Miju LabsEverything else on this site is a judgement about a market. This page is a judgement about a property right, and it is the only place where being based in Europe is worth money rather than merely being a tax and payroll question.
This page is research, not legal advice. Nothing here is a legal opinion, and none of it should reach a contract, a filing or a board paper without qualified counsel instructed in each relevant jurisdiction. The statutes and cases are named so counsel can be pointed at the primary material quickly. That statement covers this page and the three that follow it.
The asymmetry
Article 7 of Directive 96/9/EC gives the maker of a database the right to prevent extraction or re-utilisation of the whole or a substantial part of its contents, where there has been substantial investment in either the obtaining, verification or presentation of those contents. It is not copyright. It does not care whether anything in the database is original. It is a pure investment-protection right, and it is exactly the shape of a corpus of expert judgements.
Article 11 restricts the right to nationals or habitual residents of a Member State, and to companies formed under the law of a Member State with a genuine link to the economy of a Member State (EUR-Lex 96/9/EC). Contra.Work Inc., Taste Labs, Intelligence/Design Arena, AfterQuery and Mercor are US companies. None of them can hold this right over its own corpus. A European operator holds a property right in the same asset class that its best-funded competitors cannot acquire at any price.
That has a corollary about group structure, and it is the sort of thing that gets decided badly by default. The maker is the party that takes the initiative and assumes the risk of investment (Fieldfisher). If the EU company is set up as a cost-plus service subsidiary of a US parent — the standard arrangement when the sales entity is in Delaware because the buyers are in San Francisco — then the party assuming the risk is American and the right may sit nowhere at all. The European parent with a US sales entity is the structure that preserves it. That decision is made once, cheaply, at incorporation, and is expensive to reverse after the corpus exists.
The trap, which the default operating model walks straight into
British Horseracing Board Ltd v William Hill Organization Ltd, Case C-203/02 (CJEU, Grand Chamber, 9 November 2004), holds that investment in obtaining means resources spent seeking out and collecting existing independent materials, and excludes resources spent creating the data in the first place. BHB spent heavily compiling pre-race data and lost, precisely because the money went into bringing data into existence rather than into gathering data that already existed (Fieldfisher; CaseMine on the Court of Appeal judgment).
A commissioned taste corpus is, on its face, created data. You pay a designer $90 an hour to produce a judgement that did not exist before they produced it. Applied naively, BHB defeats the right at exactly the moment you most want it — when a competitor or a former buyer has lifted a substantial part of your corpus and you are trying to stop them.
This is why the page matters operationally rather than theoretically. The right is available, and the default posture forfeits it. A company that spends three years commissioning judgements, records the whole spend as "data production" in one cost centre, and only asks about the database right when a lawyer raises it in diligence will find that the ledger says the money went into creation. A litigator reconstructing the split from a general ledger three years after the fact will lose.
What to do about it, starting in month one
Article 7 protects investment in obtaining, verification or presentation. Two of those three limbs are wide open, and both describe work you were going to do anyway. The whole trick is to spend it visibly, in its own ledger, from the beginning.
| Activity | Which limb | Where it should sit | Evidence to keep |
|---|---|---|---|
| Multi-rater adjudication of disagreement | Verification | Separate cost centre | Adjudication logs, per-item rater counts |
| Calibration sessions, gold-standard re-tests | Verification | Same | Session records, gold-set versions |
| Exclusion of unreliable raters; drift monitoring | Verification | Same | Per-rater reliability history |
| Computing and publishing Krippendorff's α per axis | Verification | Same | Dated coefficient runs |
| Schema, ontology and taxonomy design | Presentation | Separate cost centre | Design docs, schema version history |
| Normalisation, indexing, cross-linking | Presentation | Same | Pipeline commits, engineer time |
| Delivery API and documentation layer | Presentation | Same | Build logs, contractor invoices |
| Curating award archives and published work into reference sets | Obtaining | Third cost centre | Collection logs, source and date per item |
| Paying a designer to produce a new judgement | Creating — the excluded limb | The residual bucket | — |
Recital 55 of the Directive confirms the direction of travel: "a substantial new investment involving a new term of protection may include a substantial verification of the contents of the database" (EUR-Lex). Verification is not an afterthought in this instrument; it is one of the three ways the right is earned.
Run verification and presentation as separately-ledgered cost centres from the first month, with their own invoices, their own headcount allocation and their own timesheets — and add a curated "obtained" layer alongside the commissioned one. Everything the manufactured oracle already requires you to build — the adjudication protocol, the calibration set, the reliability reporting — is verification spend. You are doing the work. The only question is whether the accounting will be able to prove it.
