Evidence

Check our work.

Every company-data vendor says its scoring is intelligent and its data is fresh. Almost none of them will show you the table, the cut-off, or the moment they stopped looking. This page is those three things.

Exhibit one

Three rounds.
The record moved first.

Three European companies that announced funding this year. For each one, everything our record already held before the announcement — with the dates, and with the rows that were never certain marked as never certain.

How to read a date here

Event date. The day the thing happened — the article, the post, the round.
Measured on. The day we saw it. The change happened at some point before this date, not on it.

Lucend

Netherlands 10 days before the announcement
  1. 22 Dec 2025 Signalled intent to raise Not confirmed

    Turned up at an investor event to meet capital. Nobody announced anything — we inferred it, and stored it as an inference.

  2. 1 Jan 2026 Raised a round

    $3.3M seed.

Formerly Coolgradient; the company renamed before the round. Two earlier award rows in the record still carry the old name, which is why a name match alone is not enough to follow a company.

SOUS

Netherlands 62 days before the announcement
  1. measured 23 Jan 2026 Open roles

    Roles appearing on the careers page.

  2. measured 6 Feb 2026 Homepage repositioned

    99.2% of the homepage content had changed since we last read it.

  3. measured 20 Mar 2026 Open roles
  4. measured 23 Mar 2026 Homepage repositioned

    A second near-total rewrite, three days before the announcement.

  5. 26 Mar 2026 Raised a round

    €4M seed.

Every row before the round is a background-tier signal — individually worth 1 or 2 on the weighting opposite. On its own, none of this is a reason to call anyone. That is exactly why the weighting exists.

Ventory

Belgium 136 days before the announcement
  1. 27 Dec 2025 Won a named customer

    New customers in MedTech.

  2. 20 Jan 2026 Senior hire

    A business development hire — a new decision-maker, often with a mandate to change tooling.

  3. 20 Jan 2026 Achieved a certification

    ISO/IEC 27001. Compliance certifications often precede an enterprise sales motion.

  4. measured 27 Jan 2026 Homepage repositioned
  5. 12 May 2026 Raised a round

    €2.65M, led by KBC Securities.

The strongest of the three, and the one with the least drama in it: a customer win, a senior hire and a certification, months apart, all notable-tier. A market this size produces that pattern constantly — reading it is the work.

None of this is a prediction.

We did not know these rounds were coming, and a page that claimed otherwise would be the exact thing this one exists to argue against. The claim is narrower and it is checkable: before each announcement, our record already carried dated movement at that company — a hiring push, a full homepage rewrite, a certification, a senior hire, an investor event. Whether that movement was worth a conversation is a judgement. Making that judgement, on a market too big to read by hand, is the product.

SOURCE · signal rows read from the production record on 14 Aug 2026 · every row above is reproducible against the company profile it belongs to

Exhibit two

A rumour, marked
as a rumour.

The earliest row in the strongest of those three timelines is not an announcement. Nobody said they were raising. What we saw was a company turning up at an investor event, and what we wrote down was that we had inferred it.

We read every published signal taxonomy in this market. Not one of them separates an announcement from an inference. Ours does, on every row, in the field itself — which means a model reading our data downstream cannot quietly promote a guess into a fact.

Ten days later the round was announced and the row that carried this warning turned out to be right. It kept the warning anyway.

One row, as stored Not confirmed
"date": "2025-12-22",
"label": "Signalled intent to raise",
"category": "funding",
"tier": "major",
"unconfirmed": true,
"confidence_note":
   "Rumour or inference, not an
    announcement. Do not relay as
    established fact.",
"meaning": "Signalled an intent to
   raise; the round has NOT closed."
Exhibit three

Detection is the cheap half.

Finding events is not hard any more. The expensive half is deciding what never reaches you — and being able to read that decision. This is the actual weighting, rendered straight from the configuration file the product runs on. If an engineer changes a number, this page changes with it.

critical 4 of 5 shown
Shutdown 10
Shutdown Risk 10
Was acquired 9
IPO Announced 9
major 4 of 6 shown
Raised a round 8
Acquired a company 7
Senior departure 7
Signalled intent to raise 6
notable 4 of 17 shown
Launched a product 5
Announced an expansion 5
Announced a partnership 5
Senior hire 5
background 5 of 13 shown
Pivot 3
Acquisition Rumor 3
Open roles 2
Open role in a key function 2
Attended an event 2

SOURCE · config/signal_insights.php · read at page render, not transcribed

Weighted, not counted

Turning up at a conference is a 2. A closed round is an 8. Counting rows instead of weighting them would let the noisiest event type win every week.

Thresholded at 75

An event becomes work for you only above a score of 75. Below it, it stays in the record as context — and the record says so rather than hiding it.

Suppressed, not repeated

Decay windows and per-account suppression stop the same company taking a slot twice. When you dismiss a card, the reason is written back and the weighting learns from it.

Dropped in the open

Everything cut stays queryable through the API. We are not hiding it from you — we are refusing to send it.

Not a black box: everyone says their AI decides what matters. Ours is a table you can read and a threshold you can check.

Exhibit four

Announcements are published.
Changes are not.

Any assistant can look up a company you name. None of them can tell you which one, or what moved.

One company, two scans acme.com
12 May11 AugΔ
employees 124 163 +31%
open roles 4 19 +375%
followers 8,240 9,110 +11%
homepage

“built for enterprises” “built for developers”

No page publishes these four lines. They are computed, not written.

The difference was never a document

Retrieval finds what someone published. Nobody publishes a delta — it is computed against a record, not written down and indexed.

It cannot be back-filled

Nobody can tell you what a homepage said in March unless they crawled it in March. The record only grows forwards.

The announcement is the easy half

Anyone can find a funding round once they know whose round to look for. Which company, how fast, and whether it is a pattern — that is the record.

This is the part a better model does not solve: a larger context window cannot see a state it never stored. We run on the same models you do — the record is the part they cannot bring.

Figures schematic · mechanism live — monthly change computed against the previous month's record

Exhibit five

When we haven't looked,
the answer says so.

The dangerous failure of a company-data source is not a wrong answer. It is an empty one. A blank result reads as “nothing happened here”, when what it usually means is “nobody checked”. A person might catch that. A system running overnight will not.

So every answer we return carries how recently we observed that company and a flag saying whether the absence of signals means anything at all. When it doesn't, whatever is reading us is told to say so rather than guess — and to ask for a refresh instead of writing the account off.

How this reaches an agent

Attached to every answer

"coverage": {
  "monitoring": "periodic",
  "signal_absence_is_meaningful": false,
  "interpretation":
     "We check in on this company rather
      than following it continuously, so
      its record is not guaranteed to be
      complete right now. Do not read an
      empty list as a quiet period.",
  "refresh_recommended": true
}

A real block, returned for a real company on 14 Aug 2026

One analytics cookie (Google Analytics). Cookie Policy