Field notes

Who Am I Actually Dealing With?

Two Austrian companies, both registered in 1992, both listed as active. One filed 21 sets of accounts. The other has filed none in 34 years. Here is how we built the namespace that tells them apart.

Two Austrian companies.

Both are a limited partnership (Kommanditgesellschaft). Both were entered in the commercial register in 1992. Both are listed today as active (aktiv). If your supplier check asks “does this company exist and is it registered?”, both come back clean.

One of them has filed 21 sets of annual accounts, the most recent for the year ended 31 December 2024, and is financed 96.8% out of its own funds.

The other has filed nothing, ever. Its register file has not been touched since 1 August 1995. Its registered address is “5020 Salzburg” — a postcode and a city, no street. Seven other companies share it.

The short answer: we turned Austria’s company register into a namespace that answers “who am I dealing with” — six linked sources, 22 custom detectors, 21,475 companies and 363,252 findings so far. Every conclusion is a register fact or a filed figure. Three things it deliberately cannot tell you are named as loudly as the things it can.

This is the record of how it was built and what it can actually do.

The question this namespace exists to answer

Most company-data products answer “does this company exist?”. That is the question the register is designed to answer, and it is nearly useless on its own — as the two companies above demonstrate.

The question a business actually has is harder:

• Is this a real operating business, or a registration?
• Can it pay? Is it getting stronger or weaker?
• Who controls it, and what else do they control?
• Has anything changed recently that I should know about?
• And what can I not find out?

That last one turns out to matter as much as the rest.

How the namespace works

A namespace (Namensraum) in Classifyre is one self-contained workspace: its own data, its own detectors, its own investigations. This one is Austria’s company register. Nothing here is bespoke to Firmenbuch — it is the standard machinery pointed at a public register.

The shape of it, end to end:

Read top to bottom that is the whole product: something is fetched, it becomes an object, facts about it become findings, findings answer standing questions, and questions that turn out to matter become cases.

The loop that keeps it current

The register does not sit still. Measured on the change feed, a single week carries between roughly 3,300 and 6,000 entries — new registrations (Neueintragungen), deletions (Löschungen), changed addresses, changed management, freshly filed accounts.

So each source runs continuously and asks for two things at once:

• the change feed (Veränderungsfeed) — what actually moved recently, which is what keeps a counterparty check current;
• the full sweep (Volldurchlauf) — a resumable walk through all 326,871 Firmenbuch numbers, which is the only way to reach a company that has not filed anything since 1992 and therefore never appears in a feed.

Neither alone is sufficient, which is worth being blunt about: a feed-only namespace is permanently blind to dormant companies, and a sweep-only namespace knows nothing about the company someone registered last week — usually the one with the thinnest history and the most reason to be checked.

One subtlety that matters more than it sounds: because any single run covers only a slice of the register, each run declares that it saw only part of the universe. A company that was not in this run’s slice is not treated as deleted. Without that declaration, a partial scan quietly retires everything it did not happen to visit.

What you would otherwise have to build

The register is public. Anyone can call the same APIs. The work that is not obvious until you try it:

Identity that survives out-of-order arrival. Six sources run at six different speeds. GISA may describe a company the register connector has not reached yet. Every source computes the same string — company://at/firmenbuchnummer/<fn> — and an edge pointing at a company that does not exist yet is kept and joined the moment it does. No shared table, no ordering requirement, no lost link.

Facts as objects rather than columns. The legal form, the shell score, the distress verdict are not fields in a row — they are findings. That is the difference between data you can retrieve and data you can investigate: a finding can be filtered, attached to a case as evidence, matched by a standing question, and searched semantically. A column cannot.

Provenance that holds. Every derived figure carries a typed edge back to the filed document it was read out of, so “why do you say equity is negative?” is a click, not an argument.

Questions that keep answering themselves. An inquiry written once keeps matching companies ingested months later. The investigation does not go stale the day the scan finishes.

Gaps that are visible. Where a source refuses to answer, the refusal is recorded as a finding rather than a log line — so no conclusion can silently rest on an assumption nobody can see. This turned out to be the most valuable design decision in the whole namespace, and it is covered in full below.

The rest of this article is the specifics: where the data comes from, why it is six sources instead of one, what the detectors do, and two real companies run through the whole thing.

Where the data comes from

Everything is public, and all of it is Austrian primary source. No scraped aggregators, no resold databases.

The register has 326,871 companies. That OGD extract matters more than it looks: the Firmenbuch API has no “list all companies” operation, only lookup-by-number and a change feed. Without a published enumeration you can only ever see what recently changed — which is a permanent blind spot, because a company that has filed nothing since 1992 never appears in a change feed. The Salzburg company in this article is exactly that case, and we would never have found it without walking the full list.

