Internal mechanisms · Evidence and causality

Learning mechanisms in artificial intelligence: internal systems, evidence and limits of Elena AILL

A system that changes its state is not necessarily demonstrating learning. An update can be limited to saving a piece of data, recalculating a priority, consolidating a candidate or evaluating an answer. To talk about learning within the scope of a test you need at least, when available, a baseline, an update, a measurable change, a useful metric, a check on previous knowledge and a scope that distinguishes the candidate and ordinary runtime.

Elena AILL is a documented experimental case here. The page does not present a universal theory: separate what is implemented from what is called, changes state, persists, affects output, or overcomes an ablation. The results remain limited to the relevant protocols, including negative ones.

How an AI system can learn

In general terms, a system can be described as learning when it uses data, experience, or feedback to produce a measurable change in behavior or performance. The criterion is not the presence of memory alone: it is necessary to compare a previous state with the subsequent one, measure the effect on relevant cases and check that the new result does not hide important regressions.

There are different possible procedures. A model can update parameters, a memory can acquire facts, an agent can modify a policy, a symbolic system can consolidate rules, or a candidate path can only be accepted after gates and tests. No single implementation is the universal definition.

Updating a state does not necessarily mean learning

The following classification prevents any internal mutation from being called “learning”. The same experiment can cross multiple classes: for example, a memory write can become part of a learning test only when the new content is verified together with the retention of the previous one.

Class	Operational meaning	What does it take to support it	What it doesn't prove
LEARNING	An update produces a measured improvement in the defined task.	Baseline, delta, useful metrics, regressions and scope.	General ability or unlimited learning.
UPDATE ONLY	Changes status or weight without sufficient evidence of useful improvement.	Reproducible state delta.	End-to-end learning.
MEMORY WRITE	Records recallable information or serialized state.	Writing, reading and, if declared, testing after reload.	Adaptation or generalization.
CONSOLIDATION	Reorganize or strengthen existing state.	Candidate pre/post and gate comparison.	Continuous autonomous activity.
CANDIDATE	Version evaluated but not necessarily promoted.	Identity, gate and promotion decision.	Use in ordinary runtime.
EVALUATION	Accept, review, defer, or block a result.	Causal gate decision.	Acquisition of new knowledge.
STATE UPDATE	Update context or operational variables.	Observed delta and known duration.	Persistence or learning.
UNKNOWN	Presence or call without sufficient proof of useful effect.	New probes, controls or ablations.	Any positive ability.

The internal mechanisms studied in Elena

The matrix distinguishes seven public modules and three connected mechanisms. The statuses are not equivalent: current, candidate, default-off, offline and structural indicate different perimeters.

Mechanism	Function	Status	Classification	Causality	Persistence	Limit
Persistent internal state	Record traces and candidate payloads.	CURRENT PARTIAL / CANDIDATE	MEMORY WRITE · STATE UPDATE	Limited delta and reload.	Selected candidate.	Persistence does not prove correctness.
Local feedback	Apply local signals to the candidate state.	FULL-CANDIDATE · DEFAULT-OFF	UPDATE ONLY	State delta yes; independent output effect no.	Selected status.	It is not continuously active.
Internal priority	Sort actions using relevance, salience and gate.	CANDIDATE / CONDIZIONALE	STATE UPDATE	Structural integration.	Recalculated.	It does not produce words directly.
Offline consolidation	Rearrange candidate status out of live path.	OFFLINE-ONLY · EXPLICIT	CONSOLIDATION	Effect on candidate payload.	Only after save/reload approved.	It doesn't happen automatically during the chat.
Conceptual association	Builds connections between recorded concepts.	CANDIDATES · DEFAULT-OFF	UPDATE ONLY / UNKNOWN	Connected, without uniform live selection.	Edge candidates.	Influence on non-uniform output.
Analogy and heuristic inference	Proposes candidate evidence from associations.	CANDIDATE / STRUTTURALE	UNKNOWN	Live selection not demonstrated.	Independent unproven.	It is not the same as human intuition.
Self-assessment	Accept, review, defer or block output.	GUARD CURRENT · LOOP CANDIDATE	EVALUATION	Direct effect on gates.	Recalculated for response.	It does not acquire knowledge on its own.
Conversational carryover	Transfers compatible context between turns.	CURRENT	STATE UPDATE	Limited direct effect.	Volatile.	It's not long-term memory.
Integrated cognitive stack	Orchestrate status, feedback, priority, association and evaluation.	FULL-CANDIDATE	CANDIDATE	Indirect and gated in some shifts.	Candidate state.	It does not govern the pipeline smoothly.
Checkpoint evolution scorecard	Evaluate candidates before saving.	OFFLINE-ONLY	CANDIDATES · EVALUATION	Gate decision.	Candidate only.	It does not imply automatic promotion.

