DAWM: State over Token — a research charter with falsifiers

The thesis (one paragraph)

DAWM is a next-STATE predictor, not a next-token predictor. Tokens, images, audio and documents are treated as observations of state — not the state itself. The hard, scarce part is not a new network: it is defining what world state is (its minimal trainable representation) and having that representation emerge from observation. Once state is defined and emerges, Transformer / SSM / GNN / memory networks become candidate inference engines over it.

Objective: Observation → State → next State, not Text → next Text. Internal objects of the state: Entity · Relation · State · Event · Time · Causality.

Two questions kept apart

• (A) Replace the Transformer as a function approximator? No, and it does not need to. Competing on f(tokens)→token is not the bet.
• (B) Replace the world-representation layer the Transformer implicitly imposes (world ≈ token)? This is the bet — the least-overclaimed version of the position.

The formal spine

A latent state S, a transition S_{t+1} = T(S_t, e_{t+1}) driven by outcome-free events e, and a readout O_t = π(S_t) whose observations feed back in. A belief-style schema gives a concrete instantiation: Observation → Event (an outcome-free claim) → Fluent/State (a reduced (entity, slot) → value mapping), with reduce(events) → state playing the role of the transition T. The learned-T formulation keeps the same shape but replaces the reducer with a learned one; the deterministic rule-based reducer is the baseline that any learned transition must beat.

Three falsifiable gates (the bar)

1. Carrier-independence — the same state recovered from different modalities converges to the same representation. Status: partially earned on hand-authored micro-worlds.
2. State emergence — the Entity/Relation/Event ontology emerges as latent structure from unlabeled observation without collapse, and generalizes. Status: unearned on real data; this is the active research frontier.
3. State advantage — explicit state yields a measurable, scale-growing advantage on the query classes where token models fail (compositional binding, counterfactual reasoning, long-horizon consistency). Status: untested.

The line (load-bearing)

Clarity is not progress. Real data is the arena, not the win — the win is emergence plus a scaling advantage. The deepest novelty, if DAWM succeeds, will not be a new network: it will be having defined what State is, then predicting the next one. The clean symbolic ontology is the output hypothesis, never the input — the model must grow it; if it is ever handed the ontology at scale, the result is a toy, however large.