What the law gives, what it withholds, and what is built in its place.

Technique is not property.

In Bikram's Yoga College of India v. Evolation Yoga, 803 F.3d 1032 (9th Cir. 2015), three grounds were pressed for protecting a sequence of physical movements — expression, compilation, choreographic work — and three were rejected.

Effort earns nothing on its own. Feist, 499 U.S. 340; CCH Canadian, 2004 SCC 13 ¶16; Football Dataco, C-604/10 ¶42; and 35 U.S.C. §103 settle the same rule in five independent places: the system rewards originality, not labour. Fifteen years in a seat is a fact about a person's life. It is not a fact the intellectual property system is equipped to notice.

CDPA s.3(3) separates the work from the record of the work: two objects, two questions. Directive (EU) 2016/943, at Recital 14 with Art. 1(3)(b) and Art. 3(1), excludes from trade secret protection “the experience and skills gained by employees in the normal course of their employment”, and permits observation, study, disassembly and testing. A person is not a container.

18 U.S.C. §1839(6) puts reverse engineering and independent derivation outside improper means, and under §1836(b)(3)(A)(i)(I) an injunction may not prevent a person entering an employment relationship: any condition must rest on evidence of threatened misappropriation “and not merely on the information the person knows.”

Credit is built so it cannot become property. Under the Copyright Act (Canada) s.14.1(2), “Moral rights may not be assigned but may be waived in whole or in part.”

One statute pays for non-patentable know-how.

The German Gesetz über Arbeitnehmererfindungen (ArbnErfG) of 25 July 1957 is still in force.

§ 3 defines a technischer Verbesserungsvorschlag: a technical innovation expressly not patentable and not registrable as a utility model. Non-patentability is the entry condition, not a bar. § 20(1) gives the employee a claim to reasonable compensation as soon as the employer exploits such a suggestion, where it gives the employer a preferential position similar to an industrial property right. § 22 makes that claim non-waivable in advance; it may be bargained over only once the suggestion is already in the employer's hands. The ministry's own compensation guidelines, at guideline 29, put the operative test at non-imitability, and their first example is a secret process.

The limb usually left out is not left out here: compensation ends once the innovation becomes known widely enough that competitors may lawfully use it.

The limit belongs in the same breath. § 20 confers no property right. It is a claim to money against one employer. It cannot be sold, licensed or asserted against anybody else, and it falls away when the technique becomes generally known. A legal system can pay for secret, non-patentable know-how. That is not the same as owning it.

One reading is left open: whether von Dritten nicht nachgeahmt werden können means a technique that cannot in practice be copied, or one that is kept secret. That is a question for German counsel.

The holder gets nothing.

There is no provision, in any instrument surveyed, conferring on a holder of personal data any positive entitlement in respect of it. No right to exclude. No right to be compensated for another's use. No term. No register of any such right.

The gap shows at the third party: if a recording of a worker escapes, the person with a remedy is the worker, not the holder.

The strongest such right is switched off for machine records. Under the Data Act, Regulation (EU) 2023/2854, Article 43, applying from 12 September 2025, the database right “shall not apply when data is obtained from or generated by a connected product or related service” within the Regulation's scope. Whether a heavy industrial machine with telematics is a connected product within that scope was not established, and is not asserted here.

A promise to a person in a seat therefore has to be built — in the architecture and in the paper — or it does not exist.

The movement is the identifier.

PIPEDA s.2(1), Alberta PIPA s.1(1)(k), BC PIPA s.1 and GDPR Art. 4(1) each turn on identifiability. None contains any concept resembling authorship or contribution. The person recorded is protected because they are identifiable in the record, not because they made what is in it.

Nair, Guo, Mattern, Wang, O'Brien, Rosenberg and Song, Unique Identification of 50,000+ Virtual Reality Users from Head & Hand Motion Data, 32nd USENIX Security Symposium, 2023, worked from ordinary play across tens of thousands of users, with no biometric enrolment, and identified individuals from about a hundred seconds of motion at a rate above nine in ten. The figure is taken as an order of magnitude. No study at comparable scale exists for gait, and none is read across to it.

The consequence: a movement trace with the individuating detail taken out is a movement trace with the skill taken out. There is no version of it that is both useful and stripped of identity. No claim is made anywhere that a record has been cleared of its identifiers.

Why the person's paper is elsewhere.

EDPB Guidelines 05/2020, at ¶¶21–23: “Given the imbalance of power between an employer and its staff members, employees can only give free consent in exceptional circumstances, when it will have no adverse consequences at all whether or not they give consent.”

That is why this register does not contract with the person. The person's paper is with Even Steven — not with their employer, and not with Tacit Source — signed in their own seat, on their own machine, on their own time.

California Civil Code § 1798.125(b) permits a business to offer financial incentives “including payments to consumers as compensation” for the collection of personal information, subject to prior opt-in consent, a bar on terms “unjust, unreasonable, coercive, or usurious in nature”, and a requirement that any difference be “reasonably related to the value provided”. Since 1 January 2023 a California employee is a “consumer”.

The question, in one sentence: does paying a person make their consent valid? It is open. No jurisdiction surveyed provides that it does, and one regulator has objected to paid biometric consent on reasoning that is not established. The answer here is structural. The payment follows the hour the machine runs, not the signature, and the arrangement is structured as a fee for a performance from which a recording results — not as a sale of data, the shape that invites the objection directly.

Two notice regimes, and one instructive contrast.

Canada Labour Code Part I, ss. 51–54. Section 52(1) requires at least 120 days' notice to the bargaining agent before a technological change, with five prescribed contents; s.53 allows the Board to order the change ceased for a further 120 days. Federal undertakings only.

