Novel, and not.

The Large Meaning Model is not a token predictor and not a physics simulator. It occupies a third position — and the contrast is what clarifies what it actually is.

Large Language Models are autoregressive sequence predictors trained on observational corpora; they learn what humans are statistically likely to say in surveilled public forums. World Models — the frontier represented by LeCun and others — internally simulate physical reality. Both are valuable. Neither captures what a person wants when they are not being watched.

The sharpest framing: the LMM is a socio-economic resolution model — orthogonal to text-prediction and physics-prediction. It learns to map declared natural-language intent onto actionable classification, expected outcome, and counterparty match, conditioned on the full history of declared-and-resolved signals.

The intellectual honesty of the paper depends on this column being short and the right-hand column being long. Almost every component is established. The novelty is the composition and the data condition .

Declared intent + classification + match + verified outcome, as a single attributable unit accumulated into a corpus. Declared-intent-plus-outcome is structurally different from inferred behaviour.

A corpus whose fidelity depends on the absence of surveillance . This cannot be replicated by an incumbent without dismantling the business that produced their data.

A socio-economic resolution model — neither LLM nor World Model — trained exclusively on consented declared want paired with verified result.

Pricing on verified resolution rather than impressions, with the conversion signal aligning user, counterparty, and model integrity simultaneously.

Tiered LLM agents over a taxonomy is an engineering pattern, not an invention. Standard intent-classification benchmarks (CLINC150, BANKING77, SNIPS) already exist.

A credible paper anticipates its own critics. Three claims were intentionally weakened during review because the stronger version was indefensible:

"Replace programmatic" → coexist. Programmatic will not die. The defensible position is the honest-signal layer the open web needs — the lower-volume, higher-intent complement, not the replacement.

"Near-perfect classification" → progressively self-improving. A taxonomy with ~1,500+ ambiguous categories admits no near-perfect classifier. The honest description is self-improving classification with a human-in-the-loop escape hatch.

"Agent per category" → tiered generalists. 1,500 agents is an orchestration nightmare with a thin long tail. Tiered generalist agents with retrieval get the specialisation benefit without the operational cost.

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