Terminal X's Eval Agent now matches senior investor judgment 88.6% of the time, running GPT-6.1 Sol against a 54-criterion rubric to grade fund outputs. The system catches 25% more mandate deviations than prior methods. That number matters because missed deviations in fund mandates are compliance failures, not editorial notes.
The rubric design is where this piece earns a full read. Fifty-four criteria is not a round number chosen for comfort. The authors detail how each criterion maps to specific investor review heuristics, and the gap between 88.6% match and 100% is documented honestly, with failure case breakdowns that show where the model still defers when a human would push back.
The outstanding question is what happens at scale across fund strategies with conflicting mandates. The article raises this without resolving it. That tension, between a single rubric and the actual diversity of institutional mandates, is the argument worth following into the next iteration.
[READ ORIGINAL →]