Experimental architecture · Evidence and boundaries

Non-Transformer AI architecture: components, runtime and limits of Elena AILL

Elena AILL is an experimental project with symbolic components, persistent state, n-gram decoders and PyTorch neural prototypes. In the audited generative path there are no Transformer modules or Q/K/V self-attention modules. This does not mean that the repository lacks neural networks, softmax, or content addressing mechanisms.

What is an artificial intelligence architecture

An architecture describes how input, representations, memory, learning, inference, control, and output production interact. The presence of a file or library does not prove that that component runs, changes state, affects output, or is promoted into the canonical runtime.

For Elena we must distinguish eight levels: present in the repository; loadable or used by utilities; offline-only; used by an operational route; causally proven on the output; enabled by manifest; promoted to canonical runtime; unpromoted hash-consistent candidate. None of these levels automatically implies the next.

Present technologies and role in the audited process

Technology	Present in the project	Role in the audited generative path	Notes
Transformers	Not detected in the audited path	No role demonstrated	It is not a universal statement about all historical material.
Transformer self-attention	Not detected in the audited path	No role demonstrated	No Q/K/V, multi-head attention, or scaled dot-product token-token observed.
PyTorch	Yes	Neural core excluded from canonical eager closure	Present in the repository and usable offline; it is not the audited canonical generator.
Neural networks	Yes	Offline/non-canonical prototypes	Embedding, Linear, MLP and LayerNorm are actually implemented.
Local embeddings	Yes	No canonical effects demonstrated by offline core	`nn.Embedding` it is present in the PyTorch prototype and in GLIA.
Softmax	Yes	Present in the offline core and in experimental components	The presence of softmax does not make the system a Transformer.
Content addressing	Yes	Offline/non-canonical in audited core	Similarity and softmax weigh prototypes or slots; it's attention-like, not Transformer self-attention.
GLIA	Yes	No role in audited canonical closure	HISTORICAL / OFFLINE_ONLY.
F711	Yes	Lexicalization in the verified candidate path	Create predicates and roles; it is not general reasoning.
Decoder V3	Yes	Bigram/trigram generation with beam in the verified path	It does not use Transformer logits in the observed path.
Adam / AdamW	Historical material excluded; not in the official audited route	No role in the official path verified	Adam is an optimizer, not synonymous with backpropagation.
Global backprop	Separate legacy material	Not used in the official audited route	The claim remains limited to the verified route.

Transformers and self-attention: the technical border

A Transformer's self-attention constructs queries, keys and values, then uses scaled products to combine information between tokens, often across multiple heads. In Elena's audited generative path no MultiheadAttention, Q/K/V, scaled dot-product self-attention or Transformer modules.

The offline neural core instead contains content addressing based on similarity and softmax, functionally describable as attention-like. It is not Transformer self-attention, it does not use Q/K/V, it does not use multi-head attention, and it does not use scaled dot-product token-token self-attention. This distinction appears once to avoid both an absolute statement about the absence of attention and the opposite mistake of calling every weighted selection Transformer.

The generative path actually verified

Within the documented perimeter, the verified candidate path can be summarized as:

text → text-to-meaning → memory and structured reasoning → F711 lexicalizer → V3 bigram/trigram decoder with beam → verifier → output.

The offline PyTorch neural prototype is not inserted into this chain. The path should not be presented as the current live canonical runtime: the v19 candidate is hash-consistent, but not promoted.

Component	Role	Status verified	Limit
Router / guard	Branch selection and fail-closed controls.	Candidate operational path	Activation does not demonstrate semantic correctness.
Text-to-meaning	It builds frames, predicates and roles.	Causal but partial	Incomplete open-domain coverage.
Memory / world state	It lays out facts and a structured state.	Persistence documented in limited tests	Uneven recovery for each question.
Reasoner/planner	Operates on available evidence and structures.	Candidates and controlled tests	Natural input can fail before the planner.
F711 / F900	Lexicalization of predicates and roles.	Subpath verified	PASS component does not imply PASS pipeline.
Decoder V3	Progressive surface construction.	Reason in the audited candidate path	It does not compensate for missing meaning or knowledge.
Verifier	Accept, block or produce defer.	Fail-closed verified	Stability does not equal understanding.

The PyTorch neural prototype is present, but offline

The files core/aill_model_no_attention.py and core/aill_no_attention_components.py implement PyTorch, nn.Embedding, nn.Linear, MLP, LayerNorm, tensor memory, Hebbian/plastic updates, similarity, content addressing and softmax. It is incorrect to say that Elena does not contain neural networks, does not use softmax, or does not include PyTorch.

core/__init__.py However, it defines the current boundary: PyTorch experiments are OFFLINE_ONLY / NON_CANONICAL and must be imported explicitly. The neural prototype is present in the repository but does not constitute the canonical generator of the responses in the audited path.

