Reading the dataset field by field

A ranking is only as credible as the schema behind it. We publish the two source files that build every table on this site, but a JSON key does not explain its own boundaries: totalFundingUsd is not a valuation, amountUsd is not a company total, and an empty field is not a zero. This page is the dictionary we wish every data journalist had when they opened our files.

Data cutoff 2026-11-12Page generated 2026-11-21

Why a field dictionary belongs on a rankings site

Most funding trackers publish a table and a headline number, then leave the reader to guess how the number was produced. That is a problem when the same word, funding, is used to mean a priced venture round, a government letter of intent, an at-the-market equity programme and a reported revenue forecast. All four appear in the news about quantum technology companies, and all four can be mistaken for each other by an automated collector.

We therefore treat the schema as part of the editorial product. Every field described below has a single meaning, enforced by the build script rather than by a style guide. If a value cannot be stated within that meaning, the field is left empty or explicitly flagged; it is never filled with an estimate. The point is not to make the dataset look complete. It is to make every gap visible and auditable.

This page describes the files as they are shipped, not as an idealised future version. When a field changes meaning, the change is recorded in the scan and change log alongside the date it took effect, so a figure quoted six months ago can still be interpreted correctly.

companies.json — one row per company

The registry holds one object per company. id is the stable slug used in URLs, filenames and cross-references; name is the display name; aliases collects the spellings a source might use instead. country, countryCode, region and hq describe geography at different granularity, because a company can be headquartered in one city, incorporated in another, and counted in a regional aggregate that includes both. tech and subTech are browsing labels for the three technology categories the site uses, not a technical certification.

totalFundingUsd is a curated, company-level aggregate expressed in US dollars. totalPrecision records whether that aggregate is exact, approximate, or an explicitly stated lower bound. confidence describes the strength of the sourcing behind the record as a whole. status and, where applicable, ticker distinguish private companies from listed ones. founded takes a year, and is omitted entirely when we have no defensible date; the site does not print an invented founding year to make a card look tidy.

description is a short editorial sentence, and website is a first-party link when we have one. A newer optional object, dossier, holds owner-verified bilingual company records supplied directly to the desk. Those records are kept inside the registry rather than in a parallel spreadsheet, so a company cannot appear twice under two different facts. Monetary values inside a dossier keep the units and wording supplied and are never added into the ranking total a second time.

funding-events.json — one row per report, not per round

This is the raw ledger and it is deliberately noisier than the published tables. Each object describes a single report we encountered: id, companyId and company identify the subject; date and datePrecision say when it happened and whether we know the day or only the month. round is the instrument label as reported, amountUsd is the disclosed cash figure when there is one, and amountOriginal preserves the original wording and currency so a conversion can be checked.

valuationUsd appears only when a source states a valuation; it is kept in its own column and never summed into disclosed capital. investors is a list that may be empty. summary is the headline or short description from the source, and sources plus sourceUrl keep the attribution. Several reports about one financing will all be present here. We do not silently delete duplicates, because the raw ledger is the evidence that the publication layer did its job.

The practical consequence for a reader is simple: counting rows in funding-events.json will not reproduce the number of financings shown on the site. The merge step described in our de-duplication column is what turns reports into events, and the publication gate described below is what turns events into published rows. Both transformations are deterministic and run at build time, so the same input always yields the same output.

The amount fields and what the ~ sign actually means

amountUsd is a cash amount attached to one instrument on one date. It is not a company total, not a valuation, not revenue, and not a commitment that may or may not be drawn. When a report gives a range, a rounded figure such as about half a billion, or a value in a currency we convert, amountPrecision is set to approx and the site renders the number with a ~ prefix. amountOriginal keeps the source wording so a reader can see exactly what was converted.

A missing amountUsd is not a zero. It means the report did not publish a cash figure we were willing to record. Some of these rows are policy notices or public-market instruments with no priced cash amount at all; they remain in the timeline because they are materially relevant to a company's capital history, but they cannot contribute to a disclosed-cash total. Where a valuation is reported instead of a raise, it goes into valuationUsd and is labelled as a valuation, never as funding.

