Elena AILL
An experimental artificial language-learning architecture developed around persistent internal state, local neuronal feedback, implicit memory consolidation, and one continuously evolving system.
Elena AILL investigates language acquisition, memory, reasoning, adaptation, and persistent learning through a non-Transformer experimental architecture. Its development record connects each result to a named research phase.
Elena AILL in 10 seconds
Elena AILL in 10 seconds: 849 documented phase records in an evidence-first archive; experimental paths in memory, language, reasoning, and local learning; the 10M benchmark as a sample-fact run; and preserved PASS, PARTIAL, and FAIL results. The latest locally verified runtime remains a partial candidate and does not demonstrate general language capability.
What Elena AILL is
Verified properties and research objectives are deliberately separated. The project’s direction does not become a verified capability until the corresponding phase gate supports it.
Elena AILL is an experimental research project developed through individually documented phases, each with its own objective, evidence, and outcome.
The project is organized around one evolving internal system. Local learning and consolidation are investigated through specific phase tests.
Successful restart and cold-reload tests document persistence for the state and capability examined by those tests.
Research principles
These constraints are stated conservatively. Each applies only where the relevant project report or architecture audit explicitly checks it.
No Transformer attention
Relevant architecture audits declare and test a no-Transformer, no-attention boundary. This describes the documented Elena path, not a claim that every experimental artifact has identical internals.
No retrieval-generated runtime answer
Where the corresponding audit is present, runtime answers are not generated by retrieving a stored response from a RAG or vector-database response path.
No response bank
Relevant language phases test that final answers are not selected from a bank of complete responses. The claim is limited to the paths and scans covered by those reports.
No runtime teacher rules
Teacher-runtime and prompt-lookup exclusions are stated only for phases whose reports explicitly audit them; they are not generalized beyond that evidence.
Knowledge inside the evolving system
The intended direction is for learned knowledge and behavior to remain in persistent internal state. Whether that occurs is evaluated through phase-specific learning, recall, and cold-reload tests.
Repeatable scoped tests
Results are reported against named phase gates, frozen or blind cases where available, and explicit limitations. A passed gate is not expanded into a universal capability claim.
Research areas
These subjects organize the documented experimental work and its benchmark record.
Language
Text-to-meaning, lexicalization, morphology, generation, and the boundaries between structured meaning and natural-language output.
Memory
Persistent state, recall after restart, controlled consolidation, provenance, and failure modes when evidence is incomplete.
Reasoning
Structured planning, ordered propositions, causal tests, multi-fact composition, verifier behavior, and unsupported conclusions.
Speech
The experimental path from internal meaning to generated surface language, including grammatical and semantic coverage checks.
Training
Controlled learning runs, candidate checkpoints, promotion gates, regressions, and the distinction between a trial and an accepted state.
Self-learning
Local adaptation and feedback experiments whose effects must survive their own safety, recall, and anti-forgetting tests.
Architecture
The evolving non-Transformer system, its declared constraints, component boundaries, and audits against prohibited runtime shortcuts.
Benchmarks
Phase-specific measurements, blind sets, ablations, cold reloads, and negative results preserved without presentation upgrades.
Internal technical systems
Highlighted research paths
Current research status
The current English catalog contains 849 canonical phase records. Its frozen historical snapshot contains 847 records and 838 distinct evidence records, preserving the classifications recorded at that boundary.
Phase-scoped validation; not a universal system claim.
Evidence preserved without an inferred terminal pass.
Passing and limited outcomes remain visibly qualified.
Negative gates remain part of the research record.
722 frozen-snapshot routes include one or more recorded figures. 607 preserve an explicit verdict. F896 and F898 are canonical post-snapshot records: searchable in the current catalog but excluded from the frozen 847-record counters.
Latest verified state
The latest locally verified runtime candidate is ELENA_EXTERNAL_RUNTIME_REPAIR_PARTIAL. It uses the F900, F901 and F903–F913 contract, with F902 disabled. The observed language path comprises TTM, structured memory and reasoning, the planner, F711, Decoder V3 and the verifier. It passed frozen internal controls, but the external Italian evaluation remained partial: 0/86 general benchmark answers with 100/100 deferrals. It is not promoted as a general language model or confirmed as a current public service.
Research timeline
End-to-end planner audit
The structured planner passed its controlled contract, while the natural-language pipeline remained insufficient.
Transport and surface follow-up
The follow-up isolated fact-contract transport and lexical surface gaps; its overall gate remained closed.
Scoped predicate realization
The later candidate closed the frozen predicate-realization gap and passed its own limited benchmark and reload checks.
Selected research milestones
This small selection shows why failed and later corrective phases must be read together without rewriting either outcome.
