LegacyNext · Publishing Rights Infrastructure

Know what you licensed.
Pay the right authors.

An execution layer for book, journal and news publishers licensing content into the AI economy. It records which work it is, which rights you actually hold today, what is permitted, exactly which titles sit inside a licence, and how the money reaches authors, translators and estates afterwards. Every statement is written to a blockchain, so both sides of a licence can prove the same facts years later.

The Works List Problem The Proof Layer

Core objectWork rights passport

Proof layerBlockchain-anchored

ScopeBacklist & frontlist

OutcomeAuthor-level settlement

What changed

Text licensing became an evidence problem.

Courts, regulators and AI companies all moved in the same direction inside a year and a half. The operative question is now one of proof: what was permitted, under which contract, for exactly which titles.

Nov 2025

The Regional Court of Munich I held in GEMA v OpenAI (42 O 14139/24) that memorisation of protected text in model weights is an infringing reproduction, and that the text-and-data-mining exception does not cover it. The first European ruling of its kind, and it landed on text. Not final; appeal remains possible.

Jul 2026

The same court extended that reasoning in GEMA v Suno (42 O 763/25), holding that compliance with the AI Act does not by itself bring training inside the mining exception. Two rulings, one direction: European bargaining power moves toward rightsholders.

Since Aug 2025

Under Regulation (EU) 2024/1689, providers of general-purpose AI models must publish a sufficiently detailed summary of training content and operate a copyright policy respecting reservations made under Article 4(3) of the DSM Directive. A reservation that is not machine-readable and findable per work is close to worthless in practice.

2024 onward

Publishers began licensing directly. Taylor & Francis licensed content from roughly three thousand journals to Microsoft for an initial ten million dollars plus recurring payments, with Wiley and others following, and authors reported learning about it from the press. On the collective side CCC launched an AI training licence and PLS and CLA are building comparable schemes. Every one of those licences needs work-level data underneath it to settle honestly.

20 Jul 2026

In Bartz v Anthropic the court held that training on lawfully acquired books was fair use, and that retaining pirated copies was not. The settlement that followed, approved in July 2026, put 1.5 billion dollars against a list of individual books, and assembling and adjudicating that list after the fact turned out to be the hard part. See below.

Case in point

A works list, reconstructed too late.

The largest copyright settlement in US history turned into an identification exercise. The dispute was about how the copies were obtained; the expense was working out, title by title, who was entitled to what. It is the clearest available demonstration of what a catalogue costs when that question has no answer on file.

What it took

482,460 works had to be assembled into an eligible list, after the fact, from datasets nobody had documented at work level
440,490 were claimed by the March 2026 deadline, leaving roughly 42,000 works with no one stepping forward
Unclaimed works were removed from the fund before the per-work figure was calculated
About $2,931 per claimed work, out of a net fund of roughly $1.29 billion once some $208.6 million in fees and costs came off the $1.5 billion, with a default 50/50 split between author and publisher on trade titles

What a rights record changes

The list already exists, signed and dated, instead of being rebuilt under litigation pressure
Unclaimed becomes rare, because the rightsholder of record is attached to the work rather than discovered by campaign
Splits are data, drawn from the contract at registration, not litigated per title years later
Reverted rights are visible, so a publisher licenses only what it still holds

Every catalogue carries a works list already. The only question is whether it is written down now, or assembled later by lawyers under a deadline.

The gap

The break sits between the contract and the payment.

Publishing has excellent identifiers and poor answers. ISBN, ISSN, DOI and ISNI all work. What no system holds is the thing an AI licence actually needs: which rights you hold, in which languages, today.

Editions, not works

A single title exists as hardback, paperback, ebook, audiobook, book club edition and a dozen translations, each with its own ISBN. None of those identifiers describes the work, which is the level at which an AI licence operates.

