Where every number comes from

Diaspore reads public records only: papers, grants, SEC filings, trial registries and FDA decisions. Every number on these pages is filled in from a file or a database view listed in build.json, with its source in facts.json. Hover a number to see where it came from. Each fact is dated by when it became public, and a test reads only facts dated on or before its cutoff.

What is in the database

Counts taken from the database when this site was built. The date range is the earliest and latest date on those records.

What was fetched, and when

Every fetch writes a row to the sources table first. This is the number of fetches recorded by source type, with the date of the latest.

The science corpus

The corpus is each lab's 3 most recent last-author papers at 2014, 2019, 2021 and 2025 year ends. A model read the title and abstract of each paper and filled in a fixed set of fields. It read nothing else: claude-haiku-5-5, prompt extract_v0.3. A paper without an abstract is not read, so it has no extraction: 32% of papers in the panels were left out for that reason.

Total across panels: 297,038 papers, 202,638 with an extraction. Corpus ends 2025-12-31. Nothing after a panel's end date is known to that panel. How accurate the extraction is: what is checked.

Known gaps

• SEC Form D is the only source of funding rounds. Form D does not name the round, so the stage is inferred from the amount raised.
• Acquisitions are not labels. A lab or company that was bought is not counted as a success in any test.
• Founders outside the United States, and founders whose company never filed with the SEC, are undercounted. A company link exists only when the lab head is named as an officer on the filing.
• A lab's institution is the one the paper database gave when the panel was built. It is not dated and can be a parent organisation.
• Search ranks with BM25 word statistics taken from the whole corpus, including papers published after the date you pick. This affects only the order of ties; which labs match is decided by dated papers alone.
• Search needs every word of the query to appear. There are no synonyms, so "tumour" and "tumor" are different words.
• Each lab is represented by its three most recent last-author papers at each panel date. A key paper that is older or middle-authored is missing.
• Nothing here is advice, and nothing here is a prediction until a dated forecast resolves.

The rules everything is held to

1. Public data only. Nothing from an employer, no internal views, no non-public pipeline knowledge. If the origin of an input is unclear, it is not used.
2. No hindsight. Every signal and score carries the date its facts became knowable. A backtest reads only facts dated on or before its cutoff and is scored on what happened after. The split happens in one place in the code, and an undated event is never an input.
3. No invented numbers. A cost, price, market size or effect size appears only with a source. An estimate is a range, labelled as an estimate or an assumption, with its assumptions listed.
4. Nothing publishes unverified. A memo goes out only if a person has verified every claim and every stated fact has a source. The writer and the fact-checker are separate agents.
5. Every test is shown beside baselines: random order, citation rank, a plain AI prompt, and the score without its economics layer. A result without baselines is not a result.
6. Everything is versioned. An extraction records its model and prompt, a score records its rubric, and a rubric is never edited in place.
7. People are named only from public sources, only as heads of labs on a public paper or as authors of a cited source. No page hints at unfiled patents or private plans.

Which of these numbers a person has checked, and which an AI graded: what is checked, and by whom. The long-form documents behind them: documents.