What deep research means

Why the question, not the document, is our unit of research, why every answer must be traceable, and how answers add up to an ephemeris of reasoning.

Most tools for working with documents do one of two things. Search finds the documents that contain your words. Summarisation compresses one document into fewer words. Both assume the answer is in a document, waiting to be found or shortened.

The questions that matter most after a regulatory decision are rarely like that. Did the FDA ask for the same study twice? Did the company tell investors what the letter said? Which concern survived the resubmission? No single document answers these. The answer exists only between documents: in a comparison, a sequence, or an absence. That is what we mean by deep research.

The question is the unit#

A deep research answer starts from a question and goes looking for the evidence, across as many documents as it takes. A question is a better unit than a document or a topic for three reasons:

It can be answered: “Replimune” is a topic. “What changed between the FDA's first and second letters?” has an answer: one deficiency new, two carried over, and the Phase 3 trial moved from design to early data. An answer can be right or wrong, and so it can be checked.
It cuts across documents: That answer needs both letters, and the company's account of each. A question naturally gathers the documents it needs; a document does not gather the questions it bears on.
It can be connected: Two questions that rest on the same passage are related in a precise way. A web of questions shows how a record hangs together, which a pile of documents cannot.

Every answer has an address#

Asking harder questions raises the stakes of answering them. An answer drawn from twenty documents is only useful if a reader can follow it back into each one. So every deep research answer is built from findings: a passage in a named document, at a page, found word for word; or a search of a whole document that found nothing. Our sentences connect the findings; they never stand in for them. See How deep research is verified.

This is what separates deep research from a fluent summary. A summary asks for trust. A deep research answer offers a path: from the short answer, to the quoted evidence, to the passage, to the agency's own copy of the document.

Positions and motion#

In astronomy, an ephemeris is a table of where bodies are and how they move through time. The Ephemeris does the same for institutional records. Its subject is positions: what the FDA held, what the company said, what a committee voted. And its subject is motion: how those positions changed from one letter, meeting or filing to the next.

Deep research is how positions and motion become visible. One question fixes a position (What did the FDA's first letter object to?). Another measures motion (What changed between the first and second letters?). A third tests whether two accounts of the same moment agree (Was the FDA's advice to Replimune consistent? The FDA's letter says it was; Replimune's filings say it was not).

An ephemeris of reasoning#

EphemerisX is building toward one map of this reasoning across institutional records. The pieces already exist:

The record. The Ephemeris charts every document read, as entities and relations, on the FDA and SEC maps.
The questions. Question webs ask what that record means for one company, and connect the answers through shared evidence.
The answers on demand. Ask the Ephemeris previews asking a new question, and getting an answer with its reasoning graph attached.

The method is what makes the pieces add up. Every verified passage is a fixed point, and any question that cites it is joined to every other question that does. As more questions are answered, about more companies, the fixed points are shared more widely and the questions begin to connect across cases: the same FDA concern in two programmes, the same procedural path taken by five companies. That accumulated structure of positions, motion and connection is the ephemeris of reasoning. See Connections across cases.

Why this matters#

The value is in the connections. Anyone can read one letter. The insight that a concern from a 2021 meeting opened the FDA's second letter five years later needs the whole record, read together.
The cost of being wrong is high. Investment, regulatory and editorial decisions rest on what a letter said. Answers built from verified passages can be checked in minutes, not re-researched from scratch.
The work compounds. A verified passage never has to be found again. Each new question is cheaper to answer and more connected than the last.

Where we are starting#

We are starting where the gap between two institutional records is sharpest: the FDA's Complete Response Letters and what each company told the SEC about them. Five companies have full question webs, and ten have concept-preview answers on /ask. Why this starting point: Questions between the records.