How to check whether a mechanism really has an effect

The presence of code is only the first level. Each step of the ladder requires additional evidence and cannot be inferred from the previous one.

IMPLEMENTEDThe component exists.

CALLEDThe path calls for it.

CHANGES STATEThe call produces a delta.

PERSISTSThe delta survives within the declared perimeter.

AFFECTS OUTPUTThe change causally modifies the outcome.

ABLATION VERIFIEDRemoving it degrades the result compared to the control.

implemented != called != state-changing != persistent != output-causal != ablation-verified. In the Elena audit all seven modules were implemented; call, persistence, and influence instead depend on the ordinary runtime, candidate mode, or offline path. Carryover and response guards have a direct path; the stack is indirect and conditional; association and analog do not show a uniform live effect.

Local updates and plasticity

A micro-update modifies a limited part of the state or weights using a local signal. Local feedback, Hebbian/plastic updates and candidate micro-updates appear in Elena records. These terms describe the mechanism, they do not guarantee that the final answer improves.

F69 records Hebbian consolidation with old recall 9, new sample 100 and 325 synapses in its test; the status is Validated without registered terminal verdict. F686 records reload, anti-forgetting 1 and improvements on noisy/document recall in the candidate. They are limited evidence: they do not authorize us to generalize historical or candidate components to the current runtime.

Confine. A local delta can be real and reproducible but remain UPDATE ONLY if it is not linked to a useful metric, prior checks, and an effect in the evaluated path.

Internal feedback and consolidation

Local feedback collects a signal and updates a candidate state. Consolidation instead decides how to maintain, strengthen or reorganize information already available. The two steps can cooperate, but they are not synonymous.

In the F686 perimeter the route is candidate and default-off compared to the ordinary HTTP runtime. State variation and cold reload are documented, while an independent and uniform influence on the final text is not demonstrated. Current response guards belong to evaluation, they do not automatically transform feedback into learning.

Online and offline consolidation

An online update occurs during a controlled operational flow; an offline route operates outside of the live response and requires explicit invocation. The technical/historical name dream indicates here a consolidation scheduler: it does not describe psychological dreams and does not demonstrate a continuous autonomous process during the conversation.

The offline path can reconstruct associations, compute a scorecard, and produce a savable candidate or a failed gate. Only an approved and reloaded candidate retains the update; frozen and stable are not automatically replaced.

Candidates, gates and promotion

candidate tested != candidate promoted != ordinary runtime. A candidate may exceed local metrics and remain unpromoted because scope is insufficient, a check fails, or end-to-end integration is not demonstrated.

F610 records direct recall, paraphrase recall, and anti-forgetting of 1 on a candidate test, but `candidate promoted=false`. F689 retains a candidate pass in Partial status, still without promotion. The checkpoint evolution scorecard is an offline gate: it decides whether a candidate state is savable, it does not guarantee its activation in the ordinary path.

