GPT-5.6 Sol led the reconstructed reference-based scores

Figure 08 · Reconstructed audit quality

GPT-5.6 Sol led the reconstructed reference-based scores

Active researchReconstructed scoring completeFrozen2 retained non-successesOfficial judge not run

GPT-5.6 Sol · HighGrok 4.6 Build · High

Micro reference recall

Codex

40.31%

Grok

36.43%

Macro audit recall

Codex

46.06%

Grok

41.81%

Award-weighted recall

Codex

6.814%

Grok

4.779%

Closed-world precision

Codex

50.49%

Grok

45.63%

Closed-world F1

Codex

44.83%

Grok

40.52%

Figure 08Intention-to-treat quality across 36 scheduled runs per system and 129 reference opportunities per system. Higher is better. All displayed cells are fully adjudicated.

Codex-reconstructed manual, not official EVMbench grading. Two independent Codex reviewers scored frozen outputs; a third adjudicator resolved 1 of 258 reference decisions and 16 of 211 finding decisions. Conventional accuracy is undefined.

Share figure

Read the full benchmark