The third move is the cheapest and the most under-used. Curating existing material is textbook obtaining. Award archives, published portfolios, competition entries and public reference sets, collected and structured, is exactly the activity BHB says the right protects. A corpus that layers commissioned judgements on top of a curated obtained tranche is materially stronger under Article 7 than one that is purely commissioned — and the award registers are already the seed list for recruiting, so the collection work has two justifications. It carries its own copyright and text-and-data-mining questions on the input side, which belong to Copyright is the weakest thing you own, but structurally it is the strongest limb available.
The rolling term is the part nobody exploits
Article 10 gives fifteen years from completion or from first making available. The provision that matters is the third paragraph: any substantial change to the contents, evaluated qualitatively or quantitatively — including one resulting from an accumulation of successive additions, deletions or alterations — qualifies the resulting database for its own fresh fifteen-year term (EUR-Lex 96/9/EC).
A corpus that is continuously refreshed — new panels, new phases, new axes, re-adjudicated older items — is in practice perpetually protected. The condition is evidentiary, not legal: keep dated records of each substantial refresh, with the volume added and the verification spend attached, so that the fifteen-year clock can be shown to have restarted on a specific date. This is a quarterly ten-minute filing discipline that is worth an unbounded term. It is also the strongest argument for the subscription-with-refresh shape of the held-out evaluation suite, which needs continuous re-investment for entirely commercial reasons and gets the term reset as a side effect.
What the right does and does not reach
Be precise about this, because it is easy to oversell internally.
It reaches extraction and re-utilisation of a substantial part of the contents by anyone — including parties who never signed anything with you. That is the whole point, and it is what contract cannot do. A buyer who leaks your corpus to a third party leaves you with a claim against the buyer; the database right gives you a claim against the third party as well.
It does not reach the buyer's model weights. The English High Court in Getty Images v Stability AI held that model weights are not infringing copies because they do not store the works: "by the end of that process they did not store any of those copyright works" ([2025] EWHC 2863 (Ch), 6 November 2025; BAILII). Germany's LG München I reached the opposite conclusion on memorisation five days later, which is its own problem — but no theory of the database right gives you a residual interest in a model trained on your data. Your corpus is protected by the contract you wrote and the right you preserved, never by a claim on what the buyer built with it.
It does not reach ideas or methods. Your rubric, your panel composition and your adjudication protocol are protected by copyright and trade secret respectively (Directive (EU) 2016/943 — EUR-Lex), not by Article 7.
One carve-out to have checked and then forget about, unless the product changes. Article 43 of the Data Act, Regulation (EU) 2023/2854, disapplies the sui generis right where data is obtained from or generated by a connected product or related service within that Regulation's scope (eu-data-act.com). A laptop running Figma is not a connected product in the IoT sense, so this should not bite. But the trajectory product has an obvious hardware extension — stylus telemetry, tablet instrumentation, wearable capture — and Article 43 has to be looked at before any of that ships, not after.
Worth noting for supply geography: Sweden's section 49 "catalogue rule" — protection for compilations of a large amount of information or resulting from significant investment — predates the Directive and is Sweden's implementation of it (Lexology). The Nordic jurisdictions are the friendliest on contributor terms as well as the database right, which is one input into where to concentrate recruitment — though not the deciding one, since the design talent is not distributed that way.
How much of this is actually settled
No court anywhere has decided a copyright, database-right or trade-secret case on a commissioned expert-judgement corpus. Every conclusion on this page is adjacent doctrine — horse-racing fixtures, coin prices, football fixture lists — applied to a fact pattern no judge has seen. BHB is settled law about the obtaining/creating line; whether adjudication of expert disagreement counts as verification of contents rather than creation of them is an argument, not a holding.
That is not a reason to ignore it. It is a reason to make the argument cheap to run: the entire cost of preserving the option is a chart of accounts decision and a filing habit. If it works, it is a right your competitors cannot hold. If it fails, you have lost nothing but some bookkeeping. That asymmetry is why this is the first constraints page rather than a footnote.
The same reasoning applies to how the asset is presented commercially. A right nobody has litigated is worth little in a pitch and a great deal in a diligence room, where the question is not "will you win" but "what does the buyer of this company acquire." A corpus with a documented verification ledger, a dated refresh history and an EU maker is a cleaner asset than a corpus with the same contents and no paperwork — and the difference costs almost nothing to create up front. That is the same logic that makes designing around the constraint from day one the right move in health data, and it is the reason both belong in the first ninety days rather than the second year.
The commercial companion to this page is Sell the paperwork with the data: the documentation you build for Article 7 is most of the documentation a buyer's compliance function is going to demand anyway. Build it once and sell it twice.