Terminal X puts a number on investment judgment: its GPT-6.1 Sol reviewer agrees with a senior 88.6% of the time

Startup Terminal X builds agents for investment firms that are supposed to keep the judgment of their best people inside the firm. According to a case study on OpenAI's blog, its Eval Agent running on GPT-6.1 Sol agreed with a senior human reviewer on 88.6% of 500 scored judgments. The next-best model reached 86.4%.

Terminal X splits a fund's decisions into hundreds of agents and one reviewer

Terminal X calls the problem “memory debt”. When a senior investor leaves, the firm keeps the documents but loses the reasoning behind them. So Terminal X sends forward-deployed engineers, former JPMorgan and Blackstone professionals, to sit with investment teams for weeks. They reverse-engineer trusted memos and write down processes nobody ever documented, such as how the credit desk screens a borrower.

The result is hundreds of agents organised by asset class and strategy. GPT-6.1 Sol handles the steps that need judgment and long context: the investment thesis, key considerations and fit with the fund's mandate. The cheaper GPT-6 Luna runs the high-volume steps. The agents work inside the customer's private data room, on-premise or in its own cloud, and read documents in English, Korean, Japanese and Chinese.

An Eval Agent checks every output against a 54-criterion rubric. Are the numbers and sources right, does the argument hold, does the recommendation match the house view? Each fail gets a one-line reason and a correction, and is stored as a new instruction or test case. On GPT-6.1 Sol the reviewer caught about 25% more departures from the mandate and, according to the company, cost roughly 30% less per query than the next model in line.

For funds, the audit trail matters more than the agreement score

It is tempting to read 88.6% as “AI invests like a senior partner”. It measures something narrower: how often the review agent reached the same verdict as a senior human reviewer. That still has value. A fund that can show why an agent passed a memo and which sources it used holds the kind of record investors and compliance teams will ask for.

The second layer is commercial. Terminal X is selling the conversion of unwritten experience into data the firm owns, more than any model. That also changes the junior analyst's job: every correction becomes a signal the system learns from. A fund becomes less exposed to individual departures, but it needs a vendor that keeps its Memory current.

The numbers come from the vendor and the runner-up model stays unnamed

This is a case study published by OpenAI, so both companies have an interest in it. All figures are internal: 500 judgments, a 220 basis point gap and $2.4 million in annual savings at a single private equity firm. The study does not say which model came second, how many seniors did the reviewing or how cases were selected. And agreement with a reviewer says nothing about whether the recommended deals would have made money.

Mandate drift, when a buyout fund starts recommending distressed credit, is exactly the kind of error a human misses late on a Friday. It is good that the agent catches it more often. But the study does not examine the remaining 11.4% of cases where agent and human disagreed, so we do not know who was right in those disputes.

The real test is the first partner who leaves

Terminal X reports 1,700 firm Memories, more than 1 million custom workflows and weekly active users equal to 81% of monthly actives. The next step is a Memory for each individual investor rather than each desk. The real test comes when someone whose judgment the system was meant to capture leaves a client fund, and the successor has to decide whether to trust the agent over their own notes.

Independent data would be more convincing: agreement with several reviewers, an analysis of disagreements and comparisons with named models. Until then, 88.6% remains an interesting number from a sales deck.

Lilith's verdict

A departing partner carries a box of photos out of the office, and Terminal X promises the judgment stays on the fund's servers. So far that judgment has been measured by one reviewer on 500 cases, with nothing said about the 11.4% of disputes.