Edge Score V3b vs same-window PnL: an in-sample diagnostic, not a forward test

Terminal readout, window one, closed 2026-08-30

The pre-registered window has closed. Everything below is the verdict harness’s own output under the frozen rule, published on close with no one deciding on the day whether to publish it, together with the disclosures written before the outcome was known. Each point estimate carries its interval or its denominator in the same line. A null, a partial or a negative outcome reaches this surface through the same code path and is reported the same way.

Verdict (frozen rule)

persistence_confirmedcomputed by the pre-registered rule, not chosen after the fact; the field-6 sensitivity rerun agreed, so no downgrade applies

Candidate arm, pooled forward edge

+3.83pp95% BCa CI [+2.09pp, +6.16pp] over 139 resampled wallet clusters, effective independent clusters 5.28 by Kish, NOT 139

Control arm, pooled forward edge

+0.90pppre-specified non-cleared control, NOT band-matched

Excess (T)

+2.93pp31 of 10,000 permutations at or above the observed statistic

Evidence

71,529 resolved forward positions139 active wallets; 63 of 178 frozen candidates cleared the 10-position floor

Equal-weight contrast (pre-registered secondary)

+5.79pp95% BCa CI [+3.19pp, +8.86pp] over 1,591 equally weighted wallets, so the nominal count is also the effective one; each qualifying wallet weighted equally, descriptive only, the verdict rule reads the pooled primary

Higher-evidence stratum (frozen in-sample n >= 100)

+3.98pp95% BCa CI [+2.20pp, +6.64pp]; this pool is position-weighted and this artifact does not report its effective independent-unit count, so no cluster count is shown; 60 wallets, 67,717 pooled positions

Lower-evidence stratum (frozen in-sample n 30 to 99)

+1.11pp95% BCa CI [-0.19pp, +3.05pp]; this pool is position-weighted and this artifact does not report its effective independent-unit count, so no cluster count is shown; 79 wallets, 3,812 pooled positions

Independent evidence, Kish effective clusters

5.28nominal 139 active wallets over 71,529 pooled positions; the single largest wallet supplies 41.2% of them. Post-hoc readout added 2026-08-12, not pre-registered, no verdict reads it

Leave-one-wallet-out excess (T)

[+2.49pp, +4.50pp]dropping each of 139 active candidate wallets in turn against the unchanged control arm; minimum above zero: yes. Post-hoc, added 2026-08-12, not pre-registered

Largest single capture day

9,914 positions on 2026073113.9% of the candidate arm arrived in one collection day, pooled edge that day +7.99pp. Capture day is the archive a position was FIRST seen in, not its resolution date, so a large day is an ingestion event rather than accrual. Post-hoc, added 2026-08-12, not pre-registered

Field-6 coverage by resolution month (pre-filter candidate rows)

2026-06 8.7%; 2026-07 44.3%; 2026-08 99.7%share of the candidate arm's in-window PRE-FILTER rows carrying a recorded open timestamp, by resolution month, over 87,706 pre-filter candidate rows: archived rows of frozen candidate wallets BEFORE the micro-market rule, the identity dedup and the quarantine, so this denominator is larger than the qualifying count in the Evidence row and the two must not be compared. The archive only began recording the open timestamp on 2026-07-25, so the field-6 rerun is closer to a recent-evidence subsample than an even cut of the window. Post-hoc, added 2026-08-12, not pre-registered

Field-6 sensitivity, opened and resolved in window

persistence_confirmed95% BCa CI [+2.37pp, +7.89pp]; this pool is position-weighted and this artifact does not report its effective independent-unit count, so no cluster count is shown; covers 41,151 of 71,529 (57.5%) of the candidate arm's qualifying positions, the rest carrying no recorded open timestamp; agrees with the primary