Documented evidence

The selection covers status, feedback, consolidation, carryover, candidate stack, evaluation, priority, serialization, association and promotion. “Validated” remains an editorial status and is not automatically converted to PASS.

Phase	What head	Status / verdict	What it supports	What it doesn't support
F68	Scoped anti-forgetting test.	Validated · no terminal verdict	Old recall 9; new sample 100; score 1.	General absence of forgetting.
F69	Hebbian consolidation.	Validated · no terminal verdict	Retention and 325 synapses in the test.	General learning.
F233	Cognitive state core.	Validated · PASS	Goal/action/details 6/6.	Universal cognitive architecture.
F610	Candidate lifelong ladder.	Validated · not promoted	Recall and anti-forgetting in the set.	Runtime promoted.
F668	Carryover/reasoning mouth.	Validated · PASS	Circumscribed continuity.	Long persistent memory.
F686	Feedback and candidate consolidation.	Validated · KEEP	Delta, reload and retention tests.	Continuous ordinary activity.
F688	Integrated internal loop.	Validated · no terminal verdict	Structural use of priority and consolidation.	Full causality of modules.
F689-B	Anti-shortcut follow-up.	Validated · anti-cheat pass	Empty/random controls in the perimeter.	Universality of the stack.
F698-D	Full-candidate and reload influence.	Validated · no terminal verdict	Influence in some turns.	Influence on every response.
F760	Self-evaluation contract.	Validated · KEEP_CLEANUP	Limited gate and smoke.	Unlimited correctness.
F772	Internal priority.	Validated · KEEP_CLEANUP	Technical ranking.	Acquisition of knowledge.
F778	Stacks and micro-updates.	Validated · KEEP_CLEANUP	Candidate contract.	General promotion.
F779	State serialization.	Validated · KEEP_CLEANUP	Values restored and trace.	General semantic memory.

When an update does not demonstrate learning

Negative results and boundaries are part of the test. They show why a local score, a call or a delta is not enough.

Phase	Result	Limit imposed on the claim
F688-B	Integrity is not; benchmark too narrow and leakage flags.	The first benchmark loop does not prove robust causality.
F689-A	BENCHMARK_WEAKNESS_FOUND; candidate 1, clean blind 0,8375.	A high candidate score is not enough; the phase is Failed.
F689	Status Partial; candidate promoted=false.	Candidate Pass is not the same as ordinary runtime.
F698-C	Components used, but Partial integration.	Called is not the same as output-causal.
F709-B	Default integrated stack false; generator does not uniformly realize the meaning.	The stack does not govern the entire pipeline.
F686	Delta and reload in the candidate.	It does not prove ordinary activity or independent output effect.
F853	Structural verification without training or checkpoint modification.	Cleanup/test is not new learning.
F857 + audit V2	Association linked, without uniform live selection.	Structural presence does not demonstrate influence.

Internal mechanisms and continuous learning are not the same thing

This hub describes how specific components update, evaluate, consolidate, sort, save, or gate a state. Continuous learning instead concerns the progressive acquisition of new knowledge together with the retention of previous knowledge, generalization and catastrophic forgetting. An internal mechanism can contribute to the protocol without demonstrating continuous learning on its own.

Likewise, a memory write or persistence result demonstrates retention and recall on the relevant test, not automatically learning. The general runtime map remains the responsibility of the Architecture hub; only paths that modify or evaluate state are of interest here. The language and reasoning pipelines are linked when they receive influence, but are not duplicated.

Limits, security and privacy

There is no complete public audit for all risks of a system that updates or maintains state. Open problems remain: incorrect data, contamination, unwanted updates, provenance of information, correction, isolation between contexts, deletion and privacy. The presence of gates or candidates does not authorize us to declare these points resolved.

The experiments do not demonstrate general autonomic learning, continuous activity of each module, persistence of each update, or uniform influence on the output. Technical names like dream, analogy or self-evaluation do not involve psychological dreams, human intuition or self-awareness.