Nexum Deal Debt Scenario Update

Overview

This environment is a commercial real estate diligence benchmark built from an anonymized property-level deal workflow. The main task asks an agent to review a native Excel underwriting model and a PDF debt quote summary, update the model for three bridge-financing scenarios, and decide whether the deal still clears a 20% levered IRR hurdle at a revised $132 million purchase price.

The environment is designed to test realistic financial-document reasoning rather than simple text extraction. The difficult part is selecting the correct floating-rate bridge-loan assumptions from a messy PDF extraction and applying them to the model.

What Is Exposed

Packaged source files:

• data/native/Amazon Delaware Debt Guidance 3-21-22.pdf
• data/native/Amazon Middletown - AREP Run -3.15.2022 (Base Case).xlsx
• Selected anonymized markdown deal documents under data/markdown/

The environment does not package prior eval outputs, scratch workbooks, virtual environments, or gold-source iteration workbooks.

Main Task

The principal eval row asks the model to:

• Update purchase price to $132,000,000.
• Use Section I, Initial Floating Rate Bridge Loan, from the debt guidance PDF.
• Ignore Section II, Refinance.
• Use the high end of each quoted range for Life Co Financing, Bank Financing, and Higher LTV Financing.
• Run the scenario workbook tool.
• Return levered IRR for each quote, rounded to the nearest hundredth of a percent.
• State which quote, if any, clears 20% levered IRR.

Correct scenario outputs:

Why This Is Hard

The PDF extraction is intentionally column-oriented and messy. Strong models usually need to distinguish:

• Floating-rate bridge financing from fixed-rate guidance.
• Section I bridge loan assumptions from Section II refinance assumptions.
• SOFR today from all-in rate today.
• Lender spread from lender/origination fee.
• Basis points as decimal model inputs.

Common failure modes observed in local benchmarking:

• Using fixed-rate lender-fee rows for Life Co or Bank.
• Treating 0.30% SOFR as the lender fee.
• Treating 2.80% all-in/rate-today text as SOFR.
• Averaging ranges instead of using the high end.
• Producing a directionally plausible final answer after running the wrong model assumptions.

Tools

The environment exposes document tools and a scenario workbook tool:

• read_debt_guidance_pdf
• inspect_original_debt_model
• run_debt_quote_scenarios
• search_document
• read_document

When Windows Excel automation is available, the scenario tool recalculates copied workbooks directly. On hosted/Linux infrastructure where Excel COM is unavailable, it uses a deterministic fallback calibrated to the validated workbook outputs so the benchmark remains runnable.

Rubric

The native debt scenario task is scored as:

Local Run

prime --plain eval run nexum_deal_folder_test --env-dir-path . --skip-upload -n 1 -r 5 -c 1 -m google/gemini-2.5-pro -t 5000 -T 0 --save-results --debug -A

For public PI runs, omit --skip-upload as appropriate.