Elena AILL Word Model — Sequential Generation, Word Memory and Runtime Boundaries
Elena’s Word Model research covers learned token-transition and prompt-conditioned generation state, used alongside a separate persistent Word Memory. In the ordinary runtime, a sequence-only decoder and Word Memory can influence responses; richer lemma, morphology, predicate, and argument representations belong to separate text-to-meaning and lexicalization components in the extended candidate stack. The planner remains separate, and persistence and behavioral claims are limited to documented test scopes.
Why the historical label remains
Word Model is a historical and architectural project label for the path that learned and reused word or token transition state and prompt-conditioned generation behavior. It now describes an evolving family of components, not one complete active word object.
Current ordinary runtime
The ordinary path combines a sequential no-attention decoder, a separate persistent Word Memory, and a bridge that evaluates candidate outputs. The decoder can generate from learned sequential state, Word Memory can recall indexed phrase material, and the bridge can select what continues toward the visible response.
Representation actually used
The ordinary state contains normalized tokens or strings, token-transition state, short sequential patterns, prompt-conditioned associations, and indexed phrase memory. It does not combine lemma, morphology, syntax, semantic relations, and argument structure in one representation; those richer properties belong to separate candidate components.
Component boundaries
The Word Model path covers token transitions, conditioned generation, and decoder behavior. Word Memory separately persists and recalls indexed phrase state. Candidate text-to-meaning builds structured meaning; candidate lexicalization handles morphology, finite forms, predicates, and argument realization. The planner orders structured material separately, while the bridge passes a selected candidate into the downstream response path.
Persistence and runtime updates
Decoder and Word Memory state can be loaded from checkpoints, and Word Memory preserves indexed phrase state and metadata. Observations made during one runtime process are not automatically saved or promoted, so some updates remain volatile. Cold-reload findings apply only to the exact phase and configuration tested.
Response influence
The sequential decoder can directly change generated output; Word Memory can supply recalled candidate material; and the bridge can select between candidates. Lexicalization has causal evidence in candidate gates. The historical legacy Word Model does not affect the current ordinary server path.
Current verified state
The ordinary runtime can load sequential lexical and generative state, generate from that state, recall from a separate Word Memory, and pass candidate output through a selection bridge. This does not prove a unified rich word representation, ordinary-runtime lemma or morphology coverage, unrestricted structured understanding, open-domain language competence, autonomous learning during this audit, or universal correctness.
Scoped evidence
A non-mutating runtime control produced 17 generated tokens with a passing output gate and three Word Memory recalls; the protected state remained unchanged. F898 improved its lexicalization candidate from 14/28 to 28/28, passed a 20/20 blind multi-fact evaluation, and passed its scoped cold reload. Decoder evidence includes F685 at 5/28 baseline versus 24/28 candidate, without promotion, and F875 at 14/14 required plus 48/48 blind, also without promotion. These results are not combined into a global score.
Current and historical components
The legacy Word Model is historical. The sequential decoder, Word Memory, and their bridge are current in the ordinary path. Text-to-meaning and richer lexicalization belong to the extended candidate stack. The constrained decoder is retired. Later evidence recorded for F900 and F911 informs candidate lexicalization and text-to-meaning respectively, but neither has a canonical phase route in the current site catalog.
Related language systems
Word Memory is a separate phrase-state persistence and recall layer; it does not contain the whole Word Model. The broader Language System contains the distinct text-to-meaning and lexicalization work needed for lemma, morphology, predicate realization, and argument structures.
Current and historical component status
| Component area | Status | Public interpretation |
|---|---|---|
| Historical legacy Word Model | Historical | Earlier learned word-transition path; inactive in the ordinary server path. |
| Sequential decoder | Current ordinary runtime | Generates from learned sequential and prompt-conditioned state. |
| Word Memory | Current ordinary runtime | Separate persistent indexed phrase state and recall. |
| Selection bridge | Current ordinary runtime | Evaluates decoder and memory candidates for the downstream answer path. |
| Text-to-meaning | Extended candidate stack | Separate structured-meaning component. |
| Lexicalization | Extended candidate stack | Separate morphology, finite-form, predicate and argument realization. |
| Constrained decoder | Retired | Historical path that no longer constructs final answers. |