Terminal X’s Eval Agent, powered by GPT-6.1 Sol, agreed with senior human reviewers on investment recommendations 88.6% of the time in a recent internal evaluation, the company reports. That result came from 500 scored judgments and outperformed the next-best model, which posted an 86.4% agreement rate — a 220-basis-point gap. Alongside higher agreement, the GPT-6.1 Sol-based Eval Agent flagged roughly 25% more recommendations that diverged from a fund’s mandate and executed equivalent workflows at about 30% lower cost per query.
The Eval Agent is a central component of Terminal X’s Capital Memory operating system, which the company describes as a way to capture and reproduce the decision-making processes of a firm’s top investors. Terminal X’s engineering teams—many of whom previously worked at firms including JPMorgan and Blackstone—work onsite with clients to reverse-engineer often-unwritten practices, such as credit desk screening rules or the primary sources used by co-investment teams. Those processes are encoded into a library of agents and stored as versioned firm Memories inside a private data room.
Agents are organized by asset class, strategy and investment stage, and Terminal X selects models by task. According to the company, GPT-6.1 Sol is used for stages that require deep, long-context judgment—synthesizing investment theses, identifying key considerations and assessing fit to a fund’s mandate—while GPT-6 Luna is applied to higher-volume tasks where token efficiency is a greater priority than extended context. Each Memory records past analyses, committee decisions, workflow instructions and house views; engineers then turn these Memories into repeatable, auditable workflows.
The Eval Agent grades outputs using a 54-criterion rubric that mirrors how a vice president would review an analyst’s memo. Criteria cover factual sourcing, numerical accuracy, logical coherence and alignment with the firm’s house view. For each criterion the system records a pass or fail. Failed checks include a one-line explanation and a corrected reference or corrected text, and these failures are captured as new Memories or workflow instructions so the system learns from corrections. “Every answer gets eval’d, and whatever the analyst corrects teaches the system a little more about how that firm works,” said Ken Lee, Co-Founder and CTO of Terminal X.
Terminal X says model selection and workflow design are informed by its own benchmarks and the specific demands of tasks. The company highlighted a multi-pass review workflow made possible by GPT-6.1 Sol’s long-context processing and token efficiency: an initial thesis is tested against the fund’s mandate and recent deals, uncertain verdicts are re-run, and the agent reconciles results before reporting. Terminal X reports that this multi-pass approach identified issues that a single pass on a comparable model would have missed.
Hyun Hong, Co-founder and CEO, emphasized the operational value: “GPT-6.1 Sol’s long-context reasoning and token efficiency allow us to build deeper evaluation pipelines. This helps our Eval Agents make investment judgment measurable and auditable.” Terminal X also notes a cost advantage: the Eval Agent running on GPT-6.1 Sol delivered the same workflows at roughly 30% lower cost per query compared with the next-best configuration.
The platform has been deployed at scale across dozens of clients, the company says. Terminal X reports it has built 1,700 firm Memories and executed more than 1 million custom workflows. Weekly active users reached 81% of monthly active users, a sign of frequent engagement with the platform. One private equity client estimated $2.4 million in annualized savings from automating due diligence and investment memo workflows.
Looking ahead, Terminal X plans to extend Memories from desks to individuals, capturing each investment professional’s practiced standards as well as documented rules. The company frames its mission as converting institutional judgment into data that can be measured, preserved and reused: “The firms that win the next decade will own their judgment as data. Our job is to make that judgment measurable, so it can be kept, improved, and reused,” Hyun Hong said.
Terminal X’s results highlight how specialized model selection, a rigorous rubric and versioned firm Memories can combine to drive agreement with senior investors and operational efficiencies. By encoding firm-specific judgment and continuously learning from corrections, the company positions its Eval Agent as a way to make investment decision-making more auditable and repeatable without replacing human oversight.
Source: Read the original source
Leave a Reply