The fifth check is side-matched and dependence-aware

ADR-0043 made beating always-long the fifth check so a supported verdict attests the entry, not the gate and the side. As built it does not do that. The baseline is always long, whatever sides the winner took, and its draws are means of independent boxes. On the day-boxed null control the synthetic paths drift down, a coin flip is short half the time, and the coin's EV of −0.080 beats always-long's −0.103 with p near 0.005: a coin flip passes the fifth check — make null --engine=day_boxed has printed checks that passed: plateau, leave_one_out, beats_always_long since the check was added, and nobody read it as a defect. On a world with no effect and a long-only entry the check fires at 0.065 against a declared 0.01.

Decision (adopted 2026-10-03)

1. The baseline takes the winner's side mix. The passive alternative is a market entry on every box the gate admits, under the same exits, stop and horizon — long and short — and the baseline weights the two by the winner's own share of each side. For a long-only entry it is always-long, as before. For a coin flip it is the coin's own expectation, and the check refuses.
2. The margin is measured against that baseline everywhere. The surface the sweep optimises (ADR-0045) and the fifth check use one number; a cell records which rule its baseline followed, so an older measurement reads as always-long.
3. The test resamples whole calendar days. The winner and the baseline are recomputed together on the same resampled days (ADR-0048's bootstrap), so trades that share a date, positions that overlap, and the baseline's own sampling error are all in the comparison. p is counted plus one.
4. The margin's two parts are recorded. Selection is the passive outcome on the boxes the winner chose, less the passive outcome on every admitted box; execution is the winner's outcome less the passive outcome on the same boxes. They sum to the margin. For a market entry execution is exactly nil and the whole margin is selection.
5. From the next registration, never backwards. A re-score reports what this rule would have said (ADR-0047).

Consequences

• S3's output changes once, and this ADR names the change: on the day-boxed engine the null control is now refused by beats-always-long as well.
• Q2-001's margin of −0.001 gives a p near one half under any baseline; the re-score records it. The fifth check was not in force when it resolved.
• A cell's baseline_ev means the side-matched baseline from here on.

Considered options

• Pair every trade with the passive trade on the same box, as the phrase paired on shared boxes suggests. Rejected as the whole test: for a market entry the two are the same trade and the difference is identically nil. Kept as the execution part.
• Compare with the passive outcome on the winner's own dates only. Rejected: an entry's contribution is largely which dates it picks, and that comparison removes it.
• Resample the passive pool with the winner's clustering, as the review proposes. Rejected for ADR-0048's reasons: a pool cannot reproduce positions that overlap across neighbouring entry days.

As built (M16.15, 2026-10-04)

One helper gives the baseline everywhere. probes.baseline_of runs the long passive probe for a cell, and the short one only when the cell holds a short trade; the engines' sweeps, the definition surface shown before alpha moves and the survey's readiness all take the baseline from it, so the surface and the guard cannot hold two numbers. A long-only cell's baseline is always-long, number for number: every cell of every long-only world in tests/test_probes_unchanged.py is what it was.

The comparison is ADR-0048's. The winner and its side-matched baseline are summed by calendar day and resampled together, studentised, at the winner's count; the clustered standard error is read beside it and the check passes only when both agree. A baseline drawn at any other count is refused.

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

Every row is inside tolerance at both rates. The third is the world where beats-null lets a long entry with no skill through one time in ten at 0.05 (ADR-0048, As built): the stop bites a short harder there, so being long is ahead of a coin. This check is the one that refuses it.

It still sees what is there. The signal control's planted reversal passes the fifth check in 100 of 100 seeds at the control's corrected alpha of 0.05 / 9, against a planned 0.80.

S3 changed once, as §Consequences said. make null --engine=day_boxed is refused by beats-null, the floor and beats-always-long, and passes the plateau and leave-one-out. make signal is accepted on both engines.

Execution is simulated only where it can differ. For a market entry it is nil by construction and nothing is run. For a resting order the passive market entry is simulated on each of the winner's boxes, same side, and the difference averaged; selection is the margin less that.

The refusal reads as it judges: the passive alternative at the same geometry, gate and side mix does as well. Records made before this say being long at the same geometry and gate does as well, and stand.

The survey's readiness uses the guard's comparison at the cell's own count, on the cell's trades re-run now. It still tests each candidate alone: twenty candidates are the best of thousands of cells, and nothing corrects for that here (the review's F08; the successor lab's).