Independent analysis audit for focused protocol v0.6

Scope: scientific source review and a complete hand-computable synthetic export. No main-study estimator outputs were opened. No protocol settings or production scientific code were changed.

Finding: no concrete scientific implementation defect found.

The audit checks E1 truth subtraction, mean error, sample SD and RMSE; E2 independent A/B replicate alignment, true differences and ratios, both sample-size directions, and both cross-geometry comparisons; common-78 P1 on both sides of 500/500 comparisons; selected-parent B100/B500 pairing; whole-parent profile bootstrap, endpoint/range formulas and classifications; missing groups and nonfinite failure handling.

The synthetic export contained 50,028 rows (20 illustrative replicate IDs, not simulations of the scientific models). Result-table counts exactly match the full planned analysis: calibration 1,474; profiles 352; comparisons 3,168; paired P1 changes 1,584; paired B500 changes 1,474. The synthetic bootstrap used 99 draws only to verify implementation; production remains 9,999.

Interpretation: E2 mean errors and ordering summaries are descriptive per unit; comparisons sharing parents are not independent pieces of evidence. A profile can be flat while absolutely miscalibrated. Bootstrap intervals cover within-profile contrasts approximately, without a full-grid simultaneous guarantee. P1 paired MSE-change intervals should support precision or superiority statements; reductions in mean bias alone do not establish superiority.

Reporting note: permutation means and conditional permutation Monte Carlo SE are retained in parent JSON checkpoints rather than the estimates-only long CSV. They should remain available in the reproducibility archive and be extracted if numerically reported in the manuscript. Exact source hashes and seven audit checks are recorded in focused_analysis_independent_audit.json.
