Appendix E — Complete result tables

This appendix holds tabulations that support the maps but would interrupt them: winner censuses under both rosters, the evidence-label census, and pointers to the machine-readable layer.

E.1 Winner censuses

Table E.1: Winner census, six-combination display roster (Gaussian and focused DP arms): cells won per loss family. Interpret these counts together with the near-tie and Monte Carlo-decisiveness results in the main text.

Combination	KS	Quantile	Tail	MSEL	Rank
dp_focused.GR	90	68	14	0	35
gaussian.GR	23	29	29	0	29
dp_focused.CB	6	9	10	1	30
gaussian.CB	1	10	3	3	26
dp_focused.PM	0	4	40	66	0
gaussian.PM	0	0	24	50	0

Table E.2: Winner census, full nine-combination roster. The main text uses a six-combination display roster because the focused and broad DP arms frequently tie on the primary contrast.

Combination	KS	Quantile	Tail	MSEL	Rank
dp_focused.GR	47	43	9	0	18
dp_broad.GR	44	24	5	0	12
gaussian.GR	23	29	28	0	21
dp_broad.CB	5	5	5	0	36
dp_focused.CB	1	5	5	1	18
dp_broad.PM	0	2	24	15	0
dp_focused.PM	0	2	18	51	0
gaussian.CB	0	10	3	3	15
gaussian.PM	0	0	23	50	0

The focused and broad DP arms tie in 106 of 120 primary-contrast cells. These counts are read from the generated facts object rather than repeated as caption literals.

E.2 Label census

Table E.3: Evidence-label census for the registered contrasts under the locked label rule. Source: tables/T-census.rds.

Contrast	Strong win	Win	Tie	Flag
CB vs PM (MSEL, Gaussian)	0	0	21	99
DP broad vs DP focused (KS, GR)	0	0	106	14
DP broad vs Gaussian (KS, GR)	47	18	2	53
DP focused vs Gaussian (KS, GR) - primary	45	17	1	57
DP focused vs Gaussian (MSEL, GR) - safety	22	13	26	59
Normal-cell control (KS)	0	0	2	38
Source: cell_evidence_rows.tsv. Locked label rule from the preregistration: win if the bootstrap interval lies below 1 (strong win below 0.95); tie if the interval sits inside [0.95, 1.05]; flag otherwise. The flag column is a residual category dominated by imprecision, not a harm verdict.

E.3 The machine-readable layer

The analytical display layer is generated by scripts in code/R/ from the frozen row-sets in data/tidy/ (9 tabular row-sets plus 1 JSON manifest, 8.0 MB) through the derived layer in data/derived/. Those inventory values are computed from the files at render time. CSV mirrors of script-built display tables are in tables/csv/; figures generated in figures/ have PNG, PDF, and caption sidecars, while the copies used by Quarto require the book build’s equality/hash checks. The condition-level values behind the maps are in data/derived/cond.rds (5,400 rows), and the cell-level contrasts with intervals are in data/derived/ratio_cells.rds (2,160 rows) and the frozen contrast_mcse_rows.tsv. Column schemas are documented in the repository README.

The compact provenance audit in Table B.1 distinguishes row-derived analytical findings from generated design facts, quoted operational facts, and protocol or historical records. It is intentionally narrower than the false claim that every number in the book can be recomputed from data/tidy/.