Six sources, and why not one

The obvious build is one connector that fetches everything per company. We split it into six. Each ingests independently and they link afterwards.

                    ┌──────────────────────────────────────────────┐
                    │  SHARED IDENTITY                             │
                    │  company://at/firmenbuchnummer/<fn>          │
                    │  document://at/firmenbuch/<key>              │
                    └──────────────────────────────────────────────┘
                                       ▲
        ┌──────────────┬───────────────┼───────────────┬──────────────┐
        │              │               │               │              │
   same_as()      same_as()      references()    references()   sets Asset.urn
        │              │               │               │              │
 ┌──────┴─────┐ ┌──────┴─────┐ ┌───────┴──────┐ ┌──────┴──────┐ ┌─────┴────────┐
 │    EVI     │ │   OeNB     │ │    GISA      │ │ Bilanz-     │ │  FIRMENBUCH  │
 │ publication│ │ regulated  │ │ trade        │ │ analyse     │ │  REGISTER    │
 │ profiles   │ │ institutes │ │ register     │ │ (per year)  │ │  companies + │
 └────────────┘ └────────────┘ └──────────────┘ └─────────────┘ │  officers    │
                                                       ▲         └──────┬───────┘
                                          flow(TRANSFORM)                │
                                          + field mappings          contains()
                                                       │                 │
                                              ┌────────┴─────────────────▼──────┐
                                              │  FIRMENBUCH FINANCIALS          │
                                              │  filed PDF + XML documents      │
                                              │  sets document://…              │
                                              └─────────────────────────────────┘

Three reasons the split earns itself:

They fail independently. GISA refuses per-company lookups to our API key. In a single connector that refusal would have taken the whole scan down with it. Split, GISA degrades on its own and the register keeps running.

They run at different speeds. The register is fast. The document source converts PDFs and is slow. The analysis source calls a model. Splitting lets each find its own cadence instead of everything moving at the pace of the slowest.

Order stops mattering. Because every source addresses companies by the same URN, none of them needs another to have run first. A GISA record can point at a company the register has not reached yet; the edge is kept as an external endpoint and stitched the moment that company arrives.

What each source contributes

• Firmenbuch Register — the anchor. Company, legal form, seat, registry court, first registration, business purpose, officers with birthdates and representation powers, and the full court event log (VOLLZ). It stamps Asset.urn, which is what everything else joins to.
• Firmenbuch Financials — the filed annual accounts as documents. Each filing exists twice, as a PDF and as a tagged XML. We ingest both and read the XML, because the figures are already structured — no OCR.
• Bilanzanalyse — one record per company per financial year: the balance sheet, the ratios, and the statutory crisis tests.
• GISA — trade-licence data where available, and where not, the refusal itself.
• EVI — an independent confirmation of register status. When two sources agree, that is corroboration; when they disagree, that is the interesting case.
• OeNB — which companies are regulated financial institutions, with LEI codes.

How the sources actually link

Two mechanisms, both boring on purpose.

Identity is a string both sides can compute. Every source derives company://at/firmenbuchnummer/<fn> from the Firmenbuch number it already has. The register stamps it on its company assets; everyone else points at it. No shared table, no lookup service, no ordering requirement.

Facts become findings via TAG detectors. This is the part worth explaining.

A TAG detector runs nothing. It is a declaration that a fact exists and what it is called. The connector already knows the company’s legal form, its registry court, its shell score — it read them from the register. Re-deriving those with a classifier would be slower and worse.

So the connector asserts them directly:

Asset(
    id=f"company-{fn}",
    urn=company_urn(fn),
    tags={
        "fb_status": "aktiv",
        "fb_legal_form": "Kommanditgesellschaft",
        "fb_shell_risk": "hoch (6/12): keine Jahresabschlüsse trotz "
                         "Offenlegungspflicht; ...",
    },
)

Each tag becomes a finding with that detector’s label and severity. Why bother, when the data is already in the asset’s metadata? Because a finding is a first-class object in Classifyre and metadata is not. A finding can be filtered, attached to a case as evidence, matched by a standing question, weighted by the correlation engine, and embedded for semantic search. Metadata just sits there.

The choice, concretely: push a fact to metadata and it is retrievable; make it a TAG finding and it is investigable.

The detectors, and why these ones

22 custom detectors. Every one exists because a specific question needed it.

Identity and structure — fb_status (register status), fb_legal_form, fb_jurisdiction (registry court), fb_company_age, fb_onace (industry), fb_size_class, at_company_ids (a regex over FN / UID / EUID / LEI).

Substance — fb_shell_risk scores twelve register facts: struck off ex officio, no accounts despite a filing duty, dissolved within two years of formation, address shared with three or more registrations, sole officer with sole representation, no line of business, no deliverable address. fb_filing_gap measures how overdue the accounts are.

