternatives killedCONFIRMED (on-chain: 1,837/1,837 fills within 2%; cutover confirmed)
F2No fee-clearing edge, win==price (BTC-5m)8/8 cells ≤0.1pt; 100%-cov CI [−1.5,+1.8]¢Simpson split; MDE bounds; deflationCONFIRMED (refined)
F3Net loser −$84…90kcomponents exact; deeperaccounting attacks survived; ~100% takerCONFIRMED (on-chain net −$87.1k dead-on; capital account closes)
F4Anatomy: ladder bot, change-points (flips off ~Jun 29)both-sides 36.5% exactweekly trajectory reproduced exactlyCONFIRMED
F5242-0 = artifactcap real; “hides losses” REFUTEDp≈1e-20; day-selection ≈0CONFIRMED; mechanism NAMED (§C8): activity-dependent tape truncation
F615m: real structure, mechanical prior dead (−1.1¢)feature file 0.00% mismatchpopulation kill; multiplicity; 20/20 spot-checkCONFIRMED
F7Redeem-proxy inflation (79% → 58%)exact reproduction—CONFIRMED
F8Harvest completeness (199k rows)superset harvest, +0.12%10/10 live re-queries row-exactCONFIRMED (3 sample days on-chain ≤3e-6 USD)

Label audit clean (spot-check 200/200). The matrix is now fully closed (both PENDING cells resolved same-day): the on-chain leg passed in full (below) and the §C8 mechanism is named and reproduced (see §242-0). Every finding now rests on ≥2 independent derivations sharing no code, with the Polygon blockchain as the final non-Polymarket source.

On-chain leg — ALL PASS (data/validation_onchain.md)

The third leg, run directly off Polygon — the only source Polymarket does not control. Every check passed:

  • Trading net: −87,755** (redemptions reconcile to $0.00 across 8,512 txs; buys within −0.05%).
  • Capital account closes end-to-end: deposits 26.4k − trading loss ≈ balance ≈ current cash $7.1k → a genuine self-funded net loser (no hidden inflows).
  • Fee formula on-chain: 0.07·x·(1−x) holds on 1,837/1,837 nonzero-fee fills within 2% (median rel-err −8e-6); the ~Jun-13 cutover independently confirmed from chain data.
  • 3 sample days reconcile to ≤3e-6 USD; leak check clean.
  • Last alternative hypothesis REJECTED — no loss-absorbing paired/sibling wallet: his maker counterparties are diffuse (max single-maker share 19% on one day; 174–495 distinct makers/day; the one recurring maker is itself an episodic Polymarket proxy, not a sibling).

Refinements the adversarial leg forced

  • “Zero selection edge” → “no FEE-CLEARING edge.” Simpson decomposition: the pooled zero hides two real opposite books — hedged both-sides trading has genuine gross skill (+4.4¢, CI [+0.9, +6.2], Bonferroni-surviving; concentrated in ladder-adds = within-window mean-reversion timing) while the single-sided directional book is net −4.8¢. Net of fees even the hedged edge fails the deflated bar (p=0.011). So: real gross skill existed in one sub-book, but nothing clears the fee.
  • “58.0% win vs 55.3¢ paid = +2.7¢ gross” was a mixed-weighting error (market-level dominant-side rate vs fill-level size-weighted price; both validators caught it independently). Consistently weighted, the aggregate gap is +0.22¢ ≈ 0.
  • Effective fee is 1.72%, not 2.06% — the rebate program claws back ~16% ($14,406).
  • Power limit — his tape cannot rule on OUR band. In 0.75–0.85 @ τ≤150s his data cannot exclude a +2–4¢ edge; his tape neither validates nor threatens our regime. Our purged walk-forward remains the only evidence that rules there.

