The problem is not spam, it is syndication · Why a ten-day window · The three conditions that collapse two rows · The score that decides which row survives · What the reader can actually see · Where the rule fails, and what we do about it
The problem is not spam, it is syndication
Duplicate headlines are usually not an attack and not a scraping error. They are the ordinary distribution chain of a press release: the company publishes, a wire service picks it up, specialist outlets rewrite the same facts with different headlines, and regional editions translate it. Each version is a real article with a real URL, and a naive collector would treat every one as a separate financing.
The damage is not limited to an inflated count. If three copies of a round enter a total, the total grows by three times the same money. If they enter a company profile, the reader sees what looks like a funding streak rather than one event. And if they enter a feed, subscribers get the same news three times. De-duplication is therefore a correctness requirement before it is a tidiness preference.
Our rule is intentionally conservative. It would rather keep two rows that might be distinct financings than delete a real one, because a survivor can still be corrected by an editor while a deleted row is invisible. Every merge we do perform is recorded as a count on the surviving row, so the compression is visible rather than silent.
Why a ten-day window
Reports are grouped by company, then compared inside a ten-day window measured from the event date. Ten days is long enough to catch a wire story, a weekend rewrite and a translated regional follow-up, and short enough that two genuinely separate rounds from the same company are unlikely to fall inside it. Companies in this sector do not normally price two equity rounds a week apart.
Date precision matters here. When a report gives only a month, the ledger assigns a mid-month reference point for comparison; the published date still shows the month-level precision it actually has. This keeps an imprecise row from being excluded from the comparison while refusing to pretend we know the day.
The window is a heuristic, not a law of nature. A long-delayed exclusive about a round first reported three weeks earlier will not be merged by this rule. That is one of the reasons the raw ledger remains public: a reader can see the second report and judge whether it is genuinely new.
The three conditions that collapse two rows
Two reports inside the window merge when at least one of three conditions holds. First, both amounts are undisclosed: no priced cash figure exists on either side, so there is no numeric contradiction to resolve. Second, both disclose an amount and the two figures agree within twenty percent, which absorbs currency conversion and rounded reporting without fusing genuinely different rounds.
Third, one record names a specific round while the other is a generic reported raise. A named instrument plus an anonymous summary of the same size is a common shape in syndicated coverage, and treating them as two financings would double count one announcement. In each case the merge is recorded on the surviving row as a corroborating report count.
Nothing here compares prose similarity. We do not hash headlines, measure word overlap or decide that two articles are the same because they sound alike. The rule uses company, date, amount and instrument label only. That makes its behaviour predictable and reproducible from the public JSON, which matters more than catching every last duplicate.
The score that decides which row survives
When rows merge, one of them has to carry the published event. We score each candidate: an editor-reviewed row ranks ahead of an automatically collected one, a named round ahead of a generic reported raise, a row with named investors ahead of one without, a row with a source URL ahead of one without, and a longer source description ahead of a shorter one. The highest score survives.
This ordering encodes a simple preference for the most checkable record, not for the most dramatic headline. It is also why an editor-reviewed row is not automatically the survivor: if a reviewed row lacks a named round and a source link while an automatically collected row has both, the scoring still favours the more complete record, and the provenance badge stays visible on the result.
Score ties are resolved deterministically by identifier, so a rebuild with identical input produces byte-identical output. That property is tested: our build audit runs a second render with a different build date and asserts that no page changes its content hash. A de-duplication rule that produced a different answer on Tuesdays would be worse than no rule at all.
What the reader can actually see
Three artefacts make the merge auditable. The published tables show one row per financing with a small merged-reports count where applicable. The raw funding-events.json still contains every original report, so the compression can be recomputed. And the data block at the end of this page lists every surviving event that absorbed another report, with its source link.
Company profiles go further: they show the number of syndicated reports merged into that company's rows alongside the count of published events, and they link each event back to its source. A reader who disagrees with a merge can follow the links, find the second report, and send a correction without needing access to our internals.
The feed uses the same editorial note as the page rather than the scraped headline. This avoids the odd failure mode where a subscriber sees three identical headlines on the same morning while the site claims to have de-duplicated its data.
Where the rule fails, and what we do about it
It fails when one company genuinely closes two instruments inside ten days, when a report gives a materially different amount for the same round, and when a late follow-up outside the window describes an event already published. It also fails when two different companies share a name and the ledger has mis-assigned an identifier, a problem no window can fix.
The mitigation is layered. Rows that do not meet any merge condition remain separate for an editor to inspect. Every published event requires its own hand-written note, which is the point at which a suspicious pair is most likely to be caught. Held rows never reach the totals. And the corrections path is advertised on every company profile rather than buried in a footer.
We have deliberately not replaced this with a semantic similarity model. A model would collapse more duplicates, but it would also be harder to explain, harder to reproduce from the public files, and much easier to fool with a rewritten headline. For a dataset whose main claim is that a reader can check it, a boring rule with published failure modes is the better trade.
Merged reports in the current build
The raw ledger has 111 reports. The publication layer publishes 58 distinct financings, so 53 syndicated reports were merged. Every raw report is still present in funding-events.json.
| Company | Date & instrument | Reports merged | Source |
|---|---|---|---|
| Nu Quantum | 2025-12-10 · Reported raise | 2 | sourcehow to verify |
| Quantum Art | 2025-12-10 · Series a | 1 | sourcehow to verify |
| Photonic Inc. | 2026-01-07 · Reported raise | 1 | sourcehow to verify |
| Pasqal | 2026-03-05 · Spac | 1 | sourcehow to verify |
| SpinQ | 2026-04-03 · Pre-ipo | 1 | sourcehow to verify |
| Turing Quantum | 2026-04-27 · Strategic round (state-led) | 1 | sourcehow to verify |
| QuantWare | 2026-05-05 · Reported raise | 4 | sourcehow to verify |
| eleQtron | 2026-05-07 · Series a | 1 | sourcehow to verify |
| Quantum Motion | 2026-05-07 · Series c | 3 | sourcehow to verify |
| Photonic Inc. | 2026-05-12 · Reported raise | 3 | sourcehow to verify |
| Quobly | 2026-06-04 · Reported raise | 3 | sourcehow to verify |
| Origin Quantum | 2026-06-08 · Reported raise | 2 | sourcehow to verify |
| Atom Computing | 2026-06-16 · Reported raise | 2 | sourcehow to verify |
| Taiyi Liangsheng | 2026-06-17 · Pre-Series A | 2 | sourcehow to verify |
| SpinQ | 2026-06-30 · Series D | 1 | sourcehow to verify |
| Quantinuum | 2026-09-08 · Government funding (CHIPS R&D award) | 1 | sourcehow to verify |
| Taiyi Liangsheng | 2026-10-10 · New round (undisclosed) | 1 | sourcehow to verify |
| Atom Computing | 2026-11 · Government funding (CHIPS Act LOI) | 1 | U.S. Dept. of Commerce announcement, MIT QIR Insightshow to verify |
| Q-CTRL | 2026-11-12 · Series C | 5 | sourcehow to verify |
The practical takeaway
If our event count disagrees with someone else's, check the raw ledger before assuming either side is wrong. In most cases the difference is not a missing round; it is a different decision about whether three syndicated headlines are three financings or one.