The null is drawn at the winner's count, by calendar date, with p counted plus one

Beats-null (R4.4) asks whether the winner does better than random entry under the same costs and geometry, at the corrected alpha. Three things make the answer looser than the alpha it is asked at.

The count. The day-boxed engine draws each null value as a mean over the winner's trade count. The position-boxed engine does not: its bootstrap resamples the random-entry pool to the pool's own length, so each draw is a mean over about three times as many positions as the winner holds and the null is too narrow. On a world with no effect, at one cell, the check fires at 0.140 against a declared 0.01 and 0.225 against 0.05. Q-005 and both programme 2 questions were measured on that engine.

The dependence. Both engines treat the draws as independent. The winner's trades are not: a signal fires on many names on the same day, and multi-day positions entered on neighbouring days hold the same days. With names that share a market the day-boxed null, exact when trades are independent, fires at 0.093 against 0.01 and 0.213 against 0.05. The registered questions' own measured same-day correlations run from 0.19 to 0.56.

The count of exceedances. p = exceed / B reports zero when no draw reaches the winner, and nothing at all is recorded when a check passes: Q2-001's resolution carries no p, no null mean, no standard error.

Decision (adopted 2026-10-03)

1. Every draw is a mean at the winner's trade count. A measurement records the count its null and its baseline were drawn at, and both Monte Carlo guards refuse one drawn at any other.
2. The null resamples whole calendar days, in blocks. The winner and the reference it is compared with — random entry with a coin's side for beats-null, the side-matched passive alternative for the fifth check (ADR-0049) — are summed by calendar day, and the same resampled days are used for both. Blocks are runs of consecutive calendar days, long enough to hold the overlap of positions entered on neighbouring days: a declared multiple of the hold, never less than the hold. Trades that share a date stay together; positions that overlap stay together; the reference's own sampling error is in the comparison.
3. The multiple is whatever the size table requires. The build uses the smallest multiple for which every calibration world is within tolerance (ADR-0047 §6), and records it beside the rule.
4. p is counted plus one: (1 + exceed) / (1 + B), everywhere a Monte Carlo p is computed. The draws needed to resolve the corrected alpha are unchanged.
5. A standard error is recorded beside the bootstrap. Clustered by entry date, with a correction across dates out to the block length. When the bootstrap's p and the analytic p fall on opposite sides of the corrected alpha the check refuses, naming both: the two are answers to one question, and a verdict does not rest on the one that agrees.
6. Every check leaves its evidence, pass or fail. A GuardEvidence record per check precedes each resolution, with the winner's count, EV, dispersion and standard error, the p, the reference's mean and spread, the draws and the method.
7. From the next registration, never backwards.

Consequences

• The position-boxed engine's block_bootstrap_means takes a target length; Measurement gains null_n and baseline_n, zero on a measurement made before this.
• Every p-value on a multi-day question widens. A null verdict cannot be revived by it; a pass can become a refusal.
• The synthetic engine draws independent trades by construction and keeps an independent null, which says so.
• The survey's readiness uses the same function as the guard, at the cell's own count.

Considered options

• Resample the random-entry pool at the winner's count, as the review's first fix proposes. Measured: it brings the position-boxed check from 0.140 to 0.015 at 0.01 and leaves 0.100 at 0.05 — a pool whose sides were fixed by one coin carries its own mean's noise into every draw.
• Resample the pool in date clusters shaped like the winner's, as the review's second fix proposes. Rejected: a passive pool laid out one position at a time cannot reproduce positions entered on neighbouring days that hold the same days, which is where multi-day dependence lives.
• A t-test on the clustered standard error alone. Kept as the cross-check of §5, not as the test: the bootstrap needs no distributional form for outcomes that are stopped and gapped.
• Refuse when the two p-values differ by more than a factor of two, as the review proposes. Rejected: at 1e-6 against 3e-6 it refuses a disagreement that changes nothing.

As built (M16.14, 2026-10-03)

The draw is studentised. In every resample the difference is divided by its own standard error, blocks of days as clusters, and the observed difference by its own; the comparison is of the two ratios (Götze & Künsch 1996). A draw then carries the noise of the variance estimate as well as of the mean.

Random entry is its expectation. The reference for beats-null is half the long passive outcome of every admitted box and half the short, summed by day — the expectation of a coin's side, with no single coin's luck in it. The multi-day engine's coin-flip probe is no longer drawn.

The cross-check is the larger p. §5's refusal when the two disagree means a check passes only when the bootstrap's p and the clustered standard error's p are both at or under the corrected alpha. The size table measures that decision.

What the size table reads (200 seeds a row, beats-null alone, at a declared 0.01 and 0.05; tolerance 0.026 and 0.086):

The review's two probes — rows two and three — are inside tolerance at both rates. So is every row at 0.01, the per-cell alpha every question in this lab was registered at.

The multiple is two, and no multiple closes the last two rows at 0.05. §3 asked for the smallest multiple of the hold for which every calibration world is within tolerance. On the multi-day engine, over the same 200 seeds:

A block is two holds of consecutive calendar days, or the vendored rule's length for the daily series when that is longer, and never so long that fewer than four blocks remain.

The residual, measured. ADR-0047 §6 says a fix that leaves a measured size above tolerance is not done. For the multi-day engine on names that share a market, at 0.05, this one is not done, and it is recorded as open rather than closed by choosing seeds:

• How large. The last row's world puts the stop where it cannot bind, so a long position mirrors a short one exactly and a coin's side is the null by construction. It reads 0.095 on the table's 200 seeds, 0.060 on 800 fresh ones, and 0.067 over the thousand, where the tolerance is 0.066. The true rate is about six or seven in a hundred against a declared five.
• Why. Over 400 seeds of that world the studentised difference has a spread of 1.08 where a correct reference has one. Positions entered fewer than a hold apart share days and a market; a block keeps that covariance when both entries fall inside it and loses it when they straddle its edge, and the clustered standard error discounts the same lags. The bootstrap's own 99th percentile matches the statistic's (2.41 against 2.38); its 95th does not (1.69 against 2.04 on those seeds).
• What was tried. Longer blocks: the table above. A standard error that keeps the covariance whole out to the hold (a flat-top kernel): over 400 seeds it moved the shared-market world from 0.083 to 0.077 and the independent one from 0.062 to 0.070, and was not adopted. A reference distribution with fewer degrees of freedom: no change at 0.05.
• What was ruled out. That the entry's count rises after its losses and biases the mean per trade: the two are correlated at −0.49 across seeds, and the bias that gives is two hundredths of a standard error.
• Where the stop binds, part of the excess is not an error. In that world always-short loses 0.0074 R a trade and always-long 0.0018 — the stop bites a short harder — so a long entry with no skill is ahead of a coin's side by a tenth of a standard error. Refusing that is the fifth check's work, and ADR-0049 makes it so.
• What would close it. Re-running the entry on resampled price paths keeps every dependence and is not built: these simulators answer in seconds, not milliseconds.

Both rows are strict expected failures at 0.05 that name this section, and the README lists the residual as a limitation.