The Model is the commons.

The LMM is trained on protocol-hosted infrastructure, on anonymised intent pairs, under a discipline that makes individual reconstruction provably impossible.

Training operates on anonymised intent pairs with differentially private methods applied at the batch level : individual signals provably cannot be reconstructed from the trained weights. Training is staged — a base model learns general meaning resolution, and domain-specific fine-tunes specialise on advertising, employment, civic, and health sub-corpora.

Inference is the network's second revenue primitive. Counterparties and third-party applications query the model through a metered API to perform meaning-resolution tasks no open-market model can match. The query surface exposes resolved classifications and statistical outcome distributions; it never exposes the corpus.

The paper commits to principles, not implementations — the same discipline applied throughout. Implementations may change with the cryptographic state of the art; these three statements do not.

A signal carries no durable identifier. There is no profile being assembled behind it. The author of a signal cannot be recovered from the signal, its classification, or its resolution.

Training applies differential privacy at batch level. The trained model cannot be inverted to surface any individual's declared want. k-anonymity thresholds gate any query that could narrow to a person.

A matching agent receives a classified signal, not an identity. The single largest attack surface — a counterparty re-identifying users statistically across sessions — is closed by query thresholds and the absence of a persistent conversion identifier exposed to counterparties.

Contribution is not quantity. A fair stake mechanism has to answer: who actually moved the model? These cases show why a naive per-task count fails:

The fair measure is marginal information contribution to the model . This is calculable in principle with influence-function methods and data-Shapley values — both real research programmes, both computationally expensive and still maturing. The paper commits to the principle and names the direction (influence functions, Shapley-style attribution, TMC-Shapley, Data-OOB) without locking in a specific algorithm.

Three structures were weighed. The choice trades legal exposure against defensibility of the alignment claim.

Contributors hold a contractual right to a share of model revenue, calculated by training contribution. Lowest legal exposure; behaves like ownership.

The model is held by a non-profit Foundation as protocol steward; operators implement under licence; contributors receive deferred profit-sharing distributions. IETF / W3C / Linux Foundation / Ethereum precedent.

A separate entity owns the model; contributors hold registered equity. Most defensible, most expensive, heaviest multi-jurisdiction securities burden.

Single-entity ownership is positioned as incompatible with the protocol's purpose : a model whose value depends on universal honest participation cannot credibly be one company's asset. The model is no party's property; it is the network's commons.

Four principles are constitutional — they bind future Foundation governance and cannot be amended by an operator for commercial convenience.

The fourth commitment is the load-bearing economic mechanic. Early contributors, who add information when the corpus is sparse and every signal moves the model, accrue more stake per unit of contribution than later contributors adding to a mature, saturated model. The incentive curve declines deliberately — it rewards the conviction of the early, not the volume of the late.

Current AI assumes the user already knows what they want and can state it as a prompt. This protocol assumes the opposite: the user is mid-calculation, and what they type is the visible residue of a want they have not fully resolved. The LMM is designed to map the variables that produced the declaration — not to optimise the declaration itself.

These working drafts develop the argument. The peer-citable paper is deposited on Zenodo under DOI 10.5281/zenodo.20616391.