A plateau is judged in standard errors, without the winner in its own median

guards/plateau.py asks that the winner's score exceed the median of its neighbourhood by no more than plateau_slack. Two things loosen it. The neighbourhood includes the winner, so with four cells the winner is one of the two middle values and pulls the median toward itself. And the slack is absolute — 0.10 R for every question so far — whatever the surface's scale and whatever the winner's own sampling error.

Decision (adopted 2026-10-03)

1. The winner is not in its own median. The median is taken over the neighbours that traded, the winner excluded. The neighbourhood's size is still counted with the winner in it, as plateau_cells was declared.
2. A slack in standard errors is declared with the question. A new registration declares plateau_slack_se; the check refuses when the winner's score exceeds its neighbours' median by more than that many of the winner's standard errors (ADR-0048's). It has no default.
3. The absolute slack stays. A question is refused if either slack is exceeded; the two together can only refuse more.
4. From the next registration. A question registered before this carries no plateau_slack_se; for it the standard-error half is absent and the gap in standard errors is recorded as evidence, not judged.

Consequences

• A lone spike that the old median hid — the winner 0.12 above its neighbours' median with a slack of 0.10 — is refused.
• The check's evidence gains the neighbours' median, the gap, and the gap in standard errors.

Considered options

• Replace the absolute slack. Rejected: on a surface whose errors are tiny, a slack in standard errors alone would refuse differences that do not matter in R.
• A fixed number of standard errors. Rejected: it is a number, and the question's author declares its gates.

As built (M16.16, 2026-10-04)

• The standard error is the score's. On the margin surface it is the standard error of the winner's margin over its passive alternative — the fifth check's comparison (ADR-0049) — and on the EV surface the winner's own, both clustered by date (ADR-0048). The engine supplies them in the Measurement; the guard reads nothing else.
• A nil standard error is a standard error. A winner that is its baseline has a margin of exactly nil and no error in it; its gap is nil and the check passes. A question that declared the slack and a measurement that carries no standard error at all is refused by name.
• A single cell has no neighbours to be a spike above, and passes; the neighbourhood's size is still counted with the winner in it.
• Registration refuses without it. --plateau-slack-se on both registration commands, no default; prepare says UNDECLARED and is not clean. The apparatus controls declare four, an apparatus number: the best of nine cells drawn alike sits a standard error or two above its neighbours' median by chance.
• Every plateau evidence record carries the gap in standard errors whenever a standard error exists, whether or not the question declared the slack — which is what the re-score reports for the six questions.