Ablation of eager loading

Excluding the heavy PyTorch core from canonical loading maintained equivalent outputs in 111/111 shifts e 500/500 cases; TTM and trace F711 also remained equivalent in their respective replays. Ablation showed no causal influence of the neural prototype on the canonical responses in the audited replays. It does not prove that the core is useless, that it will not be used in the future, or that PyTorch is not used in the project.

GLIA

GLIA contains previous-hidden, slot routing, embedding, linear projections and experimental content addressing. The module exists in the repository, but has been excluded from the canonical eager closure along with the heavy core: the correct public classification is HISTORICAL / OFFLINE_ONLY, not current causal component.

WithoutAdam / Big Brain

PersistentBrainState it is imported from the server and participates in the persistent state contract. However, Hebbian and broader consolidation helpers have no demonstrated independent causality on the output. The correct formulation is: WithoutAdam / Big Brain — persistent state contract + non-causal helpers in the audited path. It is not presented as a “generative brain”.

Manifest, enabled phases and promotion

The stated canonical manifest v2 lists F900 and F901 as enabled, disables F902, and enables F903–F913. ENABLED means configured by the manifest: it does not mean automatically causally verified, hash-consistent with the current checkout, or active in a live runtime.

Object	State	What it proves	What it doesn't prove
Canon Manifest v2	F900, F901, F903–F913 enabled; F902 disabled	Declared contract	Current server hashes and TTM not aligned.
Candidate v19	Hash-consistent	Declared sources and checkpoints correspond to checkout	He is not promoted to canon.
F896	Phase and telemetry not promoted	The structured planner can be instantiated in the full learned stack	It doesn't promote F896 and doesn't resolve end-to-end over natural text.
F900–F913	Mixed qualification according to manifest	Feature contract declared	Enabled is not the same as causally verified or live hash-consistent.

For text-to-meaning the updated wording is: open-domain coverage still incomplete. The most recent candidate is bound to the source and checkpoint hashes, but has not been promoted and does not demonstrate general linguistic understanding.

Documented architectural evidence

Source	What does it check	Result	Confine
Canon Manifest v2	Authorized features and hashes	F900/F901/F903–F913 enabled; F902 disabled; 2 source hashes not aligned	Declared contract, not live certification.
Candidate v19	Source hashes and checkpoints	14/14 consistent checked elements	Candidate not promoted.
Core eager-import ablation	Heavy PyTorch core influence	111/111 and 500/500 equivalent outputs	Audited replays, not universal uselessness of the core.
F235	No-attention Word Model	Validated · PASS	Circumscribed phase.
F315	Word Model Memory Bridge	Validated · PASS 10/10	Does not test the full runtime.
F896	Meaning, knowledge and planner	Oracle 32/32 planner; TTM 3/32; end-to-end 2/28	Phase and telemetry not promoted.
F898	F711 construction	Circumscribed gate PASS	Structured input and candidates.

Negative results and candidate v19 status

The frozen external Italian benchmark of external-runtime-repair-candidate-v19 record 0 runtime errors, 1 unique answer, 0/86 correct objective answers and 100/100 abstentions. The result demonstrates fail-closed stability in the sample, not general understanding, open-domain knowledge, or fixed language ability.

The main bottleneck remains text-to-meaning on complex external inputs. Making the verifier more permissive would transform correct deferrals into unfounded answers; it would not correct missing meaning, knowledge, or task handling.

F896: structured planner and natural pipeline

F896 documents 32/32 with oracle queries and frozen evidence, but only 3/32 on complete text-to-meaning e 2/28 end-to-end on applicable cases. The structured planner can be instantiated when the full learned stack is active; this does not promote the phase or its telemetry and does not demonstrate end-to-end operation on natural text.

Historical snapshot 111 rounds

The snapshot of the canonical manifest v2 records, for the historical sample of 111 turns, fallback 77/111 and one meaning completeness of 0.22. This is data from the baseline declared in the manifest, not the current state of the v19 candidate and not a live measurement.

Adam, global backprop and local updates

In the official path audited Adam, AdamW e torch.optim they are not used; global backpropagation is not used in the same perimeter. Adam is an optimizer and is not synonymous with backpropagation. In the repository there are also offline neural prototypes and legacy material: the claims should not be extended to the entire history of the project.

Local updates, Hebbian/plastic and consolidation are implemented in specific components. Their presence does not demonstrate end-to-end persistent learning, influence on every response or promotion in the canonical runtime.

What this architecture does not demonstrate

No overall superiority or equivalence to the Transformers.
No absolute absence of neural networks, PyTorch, softmax or attention-like mechanisms.
No current perfectly hash-aligned live canonical runtime certified by this page.
No promotion of candidate v19.
No general linguistic understanding or universal reasoning.
No equivalence between presence, loading, enabling, causality and promotion.

MLP, embedding, softmax, associative memory, Hebbian/plastic updates, n-gram, and beam search are known primitives. Elena's documentable specificity concerns design combination, component separation, constraints, persistence, gates, verification, and runtime boundaries. No scientifically unprecedented architecture is claimed.