British Columbia's Labour Relations Code s.54 sets 60 days, on a broader trigger, with no reinstatement remedy.

Ontario is the contrast. ESA s.41.1.1 requires an employer with 25 or more employees to have a written electronic-monitoring policy, and Ontario's own guide says the requirement does “not establish a right for employees not to be electronically monitored” and does “not create any new privacy rights”. A disclosure duty, not a limitation duty. The law makes an employer say it. It does not make them stop.

What can be checked in a delivered model.

Thudi, Jia, Shumailov and Papernot, On the Necessity of Auditable Algorithmic Definitions for Machine Unlearning, USENIX Security 2022.

The restatement: unlearning is not a property of a model. It is a property of a history, and histories are attested, not measured. Every offer to check a delivered model for compliance is offering something that cannot exist.

The positive half: a right to withdraw can be honoured today — in a database, in a retrieval index, in a small model, and in anything kept separately addressable. It cannot be honoured, in any way a second party can check, in a large generative model built the ordinary way.

Separability is therefore not a feature of this register. It is the condition on which every other promise here can be kept. Withdrawal itself is one sentence: It stops for new work at once; a job already underway finishes.

One deployed scheme already pays named contributors per use.

In it, a person contributes under their own name and is paid for each use, with payouts made weekly and internationally. The company is not named here; what matters is the rule it demonstrates.

Per-use payment to an identified person works exactly where that person's contribution stays a selectable, addressable artefact. Once a contribution is one of millions dissolved into a general model's parameters, the configuration is gone, and with it the counting.

Three qualifications are owed. It is offered in place of a right: the contributor waives royalties and equivalent claims. The payer sets the rate, on “factors that we determine”. And a contributor can check that they were paid, but not that they were paid correctly. That is why the count in this register is checkable at source rather than reported after the fact.

One limit is owed in return. In machine learning, attribution means tracing a model's output back to the material it was built from. Whether per-contributor attribution can be instrumented on a corpus small enough to individuate its items has not been attempted and reported by anyone, and no claim is made that it has. It is not needed: nothing is dissolved. A skill set is chosen by name and counted by the hour it runs — a counting problem, not an attribution problem.

A released package is fixed.

Directive (EU) 2024/2853 applies from 9 December 2026 and, under Art. 2(1), to products placed on the market after that date. A physical machine is unambiguously a product.

Art. 7(2)(c) requires a court to take into account “the effect on the product of any ability to continue to learn or acquire new features after it is placed on the market or put into service.” Recital 32: “a manufacturer that designs a product with the ability to develop unexpected behaviour should remain liable for behaviour that causes harm.” Art. 4(18) with Recital 40 provide that a substantial modification can arise “due to the continuous learning of an AI system”, making the modified product newly placed on the market at the moment the modification is made — Art. 8(2) making the modifier a manufacturer in its own right, and Art. 17(1)(b) restarting the ten-year period.

Stated correctly: continuous learning can be a substantial modification where the two-part test is met. It is untested. No judgment anywhere has addressed whether routine incremental refinement crosses the threshold.

A released package is fixed. It does not learn in the field. A skill set does not change on a machine without a person putting it there.

Two companies, because a policy is not a structure.

Data Governance Act, Regulation (EU) 2022/868. Art. 11(1) requires a data intermediation service to act as an intermediary and not to use the data it exchanges for any other purpose. Art. 11(2) imposes pricing neutrality. Art. 11(3) requires the service to be provided through a separate legal person from the provider's other activities.

The Act forces a choice between being the pipe and being a user of what flows through it, and forces that choice to be structural rather than a policy a provider states it will follow.

No assertion is made here that this register is a data intermediation service under the Regulation; that is an open question for counsel. The separation between the house that represents the person and the register that holds the number, the licence and the count is built the way the Act would require of one. Neither can quietly mark its own homework.

The ownership line sits inside that separation and is stated the same way on every track. Even Steven owns what is made of every skill, for the person or the company it came from, and holds the consent behind it; a Source's paper is with Even Steven and with nobody else. Steady Eddie and Strategic Agent collect skill sets and require them. Tacit Tessera issues the number, pins the edition, licenses to OEMs and to AI, and keeps the count. No entity in that chain owns a person's skill: the record is held under licence and never bought.

An open question, in three steps.

A reference built from a professional's practice files raises a question of title that nothing surveyed answers. It is put here in the same three steps as every other open question.

The question. When an expert has spent a career employed, whose documents are the practice files — the expert's, the employer's, or the client's?

It is open. An employed expert's working files are usually the employer's, and often the client's, under the engagement terms and the ordinary rules on work made in the course of employment. No court has decided what an individual may license out of such a body of documents, and no regulator has answered it. Nothing here asserts an answer.

What is built. The question is asked before anything is assembled, not after. Licensing a practice's own files is clean for an owner-principal, where title and consent sit in the same hands; in at least one profession it is known to be blocked; counsel rules the rest, and a licence does not cure a missing consent. Where the files are somebody else's, the Source is the company and the company signs — the reason the Records track has two kinds of Source. Authorship is recorded before any mined reference is licensed per person, because a receipt says where a passage came from and not who wrote it. And in every case the paper with Even Steven names what is being licensed and by whom, so a defect in title is a defect that surfaces before a serve rather than after one.

Paid for contributing.

Under 17 U.S.C. § 114(g)(2), of the statutory licence receipts, 45% goes to the featured artist and 5% to non-featured performers, paid directly. Under § 115(d)(3)(J)(iv), “in no case shall the payment or credit to an individual songwriter be less than 50 percent”, “notwithstanding any agreement to the contrary.”

The five per cent goes to session players who hold no copyright at all. The statute pays them because they contributed, not because they own.