# How Terminal X Uses GPT-6.1 Sol and a 54-Criterion Eval Agent to Match Senior Investors 88.6% of the Time

When a senior partner or credit analyst leaves an investment firm, the company keeps their old deal memos in a shared folder, but it loses the unwritten judgment behind why those deals were approved or killed—a costly institutional blind spot that financial AI platform **Terminal X** calls "memory debt." In a newly published production case study with OpenAI, Terminal X revealed how it encodes that institutional judgment into a two-tier architecture powered by **GPT-6.1 Sol** and **GPT-6 Luna**, achieving **88.6% agreement with senior human investors** and saving a single investment firm **$2.4 million annually** on due diligence workflows.

For any business leader or engineering team trying to get reliable, auditable work out of LLMs in high-stakes industries like finance, legal, healthcare, or enterprise procurement, Terminal X's architecture provides a step-by-step blueprint you can copy today. Instead of asking a single chatbot prompt to "write an investment memo," Terminal X splits the workflow across specialized generation agents and an independent **Eval Agent** that grades every draft against a strict **54-criterion rubric** before a human ever sees it.

Upgrading that Eval Agent and core reasoning stack to OpenAI's **GPT-6.1 Sol** delivered three immediate operational improvements over the next-best frontier model: it caught **25% more recommendations that strayed outside a fund's investment mandate**, increased human-reviewer agreement by **+220 basis points (to 88.6%)**, and ran those deep evaluation workflows at **30% lower cost per query**.

Meanwhile, for high-volume document screening across English, Korean, Japanese, and Chinese data rooms where speed and token cost matter more than deep thesis synthesis, Terminal X routes tasks to **GPT-6 Luna**—proving how practical Sol + Luna model routing cuts cloud bills without sacrificing accuracy.

## Fast Facts: Terminal X’s GPT-6.1 Sol & GPT-6 Luna Production Architecture

- **Core Business Problem ("Memory Debt")**: Built by former JPMorgan and Blackstone professionals (led by Co-founder and CEO Hyun Hong) to capture unwritten deal-screening judgment inside private, on-premise or VPC enterprise data rooms, as detailed in [OpenAI's Terminal X case study](https://openai.com/index/terminal-x).
- **88.6% Senior Investor Match**: Terminal X's agentic pipeline matches senior human investor judgment **88.6%** of the time (**+220 bps** higher agreement on GPT-6.1 Sol vs. the next-best model).
- **The 54-Criterion Eval Agent**: Every generated investment memo and thesis is graded against a **54-criterion rubric**, catching **25% more mandate deviations** with GPT-6.1 Sol.
- **Hard ROI & Unit Economics**: Delivered **$2.4 million in annualized savings** on due diligence for a single customer while reducing cost per query by **30%** compared to competing frontier models.
- **Multilingual Private Data Rooms**: Processes deal documents natively across **English, Korean, Japanese, and Chinese** with a full audit trail of retrieved sources, model versions, and step-by-step reasoning.

## Architecture Blueprint: How Terminal X Splits Work Between GPT-6.1 Sol, GPT-6 Luna, and the Eval Agent

Rather than running every prompt through the most expensive reasoning setting, Terminal X maps each stage of an investment firm's workflow to the model tier that fits its cognitive and cost profile.

| Workflow Stage | Assigned Model / Component | Specific Operational Role | Measurable Business Outcome |
|---|---|---|---|
| **1. High-Volume Data Room Screening** | **OpenAI GPT-6 Luna** | Rapidly ingests and extracts metrics from multilingual filings (EN, KO, JA, ZH) | High token efficiency and low latency on thousands of raw pages |
| **2. Institutional "Capital Memory" Store** | **Versioned Desk Memory Ledger** | Encodes how a specific credit or PE desk screens borrowers, past committee votes, and house views | Eliminates "memory debt" when senior analysts or partners depart |
| **3. Deep Thesis & Mandate Synthesis** | **OpenAI GPT-6.1 Sol** | Synthesizes multi-document deal memos, risk factors, and fit against fund mandates | **30% lower cost per query** than the next-best long-context model |
| **4. 54-Criterion Rubric Audit** | **Terminal X Eval Agent (GPT-6.1 Sol)** | Grades draft outputs against 54 desk-specific rules before surfacing to analysts | **+25% higher detection** of off-mandate errors; **88.6%** human agreement |

## Why a "54-Criterion Eval Agent" Beats Prompt Engineering Every Time

Many enterprise AI pilots fail because teams stuff 20 pages of rules into a single generation prompt and hope the model remembers all of them while writing a report. Terminal X's forward-deployed engineers take the opposite approach:

### 1. Reverse-Engineering Unwritten Rules into "Capital Memory"

Before writing a single prompt, engineers sit onsite with a fund's credit or co-investment desk to reverse-engineer their best historical memos. They document the tacit rules that are never written in an employee handbook—for example, which off-balance-sheet liabilities automatically disqualify a borrower in a specific sector, or which third-party market reports the investment committee actually trusts. Those rules are saved as a versioned **Capital Memory** record owned entirely by the firm inside its private cloud.

### 2. Separating Generation From the 54-Point Rubric Grader

Once the generation sub-agents (using **GPT-6 Luna** for extraction and **GPT-6.1 Sol** for synthesis) produce a draft investment memo, the **Eval Agent** acts as a ruthless senior partner. Powered by GPT-6.1 Sol's long-context reasoning, the Eval Agent checks the draft against **54 distinct criteria**—verifying source citations, math consistency, downside scenario modeling, and strict alignment with the fund's charter. If an answer drifts outside the fund's mandate, the Eval Agent flags and rejects it before it ever reaches a human portfolio manager.

## How to Apply the "Capital Memory + Eval Rubric" Pattern in Your Business

You do not need to run a Wall Street hedge fund to use Terminal X's pattern. Any team that produces complex, repeatable client deliverables—such as legal contract reviews, B2B technical proposals, medical underwriting, or SEO/CRO site audits—can implement this three-step workflow:

1. **Build a 25-to-50 Point Binary Grading Rubric From Your Best Past Work**: Take five of your team's best historical reports and five flawed drafts. Extract a concrete checklist of 25 to 50 yes/no criteria (e.g., *"Does every financial claim cite a primary table row?"*, *"Does the recommendation violate any client budget constraints?"*).
2. **Route Extraction to a Fast Model (Luna) and Synthesis/Grading to a Deep Model (Sol)**: Stop using one model for every step. Use a high-efficiency model like `GPT-6 Luna` (or `MAI Code 1.1 Flash` / `Gemini Flash`) to parse raw PDFs and spreadsheets, and reserve `GPT-6.1 Sol` or `Claude Opus 5.5` for final synthesis and rubric grading.
3. **Never Show Unaudited Drafts to Human Reviewers**: Run your independent **Eval Agent** automatically at the end of every generation pipeline; if the draft fails any mandatory rubric item, loop the specific failure feedback back to the writer agent automatically so humans only review rubric-verified work.