Backlist and reversion

Agreements signed in 1998 or 2012 say nothing about training, mining or model weights, and whether those uses sit inside the granted rights differs per contract, per decade and per imprint. Meanwhile out-of-print and reversion clauses have been returning rights to authors for decades. Both answers live in scanned PDFs, which is a real exposure when a catalogue is licensed in bulk.

Language and territory

Rights are split by language and territory and sublicensed onward to foreign publishers. A training licence crosses every one of those boundaries at once, and the model has no concept of them.

Dividing a lump sum

A subscription counts downloads. A training licence pays a lump sum and counts nothing, so the split becomes a policy choice rather than a measurement. The percentages owed to author, co-authors, translator, illustrator, estate and agent are contractual and correct, and they sit in documents rather than in a system that can pay from them.

First product

The verified corpus manifest.

Every training or licensing agreement ships with a signed, hashed list of exactly the works inside it. For the AI company it is proof of scope. For the publisher it is the starting point for paying authors. Neither side has to trust the other's spreadsheet.

What the manifest carries

Work identity above the edition, with ISBN, ISSN and DOI attached rather than substituted for it
Contributors by ISNI: authors, co-authors, translators, illustrators, estates
Translation and edition lineage, so a licensed English edition does not silently carry the German one
The grant as it stood: which rights, which languages, which territories, which term
Consent and reservation per use type, including author-level consent where the contract requires it
Declared provenance: human written, AI assisted or AI generated, recorded rather than inferred
Hash and signature, anchored so any later change to the list is detectable

What it lets you prove

Scope: exactly which titles were inside the deal, and which were deliberately excluded
Authority: that you held the right you licensed, on the date you licensed it
Allocation: a defensible per-title split instead of a pool with an unexplained key
Reservation: that an opt-out was machine-readable and available before training, not asserted afterwards
Author position: what each author agreed to, when an author or a society asks

The proof layer

Why this runs on a blockchain.

A rights record is only worth something when more than one party will accept it. In publishing that means an AI licensee, an author, a society and eventually a court all reading the same thing. That is the part of LegacyNext that cannot be reproduced by writing a better database.

Tamper-evident by construction

Who declared what, under which contract, for which scope, at which moment, and when it changed. A grant cannot be backdated quietly, which is exactly what an author, an auditor and a licensee each need.

One token per work

Each work carries a unique, non-fungible token with a single continuous history across every edition and translation. Issuing is limited to the party holding the right and bound to the work's identifiers, so a duplicate registration is refused at source, and a competing claim runs through a documented conflict procedure instead of settling in whichever system is consulted first.

Cross-party evidence

The licensee proves scope, the publisher proves entitlement, the author checks their own consent and split. No party has to accept another's internal system as the source of truth.

A record, not a tradable asset

The token carries a rights record and represents no financial claim on revenue. It is unique, non-fungible and not offered on a secondary market, which is the ground on which counsel assesses it under MiCA. Where a structure does carry financial exposure, that is arranged separately under the applicable regime.

Standards stay

ONIX remains the exchange language, ISBN, ISSN, DOI and ISNI the identifiers, and reproduction rights organisations the collective licensors. The record feeds those systems.

The precise claim

The blockchain registers and proves; it does not transfer copyright. Dutch law requires a deed under Article 2 of the Copyright Act, and comparable formalities apply elsewhere. The deed stays. Being exact here is what makes the rest survive legal review.

What the chain delivers is a verifiable rights record. Ownership keeps its own legal route, and holding that line is what carries the rest through a legal review.

Build vs buy

The code was never the expensive part.

A rights database is the cheap half of this problem. The cost sits in reading decades of contracts into a grant model that holds, mapping reversion, handling language and territory splits, and ONIX conformance, and that is where in-house builds run aground in year two.

The real cost

Software is a minority of the work. The years go into reading four decades of contracts into a grant model that holds, mapping reversion, handling language and territory splits, ONIX conformance, and then operating it: key management, versioning a record that stays provable across changes, reconstructing a payment line by line when an author asks three years later. A wrong split arrives as a claim from an author.

