Step one: get the two source files, not the derived ones · Step two: apply the publication gate · Step three: merge syndicated reports inside the ten-day window · Step four: recompute the totals and compare · What a genuine discrepancy looks like · Why we publish the procedure at all
Step one: get the two source files, not the derived ones
Download data/companies.json and data/funding-events.json from the site root. These are the authoritative inputs. The first is the company registry with curated aggregates; the second is the raw report ledger, including duplicates, approximate amounts and rows still waiting for an editorial note. Both are plain UTF-8 JSON under CC BY 4.0.
Resist the temptation to start with stats.json or rankings.json. Those are outputs of the build: they already reflect de-duplication and the publication gate, so comparing them with a number you derived from raw input will produce a false discrepancy. Use them only as a check on your own arithmetic after you have reproduced the transformation yourself.
Keep a local copy with the download date noted. The ledger is updated by a daily job, so a recomputation performed a week later against a newer file is not a valid test of an older published figure. The scan and change log records when each build ran and what the data cutoff was.
Step two: apply the publication gate
The raw ledger contains reports that the site does not publish. A financing is published only when an editor has written a specific note about that exact round; rows without one are held for review and never enter the ranking, the totals, the timeline, the feed or a company profile. The build prints the number of held rows on every run, and the raw file keeps them so the exclusion is visible.
If you are reproducing our number, exclude held rows. If you are building your own, keep them but label them separately. The distinction matters because a held row is not a retracted fact: it is an unjudged report. Some are revenue headlines that a keyword collector misread as a funding round; others are legitimate financings awaiting an editorial note.
The confidence column explains how to audit that decision, and each published event on the site carries its note plus a source link. You can therefore check the gate qualitatively as well as mechanically: open a note, open its source, and see whether the note describes that financing rather than a generic statement about the sector.
Step three: merge syndicated reports inside the ten-day window
Group the remaining rows by company and compare them inside a ten-day window. Collapse two rows when both amounts are undisclosed, when disclosed amounts agree within twenty percent, or when one names a specific round and the other is a generic reported raise. Keep the highest-scoring row: reviewed over auto-collected, named round over generic, named investors over none, source URL over none, longer description over shorter.
Do not delete the losing rows. Record how many reports were merged into the survivor, exactly as the site does, and keep the raw rows in your own working copy. A merge count of zero is normal for a company with a single reported round; a large merge count is a signal that the announcement was heavily syndicated, not that the company raised more money.
The de-duplication column gives the reasoning behind each condition and the cases where it fails. If your recomputation differs from ours, this step is the most likely cause: a different window, a tighter amount tolerance or a semantic similarity heuristic will all produce a different event count from the same raw file.
Step four: recompute the totals and compare
Sum the priced cash amounts per event, grouped by company, preserving approximate flags rather than converting them to exact values. Then compare that sum with the curated totalFundingUsd in the registry. They are not required to match. The registry value is a company-level aggregate that can include rounds outside the published event window; the event sum is a row-by-row account of what the site published.
Company profiles print both numbers and the difference, together with the build-time additions from auto-detected rounds. If your event sum matches the profile's event-cash figure but not the profile's headline total, you have probably reproduced the site correctly and found the exact place where the two views legitimately diverge.
Finally, compare the ranked order. Only companies with a disclosed cash figure receive a rank; a company with no priced amount appears in the directory but not in the ranking. Assigning it a position would invent a fact. This is why the number of ranking rows can be smaller than the number of companies in the registry.
What a genuine discrepancy looks like
A real error is a figure on a page that has no corresponding raw row, a rank order that contradicts the disclosed amounts, or a published event whose source does not support the instrument label. Those are worth reporting. A different total caused by including held rows, counting syndicated reports as separate rounds, or adding a policy letter of intent into priced cash is a difference in definition rather than an error.
The site is designed to make that distinction checkable. Every event has a source link and a note; every company profile has a reconciliation table, an instrument-mix table and a coverage-gaps list; the raw ledger is downloadable; and the scan log separates the data cutoff from the page generation date. None of that prevents mistakes, but it makes them findable.
