Terminal X Encodes Investment Judgment with GPT-6.1

Terminal X claims to have aligned its AI agents with the judgments of senior investors, achieving an 88.6% agreement rate measured across 500 cases, with request costs reported to be approximately 30% lower than a competing model. The system relies on versioned memos of each client's methods and an evaluation agent scoring 54 criteria, operating within private data rooms.

Quantified Results, Reported Uses, and Announced Expansion

Terminal X reports the creation of 1,700 corporate memos and the execution of over one million customized workflows. The company also specifies that 81% of its monthly active users are active each week. Among the 500 rated judgments, it measures an 88.6% agreement between its evaluation agent on GPT-6.1 Sol and a senior human examiner, compared to 86.4% for the next best model. Terminal X further announces a cost per request approximately 30% lower with GPT-6.1 Sol, for a similar speed, and cites the case of a private equity client mentioning $2.4 million in annualized savings due to the automation of due diligence and memos. The company plans to extend office memos to individuals. Its CEO, Hyun Hong, asserts that high-performing companies "will own their judgment as data," and that the platform's role is to make this judgment measurable, preservable, improvable, and reusable.

Auditable Judgment by a Verification Agent and 54 Criteria

The verification of responses is ensured by an evaluation agent, orchestrated on GPT-6.1 Sol and other OpenAI models, which reviews each output as a VP would assess an analyst's work. This agent applies a framework of 54 criteria, examining, among other things, the provenance of figures and citations, the coherence of the narrative, and alignment with internal viewpoints. Each check produces a pass or fail result; in case of failure, a concise reason and a better source or corrected text are provided and become a new entry in Memory. Typical errors include an omitted source, a faulty comparison, or a thesis diverging from the internal viewpoint. Evaluations are traceable, with each production request being auditable along with its history of steps, sources, model, and reasoning.

Architecture: Encoding Team Methods into Specialized Agents

Capital Memory aims to reduce memory debt by encapsulating investment methods into agents. For each client, agents are developed to replicate the investment practices of the company's top professionals. Terminal X teams spend several weeks on-site to reverse-engineer memos and capture informal processes, before transposing these methods into several hundred agents tailored to each asset class and strategy, with sub-agents for each step. GPT-6.1 Sol handles judgment and broad context steps, while GPT-6 Luna manages high-volume steps. Learnings are stored as versioned Memory, which compiles past analyses, committee decisions, workflow instructions, and internal viewpoints, then transforms into repeatable workflows.

Deployment and Data: Private Room, Multilingual, On-Site or Cloud

Agents can be deployed on-site at the client or in Terminal X's cloud. They operate within a private data room dedicated to sharing and managing sensitive investment data. The processes cover documents in English, Korean, Japanese, and Chinese. OpenAI models, including GPT-6.1 Sol and GPT-6 Luna, ensure reasoning within these agents.

Model Selection and Detection of Investment Style Drift

Terminal X selects its models based on its own evaluations. Among 500 rated judgments, the evaluation agent armed with GPT-6.1 Sol achieved an 88.6% agreement with a senior examiner, compared to 86.4% for the next best, representing a 220 basis point difference. The company indicates that GPT-6.1 Sol particularly outperforms in detecting investment style drift, for example, when turnaround scenarios lead to recommending distressed credit or rescue financing instead of a transaction aligned with the buyout mandate. The agent on GPT-6.1 Sol detected about 25% more such discrepancies. The efficiency of tokens allows for a deeper review process at a constant budget: testing the thesis against the mandate, then against recent transactions, re-evaluating uncertain verdicts, and reconciling results. This level of review has flagged transactions that single passages on a comparable model validated.

From Human Capital to Memory Debt: Formalizing Reasoning

Terminal X describes a competitive advantage built operation by operation but reliant on the judgment of a few seniors, hence a risk of memory debt upon their departure. To address this, its engineers, notably from JPMorgan and Blackstone, work with investment teams to encode methods and criteria within agents. According to the company, GPT-6.1 Sol helps the evaluation agent detect about 25% more recommendations that deviate from a mandate, for a cost per request approximately 30% lower than the best alternative model. Hyun Hong emphasizes the long-term reasoning capability and token efficiency to make evaluations deeper and auditable. Ken Lee highlights that each correction made by an analyst teaches the system how the firm operates, and that prolonged use brings it closer to the team's thinking.