What he actually keys on — Data Streams falsified

  • His side = sign(last on-chain Chainlink print − K) on 92% of entries; = our frozen p_cal side on 94%; = Binance-spot sign only 52%. He is a print-direction follower — a bigger, earlier, ungated, fee-paying version of us.
  • Chainlink Data Streams hypothesis falsified three independent ways: (1) he performs at the stale-print ceiling (0.795 realized vs the print’s 0.776), never above it; (2) universe print precision is flat in print age at 60–150s (0.77 @ ≤2s vs 0.75 @ 30–60s — a live oracle value is worth ≈ a 40s-stale print because the outcome is undetermined 3 min out); (3) entries are not synchronized to print arrivals (print-age at entry median 17s = random moments). The “bets against the market and wins” mystique = watching the wrong feed: he follows the print, and near strike the print beats spot 70/30 where they disagree.
  • Data Streams research track: demoted from “explains a competitor’s edge” to a label-quality concern only. (Label side re-confirmed: our on-chain settle disagrees with CLOB truth on 12.6% of windows, 35% of those within 2bps of strike — actual-resolution labels remain binding for all analyses.)
  • Enters early: median 203s-to-close; 79% of entries before our τ∈[60,150)s regime even opens.

An operator who iterates — the change-points

Three change-points in 38 days: 5-day outage/rebuild Jun 5–10; multi-asset expansion Jun 22–23 (xrp/hype Jul 1); and — decisive for competitor modeling — flip-chasing switched OFF ~Jun 29: his both-sides rate went 15.5% → 61.5% (peak, wk Jun 15) → 4.7%. That killed exactly the hedged book where his only gross skill lived. The wallet you compete with today is the directional no-flip config — the −4.8¢ book.

The 242-0 artifact — and the C7 surprise

  • Certainty: with a true 17–28%/day market loss rate, a clean 242-0 day is p≈1e-20; day-selection contributes ~nothing. The original number was a measurement artifact, full stop.
  • C7 SURPRISE — the obvious mechanism is REFUTED. The naive story (“the 1000-print cap truncated away his losses”) died under validation: the cap is real and severe (median 2,681 prints/market → ~70% of his fills invisible; ~32% of fills execute as maker → he is visible as taker on only ~9% of fills in capped tapes) — but the cap censors wins and losses equally, and the visible slice actually skews losing (capped-tape fill-level win rate 29% vs true 54%).
  • RESOLVED (§C8) — mechanism NAMED AND REPRODUCED: survivorship by activity-dependent tape truncation (outcome-correlated coverage-depth censoring). The original pipeline re-run on the frozen original tapes reproduced the original output verbatim — 242 fills / 55 markets / 242-242 winner-side, tape byte-identical. Two stages:
    • Stage A — which markets are visible at all. A market enters the wallet aggregate only if ≥1 of his fills survives the last-~1000-print cap; coverage depth is activity-dependent, and activity tracks outcome. His losses are late-flip volume storms — coverage dies ~70s before close while his entries sit ~150s out — whereas wins cruise (coverage reaches ≥127s before close). Result: 0/56 lost markets survive vs 55/129 won — the coverage-depth rule predicts tape visibility with zero errors.
    • Stage B — survivors are tautologically winner-side. The same 55 markets’ late slice is only 71% winner-side; his going 242/242 there by chance ≈ 1e-36. True window record: 129W/56L (69.7%); the tape credited +6,998.
  • The C7 “contradiction” dissolves — both measurements were correct, under different conditionals. C7 conditioned on busy markets at the fill level (visible fills skew losing: 29% vs 54%); C8 conditions on all markets at the market-visibility level (lost markets never appear at all). The same cap censors at two levels that point in opposite directions — name the conditional before comparing rates.
  • Redeem-based proxies cannot explain the 242-0 either (the tape study never used redeems; that artifact separately inflated the lifetime “79.3%”).

