Engineers do not get partial credit for confident arithmetic with the wrong units, or for naming a code that does not govern the problem. That is why PE Prep Pro created PE Precision Bench: a sealed, discipline-balanced benchmark for measuring both engineering answer accuracy and standards precision.
What the benchmark measures
PE Precision Bench gives each model the same 320 closed-book multiple-choice items—exactly 20 in each of the 16 disciplines covered by PE Prep Pro. The holdout includes 284 engineering-analysis items and 36 standards-specific items.
Each response must include a selected answer, a brief calculation or rationale, and canonical code or standards citations when the item calls for them. The official score is 80% answer accuracy and 20% standards citation F1. Unsupported citations reduce precision.
The first official snapshot
In this snapshot, DeepSeek V4 Pro ranked first with an overall score of 97.8%. The leaderboard is most useful as a map, not a trophy: discipline-level results show where model strengths diverge, and standards citation performance can lag behind correct answer selection.
Balanced coverage, limited public exposure
The source themes come from PE Prep Pro’s mobile question bank and discipline map. The official holdout remains sealed, with only aggregate results and a SHA-256 commitment published. Four separate development examples are public so researchers can inspect the task and scoring contract without releasing the test itself. This deliberate asymmetry helps limit contamination in future benchmark runs.
What the results do—and do not—mean
The benchmark measures one constrained, zero-temperature, no-tools protocol. It does not establish engineering competence, replace official codes, or determine whether a model is suitable for a project decision. Real engineering work still requires current governing documents, project context, review, and accountable professional judgment.