EphemerisX Research
How EphemerisX Deep Research Works
Following questions through the institutional record.
Institutional decisions leave a record: an FDA letter, the company's 8-K three days later, a 10-K six months on, an advisory committee vote, an approval. Each document is usually read on its own. What none of them shows is how they connect: what changed between the first letter and the second, which concern survived a resubmission, how a company's account set beside the agency's reveals what each side emphasized.
EphemerisX Deep Research reads the whole record and follows it. It does not summarize documents. It asks questions that can only be answered by reading several documents together, across sources and across time, and it answers them in the record's own words.
The research engine
A finding raises a question → the question is followed through the record → the record shows a change, a connection or a discrepancy → that raises the next question.
For example: an FDA letter says the efficacy could not be interpreted. Why? The patient population was heterogeneous. Which populations? The company's filings report three response rates. Why three? They describe three different populations. Which one did the approval use? The one the FDA's concern pointed to.
That chain is the product. A collection of disconnected facts is not.
It is what separates an investigation from a search result, a summary, a database entry or a company profile: each of those treats documents one at a time.
Four layers
- Follow a record. One asset, from FDA letter to company filing to the next filing to the decision, in date order.
- Follow a question. A finding raises a question; the record is searched for the evidence that answers it; the answer raises the next question.
- Follow a connection. Questions that cross records: one company and another, one drug and another, one FDA decision and another.
- Discover patterns. Every investigation adds another route through the record. Over time, those routes reveal connections between cases: what tends to happen after a Complete Response Letter, which concerns survive a resubmission, when an advisory committee changes the course.
One web, several entrances
Records → questions → investigations → connections → new questions.
Deep Research is not a report about a company. It is a path through the record. Sometimes the path stays within one company's record; sometimes it crosses into another's; sometimes that crossing becomes a new investigation.
Company pages, the list of investigations and the cross-investigations are different views of the same web. Every company page shows the cross-investigations that involve it; every investigation shows the cross-investigations that quote its evidence; every cross-investigation maps the company questions it connects. A reader who arrives at one question can follow the record somewhere they did not know to look.
The next question is allowed to cross the boundary. When comparing records raises a question none of them answers alone, it is recorded on the cross-investigation as a question it raises, and becomes a candidate for the next investigation. The more investigations there are, the more connections become possible.
How an investigation unfolds
An investigation starts from an anchor event, usually an FDA decision. From there, these are the moves available to it. They are not steps in a fixed sequence. The record determines which question comes next, and only the questions the record can answer are written.
| Move | The question it asks | Example |
|---|---|---|
| Anchor | What did the decision say, in its own words? | What did the FDA's first letter object to? |
| Account | How did each party describe it? | How did the company's 8-K describe the letter? |
| Before | What earlier advice, meeting or agreement does it refer to? | What did the FDA tell the company before the application was filed? |
| Change | What changed between two documents or decisions: wording, numbers, populations, requirements? | What changed between the first and second letters? |
| Discrepancy | Why do two documents report different things, and do they describe the same thing? | Why do the reported response rates range from 24.2% to 34%? |
| Dependency | What does a conclusion rest on: a trial design, a population, a manufacturer, a definition? | When did injected lesions become an FDA concern? |
| What next | What followed: a meeting, a resubmission, an appeal, an advisory committee, an approval, a withdrawal? | How did the drug go from a second letter to approval in four months? |
| Consequence | What did the company disclose changed as a result: trials, manufacturing, staff, litigation? | What happened inside the company after the second letter? |
| Open | What remains unresolved, and what is not in the record? | What remains unresolved after the approval? |
Cross-investigations
Some of the most important questions sit between records: between two drugs, two companies, two regulatory decisions. A cross-investigation answers a question that no single company's record can, through a specific connection:
| Connection | Why the records belong together |
|---|---|
| Shared concern | The FDA raised the same kind of concern about different programs. How did each record respond? |
| Divergent outcome | Similar starting points, different endings. What differed in the record? |
| Comparable evidence | The same kind of evidence, weighed in different programs: a single-arm trial, a second trial, a confirmatory study. |
| Regulatory pathway | The same regulatory step played out in different records: a resubmission class, an appeal, an advisory committee. |
| Precedent | One program's regulatory history came before, and resembles, another's. |
| Different response | Companies responded differently to the same kind of FDA finding. |
| Common dependency | Different programs rest on the same supplier, trial design or assumption. |
| Sequential event | One record's later event changes how another's earlier one reads. |
A cross-investigation is never a generic comparison of two companies. Each one rests on a connection that can be traced to the primary record in every case it draws on, and it quotes each of those records.
How questions are chosen
Candidate questions are ranked, and only the strongest are written:
- Evidence density: several primary documents bear on it.
- Time: answering it means following the record over time.
- Cross-source: it needs the FDA and SEC records read together.
- Surprise: the answer is not visible in any single document.
- Specificity: it concerns a real asset and event, not the company in general.
- Searchability: a real person might ask it.
- Follow-up: the answer leads to another question.
- Source quality: the primary documents are available.
- No speculation needed: it can be answered without guessing at motives or outcomes.
An obscure question with a real connection in the record is worth more than a popular one the record cannot answer.
The evidence rules
- Every quotation is matched word for word to the document it cites before publication, by software that stops the build if it is not.
- "Not in the document" is a search, not a judgment. It is stated only after the whole document has been searched for the terms.
- Connections are evidence. Two questions are linked only where both answers quote the same passage.
- What the record lacks is said plainly. Where a page is a scanned image with no searchable text, or a document such as meeting minutes is not public, the article says so rather than filling the gap.
- Neutral language. Each document's words are set beside the others'. Intent, honesty and adequacy are not characterized, outcomes are not predicted, and nothing is investment advice.
How an article is built
- The question, as the title.
- Short answer: the answer, first.
- What the evidence shows: the quotations, and the connection between them.
- What changed, when the question is about change.
- Follow the thread: the next questions, linked.
- The documents and passages: every source by date, the quoted words highlighted, a link to the original.
The map and the paths
- The Ephemeris
- The map of the record: every FDA and SEC document read, broken into its sentences, drugs, companies, decisions and the links between them, as one knowledge graph.
- Deep Research
- A path through the map. Each question is a route from one document to the next, and each answer is the passages along it.
- Cross-investigations
- Paths that cross between records, where one company's or one drug's record starts to explain another's.
- EphemerisX Pro
- Being built to ask these questions across the whole map at once: of any company, any drug, any decision, answered with the same evidence.
EphemerisX exists to reconstruct how consequential decisions were actually made, from the documents they produced, so that anyone can follow the reasoning and check every step. Deep Research is that method in public: free to read, and verifiable passage by passage.