Evidence · 30 · Measure & Manufacture

Slide 30 headline results

Prescribed pairs win almost everywhere

234 base × benchmark comparisons, each arm scored per benchmark. No averaging to hide behind.

13 / 13

training sets gain geometric accuracy

12 / 13

gain semantic accuracy

79%

of 234 comparisons: a prescribed arm is strictly best

98%

within noise of the best

Method	Strict best	Within noise	Scored
Base (added nothing)	8 (3%)	91 (39%)	234
Random SDF — diagnosis withheld	8 (3%)	79 (34%)	234
Own Real — more of itself	9 (7%)	45 (36%)	126
FlyT Random	7 (4%)	100 (51%)	198
FlyT Selected — diagnosis → retrieval	19 (9%)	112 (52%)	216
Prescribed	111 (47%)	219 (94%)	234
Prescribed + Deform	73 (31%)	218 (93%)	234
Either prescribed arm	184 (79%)	230 (98%)	234

Table 3 of the paper, verbatim. The gains replicate on GLU-Net: wins on all three bases tried, both accuracy families.

Interactive · try it hereOpen standalone ↗

The Full Tables

Every results table from Chapters 2, 3 and 4, including the per-benchmark and per-threshold tables the slides summarise. Bold and green follow each paper's own marking; the caption under each table says what they mean.

Chapter

Table

Chapter 3 = "Cover the Motion, Not the Look" and Chapter 4 = "Just What the Doctor Ordered", converted directly from the papers' LaTeX table sources; Chapter 2 tables come from the frozen scoring suite and the challenge reports. Tables marked "not in the submitted paper" are working tables kept for the defense.