The 15m arm — real structure, mechanically dead

History of the claim (kept as a methods lesson): the v1 “+7.5¢ band” was selection lookahead (whole-window share-weighted cushion) — caught by reproduction failure, retracted. The honest band (BTC, |cushion|≥5bps at entry, τ≤300s, first-qualifying-entry, entry-time features only):

  • His achievement is real: +3.75¢/window after our cost model (fill-level +4.46¢), 26/29 days positive, stable across life-halves, dose-response in cushion (win 90.7 / 94.5 / 98.9 at ≥5/≥10/≥15 bps on his entries). Arm-wide btc-15m is null-consistent (+1.8¢, CI [−0.3, +4.0]) — the structure concentrates in late-window cushion cells.
  • The mechanical transfer fails. Population test over all 3,613 windows (data/fifteen_population_btc15m.csv, spot-checked 20/20 against fresh sources): the trigger fires in ~94% of windows; window-selection ceiling ≈ 0 (skipped 87.9% vs entered 87.8% — he doesn’t pick windows); his side/timing beats the mechanical trigger by ~3pp (90.7 vs 87.9); priced at his own prices per matched cushion bucket, the mechanical population EV = −1.1¢, 90% CI [−2.7, +0.6] (the 5–10bps bucket, 49% of fires, is −2.2¢). The blended “+4.5¢” was an aggregation artifact.
  • Verdict for the queued 15m expansion: NO +EV seed from this tape. The value is in a side/timing model — i.e. a fresh, frozen 15m calibration fit on our own recordings, gated at the standard deflated bar. Deploy #1014 (btc15/eth15 observe containers) is the bottleneck. The tape contributes priors only: late-window > mid-window; cushion dose-response exists; eth-15m is adversely selected for him (−$22k); pre-window is dead.

Why this matters — live validation of our net-edge gate

Our validated live rule fires only when

p_side − (ask + 0.01) − 0.07·eff·(1−eff) > 0

(walkforward.rs trade_decision_slipped; live as the EMIT_MIN_EDGE / observe::net_buy_edge gate). The 0.07·eff·(1−eff) term is exactly the fee this wallet pays. His 38-day life is a ~190k-fill natural experiment: he crosses at fair prices with no fee-clearing edge and loses precisely the fee (realized −$87,540 vs model, per-fill corr 0.9915). ⇒ Empirical, large-sample proof that fee-unaware crossing is −EV — a well-capitalized operator ran the experiment for us.

The validation added a second, independent angle — the book-calibration audit: the buyable zone (ask ≤0.90, tight spread) is efficiently priced — favorite realized-win exceeds its ask by +0.8→+1.7pp ≈ exactly the fee in every ask bin (net ≈ 0 everywhere by construction); the apparent 0.95+ “overpricing” is a stale-wide-quote artifact (realized 0.468 at ask≈0.99 with 0.24 spreads — placeholder quotes, no liquidity). This independently validates EMIT_MAX_ASK 0.85 and the net-edge-vs-ask rule from the book side: our +EV comes from the p_cal gate selecting in-cell exceptions, not from broad favorite-buying. The selector kill test closed the loop: no reproducible slice of his flow beats the contemporaneous ask OOS.

Operation shape (for competitor modeling)

  • 14 asset/interval arms, 24/7 automated (no human sleep cycle). **65k/day on the 5m book (≈400× our roll, ≈20,000× our $8 fills).
  • Ladders median ~12–15 clips/market (mode exactly 20 → clip cap; max 104); median 1s inter-clip; 53.9% of multi-clip markets contain identical-size clips (bot fingerprint). Adds into weakness (ladder price drift −0.15).
  • Buy-and-hold-to-resolution — 305 sells vs 189,806 buys lifetime.
  • Iterates: outage/rebuild Jun 5–10 → multi-asset Jun 22–23 → flips OFF ~Jun 29 (current config: directional, no-flip).

