Only one format can prove an AI's answer

“I just need to know if the answers are correct.”

Every executive running AI asks some version of this. With today’s formats, nobody can give them a straight answer.

AI models state every figure with the same confidence, right or wrong. A citation could settle the question, but only if someone can check it against the exact place it came from. Raw web pages and RAG (the retrieval set-up behind most document-chat tools) can’t do that. They have no address for a line or a table cell, nothing checks a citation before it’s used, and the AI has no way of knowing whether what it cites is exact. So the user takes the AI’s word for it, or does the research over again.

ADAP (Agentic Data Address Protocol) is the only format that does all of it:

• it gives every line and table cell its own address, indexed in the file;
• the AI cites those addresses and knows each citation will be checked;
• the file verifies every citation before the answer is shown, and logs the result;
• one click opens the exact line or cell, highlighted.

Advanced GP doesn’t build AI. We built the format that makes AI answers provable.

The test

One AI model (Gemini 3.1 Pro) answered the same 100 questions on 30 SEC 10-Q filings from 15 companies, plus 20 questions that combine two filings. It got the filings three ways:

• ADAP: each filing converted into an agentic file with every line and cell addressed.
• Raw HTML: the AI is given the 30 sec.gov links, which is how people use web pages with AI.
• RAG: each filing is cut into chunks, and the 20 chunks closest to the question go to the model.

Every set-up was told to cite every claim with the document and the exact quote, which is the most help raw HTML or RAG can give.

Finding 1: only ADAP can verify and trace a citation

With ADAP, every answer ended in a known state: 99 verified through their own citations, and 1 shown false by its citation. With raw HTML the user had to take the AI’s word on 55 answers. With RAG, it was 79.

How the AI knows. The Protocol inside every ADAP file tells the AI to cite only addresses it has actually read. Before an answer is shown, the file checks every cited address. It confirms that the address exists exactly once and that the AI read it, in about 1 microsecond per answer, and writes the result to the file’s Verification Log. Raw HTML and RAG give the AI nothing to cite but a page link, a file name or a quote, and nothing checks it.

Finding 2: to compare at all, we had to do the tracing for the other formats

Raw HTML and RAG can’t verify their own citations, so our audit program did that work for them. It searched all 30 filings for every quote their answers gave:

Even a quote found exactly once doesn’t tell you which cell of which table holds the figure. Every link the raw-HTML answers wrote, all 113 of them, went to the top of a 60-100-page filing. The RAG answers gave file names only. A user of these formats starts every check with a search.

ADAP needed none of this. Its citations are addresses, and the file had already checked every one.

Finding 3: accuracy isn’t the issue. Provability is.

The AI was right about as often in every format, and just as confident when it was wrong. The difference is whether anyone can tell which answers are right.

Finding 4: errors compound across documents

Agentic AI doesn’t answer one question from one file. It chains steps across filings, models, intranet pages and databases. An imprecise figure early in the chain becomes a wrong decision later, and the error compounds at every step.

Finding 5: one click, or a search through the filing

Every link the raw-HTML answers wrote, all 113 of them, went to the top of the filing, none to a place inside it. With RAG there was no link at all.

What it costs to run

ADAP spends about a cent more per answer, because the AI reads the sections it needs rather than one page or 20 chunks. That’s still a fraction of giving the AI the filings whole, which cost $0.458 per answer. The ADAP figure is a range because the price tier of some calls wasn’t recorded.

How ADAP controls risk: two layers

Layer 1: software checks every citation. The file’s own deterministic code, with no AI involved, checks every citation and logs it. The same citation gives the same result every time.

Layer 2: a person confirms in one click. One click opens the cited line or cell, highlighted in its table, so the reviewer sees exactly what the AI reasoned from. A real cell from the wrong table passes Layer 1, and this is the step that catches it. The test’s one wrong ADAP answer cited a cell from the property-and-equipment table when the question asked for total assets. The reviewer could see the mismatch at the highlighted cell.

With raw HTML or RAG, neither layer exists. Nothing is checked automatically, nothing is logged, and the human pass starts with searching a 60-100-page filing.

A second AI is no substitute. A new AI session re-checked every answer. With ADAP addresses it was right every time. Without them it made mistakes both ways: it confirmed a wrong answer and rejected correct ones.

Why it matters in regulated work

The firm stays responsible for every AI output, and unchecked output is already costing firms money. A US court sanctioned lawyers for filing “fake quotes and citations created by the artificial intelligence tool ChatGPT”¹. Deloitte refunded over $60,000 to the Australian government over a report with AI errors².

Regulators are asking for exactly these controls:

• The Financial Stability Board: “Hallucinations can be difficult to detect and evaluate”³.
• NIST: “Review and verify sources and citations in GAI system outputs during predeployment risk measurement and ongoing monitoring activities”⁴.
• EU AI Act, Article 12: “High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system”⁵.
• Bank of England and FCA: “Only 2% of use cases have fully autonomous decision-making”⁶.

Auditing standards already favor automated controls. ISA 315: “Automated controls may be more reliable than manual controls because they cannot be as easily bypassed, ignored, or overridden, and they are also less prone to simple errors and mistakes”⁷.

ADAP puts the program first, checking every citation every time and logging it. People keep the part only people can do: judging the highlighted source in context.

This is a test result and a description of alignment, not legal advice or a compliance opinion.

Your organization’s data

The test used SEC filings. Advanced GP brings the same control to your own data: documents today, with intranet and databases next. We convert legacy formats to ADAP, connect them to the systems your teams already use, and run a pilot on your own material.

Scope and method

• One AI model, one document type (SEC 10-Q), one run per set-up, 100 + 20 questions. The cross-filing pairs are synthetic and test retrieval across files.
• Correct values were computed by program from the filings’ cells and confirmed by a second model.
• All grading was done by programs from the recorded test data, with no person or AI judging results. Every figure was computed twice, in Python and in SQL, and both had to agree.
• Raw HTML ran through the Gemini API’s URL tool, and many sec.gov fetches failed in this run. A consumer chat app may read pages differently.
• The agentic files were made with a prototype converter. Advanced GP built the ADAP format. The audit package is available on request: info@advancedgp.ai.

Sources

1. 1U.S. District Court, S.D.N.Y., Mata v. Avianca, Inc., No. 22-cv-1461, Opinion and Order on Sanctions, 22 June 2023.
2. 2CFO Dive, “Deloitte refunds over $60K for report with AI errors, Australian government says”, 21 October 2025.
3. 3Financial Stability Board, “The Financial Stability Implications of Artificial Intelligence”, 14 November 2024.
4. 4NIST, AI Risk Management Framework: Generative AI Profile (NIST AI 600-1), July 2024.
5. 5Regulation (EU) 2024/1689 (EU AI Act), Article 12.
6. 6Bank of England and FCA, “Artificial intelligence in UK financial services – 2024”, 21 November 2024.
7. 7IAASB, ISA 315 (Revised 2019), Appendix 5.

AI prices: Google Gemini API pricing, checked 2026-10-20.