DAWM State Contract: Observation to Event to Fluent

What this note records

This note states a foundational definition of "state" for a next-STATE world model and packages it with a concrete three-layer realization. Downstream work in the program assumes the definition given here. The pipeline is:

Observation  ->  Event  ->  Fluent / State
(raw modality)   (outcome-free claim)   (reduced (entity,slot)->value)

The thesis framing: a world model should be a next-STATE predictor, not a next-token predictor. Tokens, images, audio, and documents are OBSERVATIONS of state, not the state itself. The scarce, hard 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 are just candidate inference engines over it.

The contract (one sentence)

State is the minimal sufficient representation that the system maintains across observations and queries such that future predictions, future queries, and rendering-invariant inference are well-defined operations on that representation; it is distinct from the tokens / images / sounds that render it for input or output, and distinct from any single learned embedding of it.

What qualifies as state (six conditions)

A representation R qualifies as state only if it satisfies ALL of:

1. Predictive sufficiency -- R at time t is a sufficient statistic for prediction at t+k, without reference to raw observation history prior to t.
2. Rendering invariance -- R derived from any of multiple distinct renderings of the same underlying situation is the same (within tolerance) across renderings. This is the core empirical property the program must show.
3. Persistence -- R persists across observation/query boundaries; it is not reconstructed from scratch on each forward pass. Persistence is mechanical (held in memory) and semantic (maintains commitment to the situation).
4. Causal admissibility -- R admits a transition function f: given R at t and an action/observation A, R at t+1 is well-defined and respects causal structure (irrelevant observations leave the relevant part of R invariant).
5. Compositional structure -- R admits well-defined combination with other states (merging beliefs, combining temporal slices), not re-encoding.
6. Inspectability -- R admits read access for query: a query answer is derivable from R's structure, not from re-running the input pipeline.

What does NOT qualify (the anti-drift list)

Explicitly NOT state, despite adjacent-literature usage: tokens (rendering surface), token embeddings (learned encoding of rendering), hidden forward-pass activations (transient, non-persistent), KV cache (efficiency artifact, context-window-scoped, no rendering invariance), context-window contents, a semantic label (type annotation, not state), a single belief proposition (one element, not the whole), a learned latent vector lacking persistence + rendering invariance, and a clustering/categorization (statistical descriptor without causal admissibility).

"Semantic" is NOT state either -- it is a derived VIEW of state (a view function discarding rendering, retaining meaning) and serves as the MEASUREMENT INSTRUMENT for rendering invariance, not the property measured.

The three-layer realization

The thesis identified defining State as the scarce thing. A concrete, outcome-free definition was adopted as the shared contract. At a high level the three layers separate raw signal, claim, and accumulated state:

• Observation = raw image/text/audio/doc; what the perception encoders extract.
• Event = a typed, OUTCOME-FREE claim about a single (entity, slot) attribution: who claims a value for which slot, with what source provenance and a bounded notion of support. The decisive design property is that an event carries the claim and its provenance only -- never an interpretation, verdict, or outcome. Events are ordered by logical turn so that accumulation is a deterministic replay rather than a wall-clock race. This is the input event to the transition function.
• Fluent / State = a map from (entity, slot) to the current accumulated value together with an epistemic status and a record of supporting and withdrawn provenance. This is S. owns(Tom, House) is a fluent (state), NOT an event.
• Transition T = a deterministic reduction of the event stream into State, followed by a readout policy over the state that returns an epistemic status plus an answer. The key invariant: any outcome or verdict is COMPUTED by the transition/readout, NEVER baked into the Event.

(The exact field set, value enumerations, support-weight scale, and wire encoding are an implementation detail and are deliberately omitted here; the load-bearing academic content is the outcome-free separation, not the serialization.)

The iron rule

An event carries only that some source claims a value for an (entity, slot), with a bounded support -- never "what it means / whether it is right." A verdict-bearing event such as "Good Offer Submitted" is ILLEGAL (verdict baked into the event); "Offer Submitted" plus a separately computed outcome from the readout is legal. This is the schema-layer form of blocking the answer-key leak -- the same failure as the measurement traps, named at the data-contract level. It is the operational form of "ontology is output, not input."

The three falsifiable gates (current honest status)

The thesis is EARNED only when all three clear on data we did not author.

• Gate A -- Carrier-independence (state orthogonal to observation): the same state derived from different modalities should yield the same internal representation. Status: PARTIALLY EARNED, micro-world only. The strongest micro-world evidence to date is a value-sensitive convergence result with a positive multi-seed margin (clean across seeds), but ONLY on hand-authored micro-worlds. The real-data test is a binding-consistency emergence probe.
• Gate B -- State emergence (the graveyard): the Entity/Relation/Event ontology must EMERGE as latent structure from unlabeled observation, without collapse, and generalize to held-out worlds. Falsified if structure does not exceed both shuffled chance AND a raw-backbone control, or if it collapses. Status: UNEARNED. The micro-world rehearsal is the strongest candidate so far but has not been upgraded and stays a boundary case. The real-data exam currently FAILS the backbone control cleanly: the trustworthy binder under-performs naive k-means clustering of frozen features. Not a gold artifact, not at scale.
• Gate C -- State advantage (the decisive falsifier): explicit state must give a MEASURABLE advantage on the query class token models fail (compositional binding / counterfactual / long-horizon state) AND that advantage must GROW with scale. Status: UNEARNED, untested. Concretized: the baseline a learned transition must beat is a conservative rule-based reducer (the floor), specifically at its KNOWN BLIND SPOT -- frameless / semantic same-slot conflict, where it conservatively abstains. Falsified if a tuned rule matches the learned transition there.

Where DAWM and the contract meet

DAWM approaches the Fluent layer from perception UP: the first emergence target is visual/spatial object-slots (objects in images) -- one rung UPSTREAM of the symbolic (entity, slot) Fluent. The seam clicks at object-slot -> symbolic-Fluent grounding. The slot-binding confidence plays the role of the bounded support notion; the anti-collapse property (keeping distinct fluents from collapsing) is the same failure one layer down. Full typed (entity, slot) state is the grounding rung after the perceptual stage.

Binding rules contributed

• Any new representation must declare whether it IS state, IS rendering, IS view, or IS mechanism. "Sort of state" without verification is a contract violation.
• Conflating rendering with state (token embeddings as state, context windows as state) is a violation.
• Anti-tokenizer-degradation: representation work that is essentially "a better tokenizer" is a violation unless it also demonstrates rendering invariance -- the load-bearing test.
• Anti-pretraining-escape: "throw huge data + scaling and claim semantics emerged" is a violation. The thesis is an architecture hypothesis, not a scaling law.
• Compute discipline: deliberately limit compute during architecture validation; brute-force memorization should be ruled out by design, not post-hoc.

The line (load-bearing)

Clarity is not progress. This contract sharpens the target; the target (state emergence and state advantage on real data) is unhit. Real data is the arena, not the win -- the win is emergence plus a scaling advantage. The deepest novelty, if DAWM succeeds, will likely NOT be a new network -- it will be having defined what State is, then predicting the next one. The scarce thing is no longer undefined; what remains UNEARNED is making it EMERGE from real data and WIN against the rule baseline.