Implications for us

  1. Cost model empirically exact — the fee term in trade_decision_slipped matches venue reality on ~190k fills.
  2. Gates re-validated from a second angle — book-calibration audit independently supports EMIT_MAX_ASK 0.85 + the net-edge-vs-ask rule; the τ/p_cal regime remains the only place +EV has ever been demonstrated (our own walk-forward + live ledger). His tape can’t rule on our band either way (power limit).
  3. Oracle-state anxiety closed; certainty-tier poster child gone — the size_hint certainty tier must re-base on our own fills only (it was partly argued from his “100%” record), and per the same-day refit study, on actual-resolution calibrations, never frozen p_cal (which inflates conviction bins 4–7pt). Print-state closed from three directions: Phase-0 detector ≈ ask; his at-ceiling print-following; refit H4 actively harmful.
  4. NEW follow-up: he collects taker rebates (~16% of fees). Check whether OUR wallet accrues them — if yes, realized costs beat the model (free EV, zero code).
  5. Unchanged: sizing go/no-go at the week read (~Jul 9) on our own realized cash; hold-to-settlement; the 15m observe→calibrate→gate queue; all dead ends stay dead.
  6. Not a threat / no edge to copy. He supplies ~8 fills we are invisible to him — if anything he is the liquidity.

Methods lessons + incident (reuse in every future study)

Reproduce before you believe

Three headline claims died on independent reproduction this session: (1) the v1 15m “+7.5¢ band” = selection lookahead (whole-window share-weighted cushion); (2) the blended-price “+4.5¢” population EV = aggregation artifact (→ −1.1¢ at matched cushion); (3) “58% vs 55.3¢” = mixed weighting. The pre-registered 3-leg design (clean-room + adversarial + on-chain) caught every one before it could ship.

  • Capped market-tapes are unusable for per-wallet PnL in BOTH directions (maker-invisibility + recency slice) — the per-wallet-z methodology is retracted; harvest the wallet’s own /activity lifetime instead (crypto-shortterm-polymarket-data-api-gotchas).
  • A directional win rate needs CLOB-truth dominant-side labels, never redeem-presence (hedging inflates it).
  • Name the conditional before comparing rates (the C7-vs-C8 lesson): the same 1000-print cap skews losing at the fill level on busy markets (C7) and winning at the market-visibility level over all markets (C8) — two censoring levels, opposite directions, both measurements correct.
  • Frozen post-settlement tapes are exactly reproducible (10/10 live re-queries row-exact) — freeze tapes at study time; this is what let §C8 reproduce the original 242-0 output verbatim and name the mechanism.
  • Agents write findings-to-disk before digests (session-limit deaths mid-flight cost nothing this time because artifacts survived).

Incident — shared-file clobber (data/0xd02b_activity_full.csv)

dissect_discriminator.py briefly overwrote data/0xd02b_activity_full.csv with an 8h slice (restored from raw JSON; script repointed to data/0xd02b_activity_recent8h.csv). Two validators read the stub during the ~15-min window and produced numbers contradicting fresh measurements. Lesson: check file mtime when a claim contradicts a recent measurement — and give derived slices their own filenames, never a shared canonical one.

Artifacts (repo crypto_algo, not vault)

  • Canon: docs/superpowers/winner-dissection-2026-07-03.md (findings + validation matrix, updated to matrix-complete); erratum prepended to docs/superpowers/competitor-analysis-2026-07-02.md; HANDOVER.md updated.
  • Validation: data/validation_cleanroom.md, data/validation_adversarial.md, data/validation_onchain.md (ALL PASS).
  • Detail findings: data/0xd02b_anatomy_findings.md (behavior), data/discriminator_findings.md (trigger + book calibration + selector kill test), data/fifteen_arm_findings.md (15m arm).
  • Data: data/0xd02b_activity_full.csv (199,184 rows), 0xd02b_activity_recent8h.csv, 0xd02b_5m_losses.csv (1,012), btc5m_winners_backfill_pre_jun8.csv (100% coverage), discriminator_features.csv (24,183), fifteen_arm_features.csv (83,105), fifteen_population_btc15m.csv (3,614) + README.
  • Scripts: harvest_wallet_history.py, dissect_discriminator.py, dissect_15m_arm.py, build_15m_population.py, validate_cleanroom.py, validate_adversarial*.py, validate_onchain.py.

Graph View

Backlinks