Evidence needs neutrality

What you build yourself is your database. An AI licensee will not accept the publisher's own system as proof of scope, and an author will not accept it as proof of their split. A record becomes evidence when no single party controls it, which is a governance property rather than a feature added later.

Moving ground

The Munich rulings can still go on appeal, the AI Act's enforcement practice is still forming, and collective AI licences are being designed as we speak. Schema decisions taken this year will need revisiting as that settles. Buying the execution layer means someone else owns that migration.

Author trust

The reputational damage from the 2024 academic deals came from authors learning about a licence from the press. A consent register that an author can inspect is worth more than a communications strategy, and it is far harder to retrofit once trust has gone.

The boundary

You remain the licensor.

The work splits cleanly in two. LegacyNext offers itself for one half only, and that is the design principle rather than a concession.

Authority: stays with you

The licence decision: whether to license, to whom, and on what terms
Pricing: what a corpus, a journal list or an imprint is worth
The author relationship: consultation, consent, communication
Editorial and contractual judgement: what a grant actually means

Execution: where we fit

Work-level identity above ISBN, ISSN and DOI
Grant and reversion modelling, per language and territory
Consent and reservation register, machine-readable and per work
Signed corpus manifests per licence
Settlement to authors, translators, estates and co-authors, line by line
Interoperability with ONIX, ISNI and reproduction rights organisations

Who gets what

One layer, four counterparties.

The same record answers a different question depending on where you sit, which is precisely why it works better as shared infrastructure than as anyone's internal system.

Trade publishers

A backlist you can actually license, because you know per title what you hold, what reverted and in which languages. Author-level consent recorded before a deal is signed rather than explained afterwards, and settlement that reaches co-authors, translators and estates per line.

Academic and STM

Journal-level and article-level scope for corpus deals, with DOI and ISNI carried through to allocation. The 2024 deals showed what happens when authors cannot see their position. A register they can inspect is cheaper than the fallout.

News and periodicals

High volume, short shelf life and per-article rights, where retrieval and grounding matter more than training. Consent per use type, per outlet, with reservations that are actually findable by a crawler rather than asserted in a terms page.

RROs and author societies

Work-by-work consent and reservation under a collective AI licence, administration of licensed corpora, and explainable allocation of lump sums. Early relevance at mandate and clearance, rather than arriving afterwards to distribute what is left.

Boundaries

Where the limits sit.

Stating the limits early is what makes the rest credible in a room with lawyers in it.

Licensing

LegacyNext registers and executes within the authority it is given. It does not license titles, negotiate terms, or stand between a publisher and its authors.

Contract reading

Deciding what a 1998 grant covers is legal judgement and stays with your counsel. What the system does is hold that decision consistently once it is made, per title and per language, so it does not have to be made twice.

Detection

Whether a given output derives from a given text is an approximation. Signals from specialist providers are connected to rights data and presented as signals, with the causal question left open.

Identifiers

A verifiable record per work, built on ISBN, ISSN, DOI, ISNI and ONIX, which all keep their function. The scope is one catalogue at a time, which is what makes it deliverable.

Legal transfer

The passport records and evidences. A deed or other required legal basis remains necessary and untouched.

Next step

Start with one imprint.

A first deployment is deliberately small and predictable: one imprint or one journal list, five hundred to five thousand works, editions consolidated to work level, the grant and reversion status established, consent recorded per use type, a single licence or licensee, a signed manifest, one payment, and automatic allocation to authors afterwards.

That proves four things at once: rights data links reliably at work level across editions and translations, grant and consent become machine-readable, a corpus can be delimited objectively in advance, and the money divides explainably per title and per contributor.

A sixty-minute conversation about your current AI exposure and your backlist is enough to tell whether this fits. If there is a basis to continue, we agree the next step together.

Start a Conversation Technical Deep Dive

Know what you licensed.Pay the right authors.