Why can't ChatGPT answer questions about current CMBS deal performance?

ChatGPT does not have access to current CMBS performance data. Trustee reports and servicer filings are locked behind paid platforms like Trepp, Bloomberg, and Intex. Without a direct connection to SEC EDGAR filings and a system to extract and organize that data, general-purpose AI tools cannot provide current deal metrics — they can only recite stale origination numbers from their training data.

What is the ambiguity problem with CMBS metrics like DSCR and current balance?

When someone asks for 'the DSCR' of a CMBS deal, there are multiple valid answers: NCF-based vs NOI-based, weighted average by balance vs by count. Similarly, 'current balance' can mean scheduled or actual — the SEC defines both as distinct XML fields with different meanings. Our default balance uses the scheduled field (reportPeriodEndScheduledLoanBalanceAmount), not the actual field. LLMs that don't understand these distinctions will give you a single number without specifying which variant — or worse, mix them up. CMD+RVL labels every metric precisely and makes all variants available.

How does CMD+RVL compare to ChatGPT, Claude, Gemini, and other LLMs for CMBS data?

CMD+RVL tested five major LLMs (ChatGPT, Claude, Gemini, Grok, Perplexity) against five basic questions about BMARK 2019-B14. ChatGPT returned stale origination numbers from 2019 and failed to find current metrics. CMD+RVL answered all five questions with current data sourced directly from SEC filings — with a complete provenance trail for each figure.

What is AI provenance and why does it matter for financial data?

AI provenance is the ability to trace every AI-generated answer back to its original source data. For financial data, this means showing exactly which SEC filing, extraction method, and timestamp produced each number — creating an audit trail that regulators and compliance teams can verify independently.

How does CMD+RVL provide current CMBS deal data with provenance?

CMD+RVL ingests SEC EDGAR filings daily, extracts structured data using deterministic tools (no AI guessing), catalogs every data point with its source filing and timestamp, and publishes live charts on DealCharts.org — each with a complete lineage trail from filing to output.

Proof Case

BMARK 2019-B14: Five Models, One Live CMBS Question

We asked five leading AI models for the current balance and DSCR of a real CMBS deal. None gave a current, fully grounded answer. This page shows what they returned, what was wrong, and how the numbers are recovered from SEC filings.

The Test Question

“What is the current balance and DSCR for BMARK 2019-B14?”

Model Responses

Five models saw the same question. None produced a grounded answer.

All responses captured February 20, 2026. LLM answers change over time.

ChatGPT response to BMARK 2019-B14 question — part 1

ChatGPT response to BMARK 2019-B14 question — part 2

ChatGPT response to BMARK 2019-B14 question — part 3

"I cannot find a current (2025–2026) pool-level unpaid principal balance"Admits defeat in the first sentence. This number is in the latest ABS-EE filing.

"~$807 M ... ~2.18x (UW NCF DSCR)"Real numbers — from 2019 origination. Six years stale.

"Not publicly available in free sources"Both figures are in the ABS-EE filing. Public data, published monthly.

"access a paid CMBS data provider (e.g., Trepp)"ChatGPT's final answer: pay someone else.

STALE

ChatGPT 5.2

ChatGPT returns real numbers — but they're from 2019 origination documents. The $807M balance and 2.18x DSCR are six years out of date. It then admits it can't find current performance data and tells you to pay Trepp.

Why They Failed

The missing piece was a verified source trail.

The numbers were already public in monthly SEC filings.What the models lacked was a verified pipeline that could ingest the filing, extract the right fields, and preserve a source trail for every answer. Without that, they guessed, hallucinated, or refused.CMD+RVL publishes these figures on DealCharts.org with the source filing and update date attached.

BMARK 2019-B14 deal metrics on DealCharts.org — Current Trust Amount, DSCR, loans, and properties with full provenance

BMARK 2019-B14 on DealCharts.org — every metric traced to its source filing.View live page

Current Trust Amount$1,157,461,196

ABS-EE filing, Jan 2026

WA DSCR (NCF)1.88x

ABS-EE filing, Jan 2026

Number of Loans48

ABS-EE filing, Jan 2026

Number of Properties171

ABS-EE filing, Jan 2026

Could you improve this by feeding the filing directly into an LLM with RAG? Probably. But you would still need to know which filing was used, which field was extracted, and whether the same question tomorrow would produce the same answer.The real requirement is not just getting a plausible number. It is getting a number you can trace, reproduce, and defend. We use AI throughout the pipeline, but the cataloged metrics themselves are produced by deterministic extraction and tagged with source, method, and time.

Why Field-Level Precision Matters

There is no single "the balance" or "the DSCR."

When someone asks for "the current balance" or "the DSCR," there are multiple valid answers. The SEC defines distinct balance fields, and DSCR can be calculated in more than one valid way. A model that does not know the difference can give you a number without telling you which field or variant it picked.

Current Balance

Scheduled

Default

$1,157,461,196reportPeriodEndScheduledLoanBalanceAmountThe scheduled or stated principal balance, determined by considering scheduled and unscheduled principal payments received during the collection period.

Actual

$1,159,223,847reportPeriodEndActualBalanceAmountThe outstanding actual balance representing the legal remaining outstanding principal balance related to the borrower's mortgage note.

WA DSCR

NCF-based

WA by Balance1.88x

Default

WA by Count2.01x

NOI-based

WA by Balance2.06x

WA by Count2.19x

CMD+RVL labels every metric precisely and traces it to the specific SEC field it was derived from. The highlighted path is our default — but all variants are available.

Result Summary

Five models. Five questions. One grounded source.

We asked the same five questions about BMARK 2019-B14. CMD+RVL's catalog answered all five from current filings, with a source trail for each field. The models will improve over time, but a number you cannot trace is still a risk.Tested February 20, 2026

Question	ChatGPT	Claude	Gemini	Grok	Perplexity	CMD+RVL
Current trust balance?	STALE	FAIL	HALLUCINATED	STALE	FAIL	PASS
WA DSCR (NCF)?	STALE	FAIL	HALLUCINATED	STALE	FAIL	PASS
Number of current loans?	STALE	FAIL	HALLUCINATED	FAIL	FAIL	PASS
Largest loan by balance?	FAIL	FAIL	HALLUCINATED	FAIL	FAIL	PASS
Most recent filing date?	FAIL	FAIL	HALLUCINATED	STALE	FAIL	PASS

Pass

Stale

Fail

Hallucinated

Now Multiply

1,185

Active Deals

CMBS and Auto ABS deals monitored with full provenance — every number traced to its SEC filing.

7,385

Charts Published

Live charts across DealCharts.org, each with a complete data lineage trail.

Daily

Update Frequency

New filings ingested, extracted, cataloged, and published — every single day.

These models will get stronger.

The stale answers will become current. The hallucinations will get rarer. But even when the answers are perfect, you'll still need to verify them — and so will your regulators.The scariest failure mode isn't a wrong answer — it's a correct answer with no lineage. That's the one your compliance team can't distinguish from a hallucination.The practice of capturing lineage doesn't become less important as AI improves. It becomes the thing that makes AI auditable.The only question is whether you can build it without too much work, and carry it forward as the technology changes.

Now imagine this on your data.

Your firm is using AI on your data right now. When a regulator, auditor, or board member asks where the answer came from — what do you show them?SCHEDULE A FREE ASSESSMENTFree. No commitment. We connect to your systems and show you what's there.