Floor-only reading of both legs (post-hoc)

persistence_confirmed, excess +2.96pp90 of 10,000 permutations at or above the observed statistic; candidate pooled edge +3.87pp, 95% BCa CI [+2.16pp, +6.19pp]; this pool is position-weighted and this artifact does not report its effective independent-unit count, so no cluster count is shown; over 63 candidate and 1,528 control wallets clearing the 10-position floor (71,180 candidate positions); agrees with the primary reading. Post-hoc reading disclosure, added 2026-09-01, not pre-registered

Disclosures, written before this result existed

• One of three pre-specified sensitivity checks runs on only part of the evidence. Field 6 of the pre-registration specifies a rerun restricted to positions both opened and resolved inside the window. Our archive did not record a position open timestamp before 2026-07-25, so the rerun is computed only over the rows that carry one, and the share of each arm it covers is reported in its row above. That coverage is NOT even across the window: because the timestamp only began flowing near the end of it, the rerun leans on the most recent evidence, which makes it closer to a recency subsample than an independent cut. The month-by-month coverage is reported above rather than summarised into one percentage, so the tilt is visible instead of inferable. This is an omission in our archive schema, not a limit of the underlying data: the timestamp was available in the collection pipeline and was not persisted. It was corrected on 2026-07-25. That check is verdict-bearing rather than decorative: where its rerun and the two-leg computation disagree, the pre-registered rule downgrades the published verdict to partial_persistence, and the verdict row above states whether that downgrade fired.
• The pooled estimate is position-weighted, so a wallet contributes in proportion to how much it traded and the effective number of independent clusters is FAR below the nominal wallet count. The row above reports both counts and the single largest wallet's share of the arm; read the effective figure, never the nominal one, as the number of independent units behind the interval. Two further checks are reported above rather than asserted here: the pre-registered equal-weight contrast, which gives every qualifying wallet the same weight regardless of how much it traded and is the filed anti-concentration check, and a leave-one-wallet-out sweep of the excess, which was added on 2026-08-12 and is not pre-registered.
• The pre-registered equal-weight contrast, which gives every qualifying wallet the same weight regardless of how much it traded, is reported in its own row above with its own interval. Where the two differ, the pooled primary is the pre-registered statistic and is the one the verdict rule reads; the equal-weight figure is descriptive and publishes whichever direction it points.
• The pre-registered evidence split is reported above, one row per stratum, each with its own interval and its own position count. The strata are defined by frozen in-sample position count, so the lower stratum carries less forward evidence by construction and its interval is correspondingly wider. Read each stratum against its own interval rather than against the pooled headline. Nothing here claims the effect is uniform across them.
• Forward evidence is supposed to accrue as positions resolve, and some of ours did not arrive that way. A collection day on which a large block of already-resolved positions is first captured is an ingestion event, not accrual, and the largest such day is reported above with its share of the arm and its own pooled edge. It is not a protocol violation: the window, the frozen sets, the exclusion rule and the verdict rule are unchanged, and the same collector serves both arms. It does mean the evidence is lumpier in time than a steady-accrual reading would suggest, and it is disclosed here because a reader who finds a large single-day block after publication is entitled to have been told first. This disclosure was added on 2026-08-12.
• This measures persistence of realized edge over entry prices. It is not validated forecasting ability: realized edge is partly price structure and fill quality, the control is pre-specified but not band-matched, and this is a cohort average that licenses no statement about any individual wallet.

Window one, what the readout means. Written 2026-09-01, after the terminal read

Figures below are fixed to the 2026-09-01 computation (snapshot, sha256 bd18a521...c47b), which is the frozen terminal artifact. The first closed-window computation, on 2026-08-31 (snapshot, sha256 877812c5...7180), read the same verdict: candidate +3.83pp, 95% BCa [+2.09pp, +6.18pp], on 71,486 candidate and 441,347 control positions; excess +2.94pp, 26 of 10,000 permutations at or above it; Kish 5.28; equal-weight +5.78pp, 95% BCa [+3.11pp, +8.86pp]. The floor-only reading below is a dated addendum (sha256 cbf57c28...3448) computed over the same frozen lineage, not an edit to the frozen bytes.