When you do find one, send the URL, the figure, the source you compared it with and the date you downloaded the JSON. A correction that can be reproduced from a dated copy of the input is much faster to verify than a general claim that a number looks wrong.
Why we publish the procedure at all
There is a commercial argument for keeping the transformation opaque: readers who cannot recompute a number have to take it on trust. We take the opposite view. A funding ranking that invites scrutiny is more likely to be cited, more likely to be corrected, and less likely to quietly drift into a marketing artefact.
It also disciplines the editorial side. Once a rule is written down and the input is public, changing the rule to make a headline look better leaves a trace. The daily build, the hash-based sitemap dates and the repository history all make silent edits harder than a visible correction.
Finally, it is the most direct answer to the suspicion that a data site is just an automated aggregator. An aggregator cannot explain why a row was held, why three headlines became one event, or why a letter of intent did not enter the total. A dataset with a published procedure can.
Provenance of the rows in the current build
19 of 58 published events were first detected by automated news monitoring; they keep a visible ⚡ provenance flag and a clickable source URL. The other 39 were entered directly by the editorial desk. Every published event — regardless of how it entered the ledger — has its own hand-written note; rows still waiting for one are held back and never counted here.
| Company | Instrument & date | Disclosed amount | Source |
|---|---|---|---|
| Q-CTRL | Series C 2026-11-12 · Asia-Pacific |
~$290M⚡ auto-detected | sourcehow to verify |
| Taiyi Liangsheng | New round (undisclosed) 2026-10-10 · China |
Not disclosed⚡ auto-detected | sourcehow to verify |
| Quantinuum | Government funding (CHIPS R&D award) 2026-09-08 · United States |
~$100M⚡ auto-detected | sourcehow to verify |
| SpinQ | Series D 2026-06-30 · China |
~$140M⚡ auto-detected | sourcehow to verify |
| Taiyi Liangsheng | Pre-Series A 2026-06-17 · China |
~$42M⚡ auto-detected | sourcehow to verify |
| Atom Computing | Reported raise 2026-06-16 · United States |
~$300M⚡ auto-detected | sourcehow to verify |
| Origin Quantum | Reported raise 2026-06-08 · China |
~$420M⚡ auto-detected | sourcehow to verify |
| Quobly | Reported raise 2026-06-04 · Europe |
~$124M⚡ auto-detected | sourcehow to verify |
| Quantinuum | IPO 2026-06-03 · United States |
~$1.7B⚡ auto-detected | sourcehow to verify |
| Photonic Inc. | Reported raise 2026-05-12 · North America |
~$200M⚡ auto-detected | sourcehow to verify |
| eleQtron | Series a 2026-05-07 · Europe |
~$61.6M⚡ auto-detected | sourcehow to verify |
| Quantum Motion | Series c 2026-05-07 · Europe |
~$160M⚡ auto-detected | sourcehow to verify |
| QuantWare | Reported raise 2026-05-05 · Europe |
~$164M⚡ auto-detected | sourcehow to verify |
| Turing Quantum | Strategic round (state-led) 2026-04-27 · China |
Not disclosed⚡ auto-detected | sourcehow to verify |
| SpinQ | Pre-ipo 2026-04-03 · China |
~$84M⚡ auto-detected | sourcehow to verify |
| Pasqal | Spac 2026-03-05 · Europe |
Not disclosed⚡ auto-detected | sourcehow to verify |
| Photonic Inc. | Reported raise 2026-01-07 · North America |
~$131M⚡ auto-detected | sourcehow to verify |
| Nu Quantum | Reported raise 2025-12-10 · Europe |
~$60M⚡ auto-detected | sourcehow to verify |
| Quantum Art | Series a 2025-12-10 · Middle East |
~$100M⚡ auto-detected | sourcehow to verify |
The practical takeaway
If you can reproduce a figure from the raw JSON, you can cite it with a clear conscience. If you cannot, the scan log, the methodology columns and the corrections address are the three places to start asking why.