The floor is cleared by a lower confidence bound, and power is planned against a declared alternative

guards/clears_floor.py refuses when the winner's EV is below the declared floor. The winner is the best of k cells, so its point estimate is biased upward, and a true effect sitting exactly at the floor passes about half the time. The power plan (hypothesis.py) sizes N to tell the floor from zero, two-sided: nothing is sized to show an EV above the floor. The dispersion the plan uses is the survey cell's or the template's, while the winner may be any cell of the sweep — Q2-001 planned at 2.22 R on its hold-5 cell and won at hold 20, where its ungated sibling measures 4.09 R.

This is the one fix that changes what supported means, and the author adopted it by name.

Decision (adopted 2026-10-03)

1. The floor is cleared by the lower confidence bound. The check refuses unless winner.ev − z(1 − alpha/k) · se is at or above the declared floor, where k is the declared sweep size and se is the winner's standard error clustered by entry date with a correction across dates out to the hold (ADR-0048's standard error). Simultaneous bounds at alpha/k cover all k cells, so the bound is valid after choosing the winner. The frequency half of the floor is unchanged.
2. Power is planned against a declared alternative. A question declares, at registration and beside its floor, the EV it wants power at — strictly above the floor, with no default, printed by --schema as a key and never a value. Required N is one-sided: ceil(((z(1 − alpha/k) + z(power)) · sigma / (alternative − floor))²). The vendored core/power.py is untouched; the formula lives in this lab's code.
3. The plan uses the widest cell. A question registered from a survey plans on the largest dispersion among its sweep's cells, naming the cell.
4. The winner must be powered at its own dispersion. At measurement the required N is recomputed with the winner's own standard deviation and its effective count; a winner too dispersed for its count is refused. This reads the measurement partition and can only refuse.
5. The controls plant at an apparatus alternative. controls.toml declares it — an apparatus number, never the author's — and S10 becomes a planted effect at the alternative is accepted.
6. From the next registration, never backwards. A question registered before this carries no alternative and is read as such; a re-score reports its bound and judges it against its recorded floor (ADR-0047).

Consequences

• Fewer questions are affordable: at a corrected alpha of 0.01 and power 0.8, a gap of 0.10 R between alternative and floor needs 1,446 trades at a dispersion of 1.2 R and 4,951 at 2.22 R. That is what clears the floor costs when it is meant.
• Q2-001's bound at the plan's dispersion is +0.151 with independent trades and below the floor under any clustering; the re-score records it.
• The quickstart's known answer (714) stays as the M0.15 figure, labelled as the rule before this one.

Considered options

• Keep the point estimate. Rejected by the author's adoption: it attests the best of k draws, not the cell.
• A separate winner's-curse correction. Rejected: the simultaneous bound already covers selection among the k cells.
• A default alternative. Rejected: a default is a recommendation (R9).

As built: the dispersion half (M16.16, 2026-10-04)

• §3, the widest cell. A question registered from a survey takes its plan's dispersion from the largest among its family's cells that the survey measured and did not refuse, and its provenance names that cell. prepare prints both numbers. On the test fixture the chosen cell measures 0.49 R and the family's widest 1.02 R, and a floor that was powered at the first is refused as underpowered at the second.
• §4, the winner's own dispersion. The engines put the winner's standard deviation, the standard error of its EV clustered by date, and its effective count — the count of independent trades that standard error is worth, never more than it holds — in the Measurement. At REGISTERED → MEASURED the required count is recomputed on the question's own formula at the winner's standard deviation and refused as underpowered at the winning cell's own dispersion when the effective count is below it. It reads the measurement partition and can only refuse.

As built: the bound and the alternative (M16.21, 2026-10-04)

• §1, the bound. clears_floor refuses unless winner.ev − z(1 − alpha/k) · se is at or above the declared floor; se is the winner's standard error clustered by date, which the engine puts in the Measurement. A measurement with no standard error is refused by name. The frequency half is unchanged.
• §2, the alternative. PowerPlan.alternative_ev_net_r; a registration without it is refused, naming the key and no value, and so is one whose alternative is not above its floor; --alternative-ev on both registration commands, no default; prepare says UNDECLARED. The required count is inference.one_sided_n. The review's figure checks: 1,446 at 1.2 R, a gap of 0.10 R, a corrected alpha of 0.01 and power 0.8. python -m occams --schema lists the four keys a registration declares — the floor's pair, the alternative, the plateau's slack in standard errors — each with its type and none with a value. whatif prints the count for four gaps beside the count under the rule before.
• §5, the controls. controls.toml declares an apparatus alternative of 0.30 R, twice the control's floor, and the signal control plants there net of the spread: 732 trades a cell required, 1,000 held. Before this the signal planted 0.165 R before costs — 0.115 net, under its own floor — and was accepted because the best of nine cells cleared 0.15 by its point estimate. That acceptance was the defect.
• §6, never backwards. A question with no declared alternative is judged by its point estimate, as it was registered to, and its bound is written into the evidence beside it. A re-score asks for the bound by name.
• What the size table reads. With the true EV exactly at the floor the bound is cleared 0.010 of the time at a declared 0.01 and 0.065 at 0.05, over 200 seeds. With the true EV at the alternative and the count the plan asks for it is cleared 0.75 of the time against a planned 0.80, over 100 seeds, inside tolerance. Across seeds the null control is accepted by all five checks 0 times in 200 on either engine; the signal control 92 times in 100 on the synthetic law and 100 in 100 through the day-boxed engine.
• What it costs is real. A survey family that was powered at its chosen cell and a floor told from nil is not powered at its widest cell and a floor to be cleared from above. That is what the consequence above says, measured on the fixture.