Window one of AsPredicted #294147 closed on 2026-08-30. The rule filed before the window opened returned persistence_confirmed. Five dated notes were added while the window was open, all linked on this page: two corrected the implementation toward the filed rule (2026-06-07, 2026-07-03), one disclosed the parent extract’s record-retention limit (2026-07-27), one disclosed two tooling defects (2026-07-30), one corrected a label (2026-08-14); none changed a registered set, hash, floor, seed, statistic or verdict rule. Start with its weight: of the candidate arm’s 71,529 forward positions from 139 active wallets, one wallet supplied 41.2% and the five largest 62.2%, so the position-weighted headline rests on the equivalent of about 5.28 equally weighted wallets (Kish effective sample size), not 139. That figure is post-hoc, added 2026-08-12 with the window open and an interim result visible, and gates nothing.

What was filed. Does an in-sample-selected wallet set keep a positive realized-edge difference after its freeze? Filed 2026-06-01, the day before the window opened (receipt verified 2026-07-25): 178 candidate wallets that cleared a Benjamini-Hochberg screen at q = 0.10 on the 2026-04-25 tape, 3,693 non-cleared controls, one measure (realized outcome minus the entry price, computed dollar-weighted; see the 2026-08-14 note), two required legs (candidate pool beats control pool by wallet-label permutation at the one-sided 0.025 level; candidate 95% BCa lower bound above zero), floors of 40 qualifying wallets and 1,000 candidate positions. Only the sign and pass rule were registered, no magnitude.

What it showed, as computed 2026-09-01. Candidate +3.83pp, 95% BCa [+2.09pp, +6.16pp], on 71,529 positions from 139 active wallets, 63 of 178 frozen candidates clearing the 10-position floor; both floors met. Control +0.90pp, 95% BCa [+0.64pp, +1.33pp], on 443,880 positions from 2,406 active wallets. Excess +2.93pp; 31 of 10,000 label permutations at or above it. Both legs are computed over every frozen wallet with at least one qualifying position, a reading fixed in the harness on 2026-06-06 before any forward row existed; restricted to the 63 and 1,528 wallets clearing the 10-position floor, both legs also pass (reported in its own row above). The filed both-in-window rerun is approximate: open timestamps began 2026-07-25, so it covers 41,151 of 71,529 candidate positions (57.5%; 84 active wallets, 46 clearing the floor; rows resolved before 2026-07-25 that carry an open timestamp were captured late). Both legs pass there (excess +3.66pp, 62 of 10,000 permutations at or above it; 95% BCa [+2.37pp, +7.89pp]), so the filed partial_persistence downgrade is not triggered.

The concentration objection, and two checks. The filed check for exactly this objection, the equal-weight contrast, gives each floor-clearing wallet one vote (63 candidate, 1,528 control) and still separates the arms: +5.79pp, 95% BCa [+3.19pp, +8.86pp]. It is a different estimand from the pooled excess, so its larger point value is not evidence that the pooled result is stronger than it looks; what it shows is that removing the position weighting does not remove the separation. Concentration limits the pooled headline’s precision; on this check and the leave-one-wallet-out sweep, it does not supply its sign. A post-hoc leave-one-wallet-out sweep (not pre-registered; point estimate only) keeps the excess between +2.49pp and +4.50pp across all 139 drops.

