# DEVIATION D7 (addendum to D6): arm-P secondary is run as §4 orders

Written: 2026-10-28 06:45:05 MDT (28 October 2026), BEFORE the arm-P secondary run. scores/secondary_P.json did not exist when this was written.
Clarified by: Abe (prereg owner) at 06:44 MDT, 28 October 2026: D6's "the 54-cell version will NOT be run" meant
"do not swap the 54-cell analysis in as the arm-P primary (or for its planted/shuffle controls)". It did not cancel
the §4 secondary. Recorded by Grok Bot (executor) for Ford.
Applies to: PREREG_HLE_B2_v2.md (LOCKED, fabb2cc3dcf571d2b412f0d221404529cef9b4c798cc98993cd6ade6b0973680), incorporating
v1 §4; supersedes only the "arm-P secondary HELD" paragraph of DEVIATION_D6_ARM_P_STRATUM.md
(sha256 01a00c90f69ccac29c5d08f251128ca4dda2c162ed34045b1859915439be7af6). Everything else in D6 stands.

- The arm-P secondary (explicit --stratum-key category; 6 pooled + 48 category cells = 54 cells, Westfall-Young
  adjusted within itself at alpha 0.025) is run exactly as v1 §4 orders.
- It is localisation only ("secondary, for localisation", §4).
- It is run after the primary verdicts are known (arm G PASS BORDERLINE, arm P PASS; batch-2 v2 PASS), which is
  where §7.6's fixed order puts it (primary per arm, then secondary per arm).
- It cannot change any verdict, in either direction (§4: "The secondary runs never change the batch verdict").
- The arm-P planted, shuffle and primary runs stay the pooled 6-cell runs of D6; none is re-run.

Exact command (DGX, cwd ~/honesty-tools/runs/epoch-crosswalk/batch2/v2, nice -n 19 ionice -c3, python3):
  python3 ../code/nulltest_paired.py data/B2P_pairs.jsonl --pair-key id --alpha 0.025 --json --jobs 4 --stratum-key category > scores/secondary_P.json
Start/end and output sha256 are appended to logs/SCORE_RUN.log.