The site also keeps a distinction between the registry aggregate and the sum of event rows for a company. They are not required to agree: the aggregate is curated at company level and can cover rounds outside the window of published events, while the event list is a row-by-row record of what we published. Company profiles show both numbers and the difference, rather than picking whichever looks better.

Provenance: autoCollected, reviewed and the editorial gate

autoCollected marks rows first found by automated news monitoring. reviewed marks rows an editor has examined directly. These flags are provenance, not a quality score: an automatically discovered row is not automatically trusted, and a manually entered row is not automatically correct. The visible badge on the site tells a reader how a row entered the ledger, so that the distinction can be checked rather than inferred.

The stronger guarantee is the publication gate. A financing is published only when an editor has written a short, specific note about that exact round in the notes file. Rows without a note stay in the raw ledger with a visible held-for-review status at build time; they do not enter the ranking, the totals, the timeline, the feed or a company profile. This turns a promise about human review into a mechanical constraint.

A note is not a fact-check certificate. It is a receipt that someone opened the source and formed a judgement about that particular financing, and it is written to be distinguishable from every other note. Where a source is weak, the note says so. Where a headline looked like funding but turned out to be a revenue forecast, the row is held. The confidence column and the scan log explain how to audit these decisions.

Derived files and what they are not

stats.json holds the aggregates that drive the hero numbers, snapshots and charts. rankings.json is the de-duplicated, rendered view used by the tables. company-intel.json is a generated export of the owner-verified dossiers that already live in companies.json, produced at build time so the two can never drift apart. feed.xml carries the same editorial notes as the pages, never raw scraped headlines. None of these files is an independent source of truth.

If you want to challenge a figure, start from the two source files rather than from a derived one. A derived file can legitimately differ from the raw ledger because it reflects the publication layer, and the difference is explained on the methodology and scan log pages. What would not be legitimate is a derived file that quietly invents a value which has no corresponding raw row.

The schema is versioned through the repository's change history. Each file is plain UTF-8 JSON with no required authentication, no rate limit and no licence restriction beyond attribution under CC BY 4.0. That is a deliberate choice: a dataset that is easy to download is also easy to disagree with, which is the only reliable way we know to find our own mistakes.

A worked reading of one record

Suppose a row says a company raised an approximate amount in a round labelled Reported raise, with an empty investors array and a source link. The correct reading is: one report exists, it did not name a specific round or investor, the figure is approximate and the original wording is preserved for checking. The row is a report, not a confirmed financing, until it passes de-duplication and the editorial gate.

Suppose instead a company has three rows in the same ten-day window with the same approximate amount, from three outlets. The correct reading is one financing with three corroborating reports, not three rounds. The published table will show one row with a merged-reports count, and the raw ledger will still show all three. Neither view contradicts the other; they answer different questions.

Files you can download right now

FileWhat it contains
companies.jsonCompany registry, verified dossiers and cumulative disclosed totals
company-intel.jsonEditor-verified China company records, exported from the registry
funding-events.jsonRaw report ledger, including duplicates and candidates awaiting a note
stats.jsonAggregates used by the hero numbers and snapshots
rankings.jsonDe-duplicated, rendered view used by the tables

The practical takeaway

The useful habit is to read a number together with its flags. A disclosed amount without its precision marker, instrument label and provenance is just a number; with them, it becomes a claim a reader can check. If you find an interpretation on this site that the schema does not support, the corrections address is on every page.

More from the research desk

Global funding rankingCompany profilesMethodologyScan log

Compiled by QuantumRoundup Research Desk · Reviewed by the Editorial Desk (every published event has its own hand-written editorial note)

Data cutoff (latest reviewed event): 2026-11-12 · Page generated: 2026-11-21 · Editorial policy · Corrections log · Methodology · build 4034dd8677a5