Terminal X's Eval Agent Matches Senior Investor Judgment 88.6% of the Time With GPT-6.1

Terminal X has published a case study with OpenAI describing an Eval Agent built on GPT-6.1 Sol that scores investment work against a 54-criterion rubric. The headline number: its judgments agree with senior investor judgment 88.6% of the time.

The second number matters as much as the first. Terminal X says the agent catches 25% more deviations from a fund's mandate, which reframes the product from a drafting assistant into a control layer rather than a faster way to produce documents.

Rubrics are the mechanism. Instead of asking a model whether an analysis "looks right", the Eval Agent checks a stated list of 54 criteria and returns a graded verdict on each one. That structure is what makes an automated judgment auditable: a disagreement with a human reviewer points to a specific criterion rather than a vague objection.

For funds, mandate compliance is exactly the kind of task where generic model quality is not enough. A memo can be fluent and still recommend a position the fund is not allowed to hold. Catching that is a rule-shaped problem, and rules are what a rubric encodes.

The 88.6% figure should be read carefully. Agreement with senior investors is a measure of imitation, not of correctness; if the senior reviewers share a blind spot, the agent will reproduce it efficiently. The vendor's own framing is about matching expert judgment, not about outperforming it.

The economics are the quiet argument. Senior investment review is expensive, scarce, and slow, and it does not scale with deal flow. An agent that reproduces most of that judgment at machine speed changes how much work a small team can supervise, even if a human still signs off.

What to watch is whether the rubric-based approach generalizes beyond one fund's mandate. Rubrics are portable in principle, but the criteria that define a strategy are often tacit, and the hard part of deploying this pattern elsewhere will be writing the 54 lines, not running the model.