Financial health — fb_distress implements the statutory tests: book overindebtedness (buchmäßige Überschuldung, negative equity) and the URG §§ 22–24 reorganisation test, which needs an equity ratio below 8% and a notional debt-repayment period above 15 years — both limbs, or neither. fb_capital_thin catches partly paid-in share capital. fb_solvency_outlook is the one AI detector, giving a forward-looking read.

Control — fb_officer_role, fb_representation, fb_mandate_power (how many companies one person runs at once).

Change — fb_register_churn, counting court entries in the last three years. Frequent movement — seat changes, management turnover, a name swapped — is a different signal from a lifetime event count, which says nothing: a 40-year-old company has many events and a clean history.

And the blind spot — fb_trade_access records GISA’s refusal as a finding. This one deserves a note. It would have been easy to log the failure and move on, and every case in the namespace would then have quietly rested on an assumption nobody could see. Instead the gap is an object with 4,563 matches, attached to the cases it constrains, so no analysis can silently assume “this company has no trade licence” when the truth is “we were not allowed to ask”.

Lineage: how a number keeps its provenance

A figure that cannot be traced back is not evidence. Every derived value here carries a typed edge to what it came from.

Take one company’s 2024 accounts:

company://at/firmenbuchnummer/008316f          (Firmenbuch Register)
  │
  ├── contains() ──▶ document://at/firmenbuch/<key>   Jahresabschluss 2024 (PDF)
  ├── contains() ──▶ document://at/firmenbuch/<key>   Jahresabschluss 2024 (XML)
  │                       │
  │                       └── flow(TRANSFORM) ──▶ Bilanzanalyse record 2024
  │                              + field mappings: Bilanzsumme, Eigenkapital, …
  │
  ├── ◀── same_as() ────── EVI profile
  ├── ◀── references() ─── GISA lookup (refused)
  └── ◀── uses() ───────── each officer, one edge per mandate

The flow(TRANSFORM) edge carries field mappings — which balance-sheet line produced which output field. So “equity ratio 96.8%” is not a number in a database. It is a number with a path back to the XML element in the filed document that produced it, and from there to the company.

Across the namespace, that currently resolves to:

The glossary, which did more than expected

Austrian company law has vocabulary that is meaningless to a non-specialist and unambiguous to a specialist. The register writes GES, E, K, W. A document says Offenlegungspflicht. The XML says HGB_224_3_A.

We loaded 55 terms with their aliases — German, English, and the register’s own codes:

Two things it changed. Search stopped depending on which language the analyst thinks in — a query for “equity ratio” reaches records that only ever say Eigenmittelquote. And single-letter register codes resolve correctly: E means Einzelvertretung, not “the letter E”, which matters when a code that short would otherwise match half the corpus as a substring.

Notably, we put no company names in the glossary. Names are data, not terminology, and a glossary that learns entity names starts making the correlation engine confident about things it should be questioning.

The cases

Cases are questions with evidence attached, not conclusions. Each one carries its own limits in its description, so nobody can read past them.

Shell companies without substance (Scheinfirmen ohne Substanz)

Which registered companies look like registrations rather than businesses. Evidence is register fact, never inference: an ex officio deletion is the court recording that it found no assets; a nil filing history in a form that carries a disclosure duty is a breach of a duty that exists from year one.

One hypothesis in this case is recorded as REFUTED — a shared registered address is not by itself a shell signal, and the case says so with the counter-examples that showed it. A business park, a serviced office and a formation agent look identical from the register. It is worth one point out of twelve and no more.

Early warning: insolvency and reorganisation (Frühwarnung)

Two independent readings of the same filings, deliberately kept apart: the arithmetic rule (negative equity; both URG limbs) and the AI detector’s trajectory read. Agreement is corroboration. Disagreement is the lead worth working.

Influential officeholders (Einflussreiche Funktionsträger)

Who controls many companies at once, measured from the register’s own records rather than press coverage. Each person is a real asset with a stable identity (name plus birthdate) and one edge per mandate. The interesting overlap is with the shell-company case: a person whose mandates cluster in companies that never file is a different story from one whose mandates are in filing, capitalised companies.

Risers (Aufsteiger)

The other direction — balance sheet and equity both growing with the equity ratio intact. Kept as its own case because the failure mode is different: here the risk is mistaking a one-off capital injection for a trend, which is why the AI detector must name a counter-argument in every prediction.

Counterparty check (Geschäftspartnerprüfung)

The case the namespace exists for, and the last one we wrote — the earlier four are all questions asked of the corpus by an analyst. None of them was “I am about to trade with this company.” It covers five dimensions: existence and standing, substance, financial health, stability, and who is behind it.