What this does not show. It is persistence of realized edge over entry price, not forecasting ability: realized edge is partly favorite-longshot price structure and fill quality, the control is pre-specified but not band-matched, and the residual test that would net that out is not yet filed. It is a cohort average, licensing nothing about any single wallet: at nominal q = 0.10 about 17.8 of the 178 in-sample clears are expected false discoveries, and the 2026-07-30 amendment on this page says why the true share may exceed it. Evidence over-weights shorter-dated markets, and both arms are conditioned on forward activity. Evidence is admitted on each market’s scheduled end date, the quantity our archive records as its resolution date, which can differ from when it actually settled. That is a deviation from the filed word “resolve”, mitigated but not erased by being frozen, arm-uniform and fixed before the window opened. Recomputed 2026-09-04 on each market’s actual settlement time, taken independently from the venue’s closedTime and the Polygon settlement block (the two disagree on window membership for 1 of the 211,518 markets both cover and on the settlement day for 16; median difference zero hours; the counts here are on the Polygon settlement block, which covers all 211,520): 1,549 of 71,529 candidate positions (2.2%) and 23,213 of 443,880 control positions (5.2%) had settled outside the window they were counted in. On the settlement-defined pool the frozen rule returns the same label, persistence_confirmed, with the pooled contrast at +2.96pp (was +2.93pp), leg A p = 0.0029, and the candidate BCa interval +2.08pp to +6.32pp over 127 active wallets (was 139); the Kish effective-cluster caveat was computed on 139 wallets and has not been recomputed on 127. The 10.7 percent figure previously reported here measured a different quantity, positions first archived before their scheduled day, and is superseded by these counts. Window one’s collector admitted only positions scheduled inside its 90-day window, so long-horizon markets were never in its pool; the same design means positions scheduled outside the window that settled inside it were never captured, so that direction is unmeasured here, and this verdict stays not quotable as a result until it is. The +2.93pp excess is about a third of the +8.7pp in-sample gap the set was selected on, a selection-biased figure quoted from the filing without an interval. The 2026-07-03 mid-window correction quarantined 167 rows, excluded here. The per-arm count of positions first captured after the 2026-07-30 collector fix, promised for this readout by the amendment below, is not yet reported; it will be added to this page when measured, with its date.

Window two is filed: AsPredicted #303724 (receipt verified 2026-07-26) is filed to run the identical design on the frozen sets over 2026-08-31 to 2026-11-28, contiguous and, as filed, sharing no position, new seeds, same floors; its first closed-window computation runs on 2026-11-29, the computation of 2026-11-30 is its terminal read and freezes the artifact (a two-day late-capture allowance fixed on 2026-09-02, before any window-two evidence had been read), and it publishes whatever the rule returns. Window-two collection began on 2026-09-02, two days after the window opened. Resolutions from 2026-08-31 and 2026-09-01 are eligible for recovery on later collection passes: every candidate wallet is fetched on every pass and control wallets on a rotating slice, and the first pass had already recovered 1,076 candidate rows dated to those two days. Not every such row will come back: the collector reads a wallet’s most recent activity, and a busy wallet’s older rows can fall out of it. The recovered count will be reported with window two’s read.

Provenance. This narrative was gated by this project’s automated independent-review agent on 2026-09-01 (pass with required edits, applied), an in-house AI check with a non-overridable veto rather than a third-party or human review, against draft sha256 253706b6...280f; the text published here is draft sha256 30fe3296...c958, which differs by the two filled placeholders, by the edits a second automated review required on 2026-09-02, by the 2026-09-03 correction of the boundary-deviation figure (10.7% on the unit the test scores, replacing 9.1% computed on wallet-market pairs), and by the 2026-09-04 settlement-time recomputation that superseded that figure. The founder, the only human in the chain, approved publication on 2026-09-02. No signature statement is reproduced: no human signed a verbatim text, and this line records what was reviewed and what was approved rather than putting words in anyone’s mouth.

Dated notes for window one, in the order they were written

Every note written for this window while it was open, plus the one reading note filed after its close. None changed a registered set, hash, floor, seed, statistic or verdict rule. The 2026-07-30 amendment is quoted in full in its own block below; the others are restated here from the dated internal record.

