Reconstructed effects and uncertainty

Table 04 · Paired inference

Active researchReconstructed scoring completeFrozen2 retained non-successesOfficial judge not run

GPT-minus-Grok paired effects across 12 audit clusters with hierarchical-bootstrap intervals and exact audit-level sign flips.
Paired reconstructed endpoint	GPT minus Grok	95% CI	Exact p
Award-weighted recall, primary	8.83 pp	-6.88 to 25.48 pp	0.1484
Micro reference recall	4.26 pp	-5.32 to 14.35 pp	0.3359
Closed-world precision	6.95 pp	-11.17 to 21.94 pp	0.4150
Closed-world F1	7.01 pp	-2.71 to 16.30 pp	0.1055

Table 04GPT-minus-Grok paired effects across 12 audit clusters with hierarchical-bootstrap intervals and exact audit-level sign flips.

The primary interval crosses zero. These are reconstructed manual estimates, not official EVMbench results.