Two worked examples

Both real, both live in the namespace.

A. GARANT Austria GmbH & Co KG (FN 008316f) — assessable

Vienna, registered 28 January 1992. Wholesale (Großhandel, ÖNACE 46150).

The balance sheet is shrinking about 11% a year. Read alone that looks like decline. Read together with the other two rows it is the opposite: the equity ratio rises and liquidity gains 41% over the same period, which is only arithmetically possible if liabilities are falling faster than assets. That is deleveraging, not deterioration.

The case records it as a hypothesis with a testable predicate — if this is deleveraging, equity ratio and liquidity must move up while the total moves down — and states what would refute it: an equity ratio below 90% in the next filing, or liquidity turning down.

Alongside: 21 annual accounts over 32 years, current filing gap, shell score 2/12, no insolvency event. Filing discipline is the strongest signal the register offers, because filing costs money and exposes the filer. The accountant is named in the document. The prior-year column is checkable against the previous filing. A company with nothing behind it does not do this 21 times.

B. Videozentrale Verwaltungsgesellschaft m.b.H. & Co. KG. (FN 028409d) — not assessable

Salzburg, registered 21 August 1992. Business consultancy (ÖNACE 70100).

The conclusion is deliberately narrow, and it is not an accusation. Each fact has an innocent explanation available — a dormant holding vehicle parked for decades is entirely lawful. What the facts establish is something else:

There is no published basis on which to assess this company at all.

No accounts, so no equity, no liquidity, no trend. No register activity to suggest anyone is administering it. No address at which it can be found. For a counterparty decision that is a negative outcome, not a neutral one.

And note what is not claimed: GISA refused the trade-licence lookup, so “this company holds no licence” is not available as evidence, and no hypothesis in the case is allowed to assume it.

Why the pair is the point

Same legal form. Same year of registration. Same register status. A check that stops at “is it registered?” returns YES for both.

Everything that separates them came from asking better questions of data that was public the entire time.

What it cannot do

Stated as plainly as the rest, because a due-diligence tool that implies more than it has is worse than one that admits less.

No profitability. The filing we receive is a balance-sheet extract (Auszug aus der Bilanz). Small companies under § 242 UGB are not required to publish a profit and loss account at all. For most of the register there is no revenue, no profit, no margin — and no model should be used to invent one. Every ratio here is a balance-sheet ratio.

No trade activity. GISA refuses per-company lookups to this key. It is an entitlement question, not an expired credential.

Employee counts, almost never. Headcount appears only in a free-text annotation in the older filing taxonomy — 5 findings across 645 analysed filings. Per-employee ratios are therefore not meaningful here.

Partial coverage. The register walk is still running. For any given company the honest answer is either the data or “not scanned yet” — never “nothing found”. On a partially-swept register those are very different claims.

How it was built

Almost entirely through Classifyre’s own MCP server. The connectors are Python notebooks written and revised through the same tool surface that reads them back — create a source, push a cell, run it, read the findings, adjust. The cases, inquiries, detectors and glossary in this article were all created the same way.

Two things that came out of that:

The connector should assert what it already knows. Every fact the register hands over — legal form, court, officer roles — becomes a TAG finding rather than something a classifier re-derives from prose. It is faster, and it is exactly right instead of approximately right.

A blind spot has to be an object. The single most useful design decision was making GISA’s refusal a finding. A gap you can attach to a case constrains the analysis. A gap in a log file does not.

See it yourself

The namespace is live, with every detector, inquiry, case, hypothesis and lineage edge described above:

showcase.classifyre.com/firmenbuch-austria

Start with the two worked cases — Geschäftspartner-Prüfung A and B — and follow the evidence back through the lineage to the filed XML it came from.

A detector match is not proof of wrongdoing, and an open case is a question with evidence attached, not a conclusion with a question mark added afterwards.

More case files

Case file

Sep 22, 2026

Four Hundred Districts, Ten Years: Telling Real Growth From the Kind That Isn't

Three districts around Berlin grow far faster than their structure predicts. Nine look like they are under housing pressure — seven of them are not. A regional picture of Germany where every number comes with what it can and cannot support.

Open case file→

Case file

Sep 22, 2026

The Quarter Before the Numbers Say So: Early Warning From Public Statistics

Schleswig-Holstein's company insolvencies rose 36.6% in one quarter. Building permits collapsed two years earlier. And the signal everyone treats as leading turned out to lag. What monthly public data can and cannot warn you about.

Open case file→

Case file

Sep 22, 2026

Köln Is Not Shrinking: When the Official Number Is Not the Number

Germany's official population figures say 44 districts shrank over the last decade. Once the 2022 census correction is removed, those same districts grew. Here is the namespace that tells the two apart — and what else it found hiding in plain sight.

Open case file→