2026-06-07

Correctness fix to the shared position aggregation, before any forward readWin or loss on positions derived from SELL fills was being resolved against the wrong outcome index and came out inverted; BUY-derived positions were unaffected. The window had opened on 2026-06-02 and the collector's first successful run was 2026-06-07, under the defect, producing a single archive. That archive was deleted and regenerated under the corrected code, which resolves win or loss against the side actually held. No statistic, frozen set, floor, seed, window or verdict rule changed, and no forward result had been read or published.

2026-07-03

Resolution-screen correction, mid-window, toward the filed exclusionsField 6 of the filing excludes voided, disputed, invalid and ambiguous resolutions. The scoring in use since the window opened had not implemented that intent for markets closing on an exact tie, markets under a UMA dispute, voided markets, or UMA states still pending, so some rows carried an outcome the filed exclusions should have kept out. From 2026-07-03 the screen excludes those cases. Every archived market was re-fetched and re-screened; 168 position keys (0.18% of archived rows) were quarantined, and every verdict computation from that date, including the terminal read, excludes them (167 archived rows matched). No floor, seed, window, statistic or decision rule changed. The pinned methodology hash covers configuration rather than this code path, a gap recorded for future filings.

2026-07-27

Window-one extractor disclosure (record retention)Window one's parent extract retained each wallet's newest 2,000 activity rows, and the public data API demonstrably does not serve complete history for some hyperactive wallets even without that cap (measured example: 2,614 API-served resolved conditions vs 3,912 chain-derived for one high-activity wallet). Positions missing from public surfaces skew toward losses, so in-sample edges on truncated records are biased upward. This window's extractor removes the cap, walks each wallet to end-of-history with truncation recorded per wallet in a completeness manifest, and every provisional candidate is checked against its on-chain record under the frozen rule above. The window-one and window-two Arm B frozen sets are unchanged and their forward tests are unaffected (sets frozen before their windows; measurement arm-uniform); their candidate set should be read as the screen's output on the 2026-04-25 extract as it existed. The defect's expected direction is toward attenuating forward persistence, not manufacturing it; that observation must never be strengthened into a claim that the true effect is larger. The in-sample FDR-guarantee line is softened accordingly: nominal q = 0.10 assumes valid per-wallet records, and loser-skewed truncation weakens that guarantee for the screen stage.

2026-07-30

Selection-extract and collector defects, disclosed with their handlingTwo instances of one defect (a missing page-size parameter on an event-metadata lookup) were found while preparing a later filing, one in the extract builder behind the 2026-04-25 selection and one in the daily forward-position collector. Both are disclosed in full, verbatim, in the amendment block below, together with what each can and cannot do to the forward tests and the follow-up measurements they name.

2026-08-14

Entry-weighting terminology correction, before either window's readThe filings describe the entry price as volume-weighted. The frozen computation, pinned by content hash since before window one opened and unchanged since, is dollar-weighted: each fill's price weighted by the dollars filled at it. The two definitions diverge only when a position's fill prices vary, and the dollar-weighted edge is the smaller. Measured on 150 wallets' raw fills, the divergence in realized edge was median 0.36pp, 90th percentile 1.25pp, maximum 4.06pp, one-way. No code, analysis or archive changed; every verdict operates on the frozen dollar-weighted quantity.

2026-09-01

Reading note: the wallet unit of the two verdict legsField 5 of the filing runs the two legs over "qualifying wallets"; field 7 defines the 10-position floor as the condition for counting toward the qualifying-wallet floor and the equal-weight secondary. The harness, fixed on 2026-06-06 before any forward row existed, runs both legs over every frozen wallet with at least one qualifying forward position and applies the floor where field 7 places it. This is a reading disclosure, not a deviation: no registered object changed and the reading predates the data. To make the other reading checkable, a floor-only rerun over the floor-clearing wallets is reported as a labelled post-hoc row; no verdict reads it.

Registry dataset

The frozen per-wallet registry behind this screen is being prepared for public release and is pending a data-release review. Nothing is downloadable today. Leave an email and the dataset link is emailed after release.