Meaning, Knowledge and Planner Audit
An important insufficient result: the structured planner worked on its controlled structured input, while the natural-language end-to-end path remained too weak for validation.
Read the evidence → F898 · ValidatedExisting F711 Predicate Realization
A later scoped correction: predicate realization reached 28/28 on the frozen supported meanings and retained 20/20 on a separate blind multi-fact set.
Read the evidence →Known limitations
- Performance remains uneven on unsupported or differently phrased natural-language inputs.
- End-to-end behavior depends on the quality of text-to-meaning processing as well as planning and surface realization.
- A phase-specific validation proves only the benchmark and gate that were run.
- Successful benchmarks do not imply human-level language, reasoning, consciousness, or general intelligence.
- Later evidence can narrow, correct, or supersede the interpretation of an earlier phase.
Evidence-first methodology
Public presentation follows the result instead of changing it. Evidence boundaries are part of every final editorial page.
Original status retained
Validated, documented, partial, mixed, failed, and inconclusive outcomes remain distinguishable.
Numbers stay scoped
Benchmark values are reproduced from their evidence packages and are not upgraded for presentation.
Failures remain visible
A negative result is part of the research record and is not rewritten as a successful milestone.
Public boundary enforced
Public pages explain results and limits while withholding source code, private paths, checkpoints, prompts, credentials, and reserved implementation details.
Research library
Focused thematic syntheses complement the phase-by-phase evidence archive.
Experimental architecture
A documented view of Elena AILL’s evolving non-Transformer research architecture, separating current evidence, historical components, and research objectives.
Research lineage
A non-linear history of Elena AILL: foundations, speech, memory, local learning, integration, failures, corrections, and later scoped validation.
Language system
The documented path from internal meaning to surface language, including Clean Speak, SPEAK-NEW, text-to-meaning, planning, lexicalization, and known failures.
Memory research
Persistent state, restart consistency, recall, consolidation, anti-forgetting, provenance, and the limits of what a reload test proves.
Reasoning research
Structured planning, ordered facts, causal and multi-fact tests, verifier behavior, and failures between controlled meanings and natural-language execution.
Training and candidate evaluation
Controlled learning runs, SenzaAdam history, candidate checkpoints, baseline comparisons, promotion gates, and regressions.
Self-learning experiments
Evidence about local adaptation, feedback, autonomous-weight experiments, safety gates, recall, and anti-forgetting without claiming unrestricted autonomy.
Benchmark hub
A scoped index of benchmark families across language, memory, reasoning, speech, holdout, blind, cold-reload, candidate, baseline, oracle, and end-to-end tests.
Research milestones
Historical milestone labels and later phase gates, presented with their documented meaning and without converting project history into a success-only timeline.
Visual evidence archive
An organized historical view of approved public Stage 2–6 figures and benchmark images, reused without duplicating files.
Evidence-first methodology
How phases, verdicts, metrics, source figures, failures, candidates, and public boundaries are represented without upgrading the evidence.
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.
Word Memory
Evidence about persistent lexical state, recall, consolidation, restart checks, and anti-forgetting, with persistence kept distinct from retrieval correctness.
Word / World Document
The historical Word / World Document research label, covering documented ingestion, structural reading, session memory, grounded answers, and runtime boundaries.
Internal learning systems
Verified technical mechanisms for persistent state, local feedback, internal priorities, offline consolidation, association, carryover, and answer evaluation.
GLIA — Historical offline research
GLIA identifies a historical, offline Elena AILL research track concerned with stability and memory-related mechanisms. Its documented results remain part of the project lineage, but its current ordinary-runtime status is not publicly confirmed.
SenzaAdam and Big Brain — Persistent-memory research
SenzaAdam and Big Brain name a documented line of persistent-memory and local-update experiments, including phases that tested operation without Adam and with global backward disabled. They are not presented as a universal current engine.
Clean Speak and SPEAK-NEW — Controlled output research
Clean Speak is a documented experimental path for controlled language output; SPEAK-NEW belongs to its historical lineage. Their phase evidence is preserved without claiming that either label identifies the complete current response runtime.
10M persistent-memory benchmark run
A persistent-memory streaming run configured with 10,000,000 sample facts. Its 10,833,323 learned-fact count is cumulative because the run resumed prior state; both values count facts, not model parameters.
Who is Elena? Why Elena AILL exists
The human question behind Elena AILL: whether an artificial system can learn over time, preserve useful experience, and evolve without turning a research ambition into an unsupported claim.
Public research boundary
What the public archive includes, what it excludes, and why benchmark evidence does not imply a live product, unrestricted capability, or deployment readiness.
Explore the phase archive
Browse the current English evidence archive, including validated, documented, partial, mixed, and failed phase records.