The Intent Exchange

The system never interrogates the person. It helps them articulate a want, then resolves it. Structured data emerges as a byproduct of a genuinely useful exchange — not as the goal of a form.

The user states a want in natural language — text or voice. No tracking, no profile building, no visible form fields. The signal is consented and time-bound at the moment of creation. The person can see, correct, expand, or delete anything the system extracts.

Rather than one agent per taxonomy category — an orchestration nightmare with a thin long tail — a small number of tiered generalist agents (L1 top-level taxonomy → L2 specialist → L3 localisation) retrieve over category-specific examples. A human-in-the-loop escape hatch resolves ambiguous edge cases. Classification is progressively self-improving , not "near-perfect."

Brands, employers, public institutions, and service providers run matching agents that consume classified signals and return an offer, job, service, or answer. They never see the author. Multiple counterparties compete on relevance for the same signal.

The user approves what surfaces. Value flows only on outcome — pay-per-click, then pay-per-qualified-lead when the user signals "I want to buy now," then pay-on-conversion settlement. No CPM. The outcome is recorded and fed back as training signal.

The atomic unit of the corpus is the intent pair. It records what a person said they wanted, how the network interpreted it, what was offered, and what the person actually did. The pair is to the Large Meaning Model what the block is to a blockchain: a small, self-contained, verifiable unit whose accumulation becomes the asset.

The user's want, in natural language, consented and time-bound.

The schema node the agents resolved the signal to.

The counterparty offer returned and surfaced to the user.

Click / lead / conversion / verified — the real-world result.

Structurally, the outcome gradient on the pair is a reinforcement-learning-from-human-feedback reward signal : the network learns not just what people say, but whether the resolution of what they said was right . The corpus is, in industry terms, Zero-Party Data — neither inferred (third-party) nor merely transactional (first-party), but explicitly declared and cryptographically attributable.

The model improves along three axes simultaneously — and compounding does not require an increasing compute budget. It requires increasing network participation.

Every signal adds a record to the corpus.

Every outcome corrects or reinforces a prior classification.

Every new domain — employment, civic, health, education — adds a dimension of meaning.

Signals declared Intent pairs LMM training Better matching User trust More signals more domains LMM inference API — revenue A surveillance-based competitor cannot copy this corpus without first building a surveillance-free environment — which is structurally incompatible with the business model that produced their existing data. The moat is the absence of surveillance, not the presence of scale.

Resolution fees fund the network in the near term: lead fees and conversion fees flow to the operator each time a signal resolves into a real outcome. Inference is the second primitive — brands, platforms, and third-party applications query the trained LMM 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.

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