How deep research is verified

How deep research is verified: its sources, how each quote is matched to its document, how “not in” claims are searched, and when two questions connect.

Deep research makes one promise: everything in quotation marks is in the source, word for word, and everything we say is absent was searched for in the whole document. This page explains how that promise is kept, and the reasoning behind each rule.

The sources#

FDA Complete Response Letters, as the FDA publishes them on download.open.fda.gov.
The company's SEC filings on sec.gov: 8-Ks and the press releases filed with them, and 10-Ks.

EphemerisX does not host copies. Every citation links to the agency's own copy, so the document you open is the one the answer was checked against.

Some FDA documents are not public, such as letters from before the FDA began publishing them, or meeting minutes. When a company quotes one in a filing, the answer quotes the filing and says so: “the 2023 letter, as Aldeyra quotes it”. It never presents a company's account of a document as the document.

The question#

Research starts with questions, not documents. For each company we write questions that the record can answer, and that a careful reader would ask:

What did each letter object to? In the FDA's words.
How did the company describe it? In the company's words, filed days later.
What changed between letters? Which concerns carried over, which are new, which disappeared.
What happened next? Meetings, resubmissions, committees, approvals, disputes, lawsuits.
What remains unresolved? What the record does not yet answer.

Then the questions particular to the case: a manufacturer's warning letter, a name the FDA rejected, a response rate reported three ways. Each question is placed in a period (when it happened) and a theme (the FDA's position, the company's account, the evidence), which become the rows and columns of the timeline.

Quotes are matched, not retyped#

A quote is never typed out. It is recorded as a start phrase and an end phrase in a named source. When the research is built:

The source is fetched from the agency and its text read: the PDF's text layer for an FDA letter, the filing's text on sec.gov.
The passage is located from the start phrase to the end phrase, and the source's own text between them is taken, with its page number.
Every quoted span is checked to sit inside its passage, and every highlighted span too.
The link is made to the PDF page, or, for a filing, to the passage itself where the browser supports text links.

If any passage cannot be found, or any quote is not inside its passage, the build stops and reports every failure. Nothing is published with a failed check. Because the quoted text is taken from the source rather than from a draft, a quote cannot drift from the document it cites.

“Not in” statements#

Saying what a filing does not mention is a strong claim, so it is tested, not asserted. Each “not in” statement records the terms that would show the subject was mentioned. At build time the whole document (every page, not an excerpt) is searched for each term, ignoring case. If any term is found, the build stops.

The terms are shown to the reader. On a question page the statement names the search; in the concept preview the dotted node lists every term (see Searched, not found). You can repeat any search yourself in the source document.

Connections#

Two questions are connected only when both of their answers cite the same passage, or the same search. Each connection names that passage, and the build checks that both answers really cite it. A connection drawn between questions that share no evidence fails the build.

This rule keeps the web honest. It would be easy to join questions because they feel related; joining them only through shared evidence means every line on the web can be followed to a passage, and the shape of the web reflects the shape of the record.

Our words, and theirs#

Ours (editorial)	Theirs (quoted)
Questions, summaries, short answers, connecting sentences, node labels and connection labels	Everything in quotation marks, and every passage in The passages quoted or the evidence panel

Our wording stays close to the documents', and is marked where it could be mistaken for theirs (Node label: our summary). Every page's footer says the same thing.

What deep research does not do#

It does not predict. No answer says what the FDA will do next or whether a drug works.
It does not resolve disagreements. When the FDA and the company describe the same meeting differently, both accounts are quoted and the difference is the finding.
It does not fill gaps. When the record is silent (the FDA's approval documents are not yet among a company's sources, for example) the answer says “not in this record”.

Why this method#

Regulatory and financial decisions are made on records like these, and the cost of a misquoted letter is real. A method that makes errors impossible to publish is worth more than one that makes them unlikely: so the checks run on every build, on every quote, and they stop the build rather than warn.

The method also makes the research cumulative. Each checked passage is a fixed point: once it has been located and verified, any number of questions can rest on it, and any new question that cites it is automatically connected to the old ones. That is how a set of answers becomes a map. See An ephemeris of reasoning.