Getting started

AILL Research

Real Applied AI benchmarks, persistent internal memory, native text output, and public-safe research artifacts.

Elena AILL is the main research project. The current active runtime path is SenzaAdam Big Brain + Clean Speak: Big Brain provides persistent internal memory and context, while Clean Speak is the controlled text output layer.

Earlier SenzaAdam Brain + SPEAK-NEW work remains documented as validated lineage. GLIA, Hebbian, synaptic, and dense AILL experiments remain preserved as research history, not as the current public runtime claim.

Elena AILL does not grow only by becoming a bigger Transformer. The current active path focuses on local memory updates, bounded active cells, cold reload validation, anti-forgetting, and internal memory recall without external RAG or vector database runtime.

Research evidence only Real PNG charts only Current path: Big Brain + Clean Speak Reads, writes, codes, documents

Milestone 10M PASS

Stage 6 0.97561

Selfdoc 10/10

Demo Private

◆

10M parameter milestone validated — this preview is anchored to validated benchmark evidence and keeps the demo behind the existing quality gate.

What is Elena AILL

Elena AILL is the main research project and the public page is written as an AILL Research status surface plus an Applied AI benchmarks release. Every public claim is tied to benchmark evidence, report language, or PNG artifacts already generated in the workflow.

The system is built from scratch and documented conservatively: the overview introduces the runtime, while the architecture and public-surface pages explain lineage, constraints, and release boundaries in more detail.

Current active runtime

The active project path is test-modello/aill_originale. Big Brain is the persistent internal memory and context backend; Clean Speak is the final controlled text output layer. Big Brain is not allowed to generate the final answer directly.

SPEAK-NEW remains in the documentation for continuity, SEO, and research traceability. It is preserved as earlier validated memory-to-language lineage behind the current Clean Speak direction, not as the current runtime wording.

Current

SenzaAdam Big Brain

Persistent internal memory based on local cells, synaptic associations, bounded active paths, cold reload checks, and anti-forgetting validation.

Current

Clean Speak

The output layer that turns internal memory and context into controlled Italian text and decides when to answer or ask for clarification.

Lineage

SPEAK-NEW lineage

Earlier validated output and runtime lineage preserved for continuity. It remains documented, but the current website formula is Big Brain + Clean Speak.

Research-only

Archived and legacy paths

Older dense AILL, archived branches, and external demo or tool paths remain preserved as historical research, not as active runtime.

Validated Applied AI benchmarks for the 10M persistent-memory milestone

This benchmark bar is intentionally compact but specific. It summarizes the verified 10M SenzaAdam result using the same measured fields that appear in the public reporting workflow: learned facts, created cells, reused cells, recall, anti-forgetting, active path size, and persistent brain artifact size.

That matters for Applied AI indexing because the page does not ask the reader to trust a vague claim such as memory improved or performance scaled. It exposes the exact benchmark vocabulary used to describe what changed, what remained stable, and what the public package can defend.

Sample facts10,000,000

Learned facts10,833,323

Created cells394,856

Reused cells10,438,467

Recall0.9883

Anti-forgetting1.0000

Active cells avg1.00

Brain file2,358.25 MB

StatusPASS

OptimizerNo Adam

Global backwardDisabled

10-second summary

What Elena AILL has achieved so far — in 10 seconds

Elena AILL is an independent AI research project built from scratch in PyTorch, focused on persistent internal memory, bounded context, and modular training.

Developed and tested on consumer hardware, including an NVIDIA RTX 4060 Ti 16GB.

Validated milestones

1M + 10M completed

10,833,323 learned facts validated.

Memory signals

Memory signals stayed strong

recall_score 0.9883 · anti_forgetting_score 1.0000

Active validated modules

Big Brain + Clean Speak

SenzaAdam Big Brain + Clean Speak is the active public formula, with earlier research lines preserved elsewhere in the documentation.

Current boundary

100M deferred, public scope controlled

100M was deferred, not failed, and the release scope stays focused on evidence, documentation, and validated artifacts.

Research status release: validated Applied AI benchmarks, benchmark metrics, and real generated PNG artifacts.

Core modules

Module	Role	Status
SenzaAdam Big Brain	Persistent internal memory and context layer based on local cells, bounded active paths, cold reload checks, and anti-forgetting validation.	Active
Clean Speak	The output layer that turns internal Big Brain context into controlled Italian text and decides when to answer or ask for clarification.	Active
SPEAK-NEW lineage	Earlier validated output lineage preserved for continuity. It remains documented, but the current website formula is Big Brain + Clean Speak.	Active
Archived and legacy paths	Dense AILL history, older branches, and archived runtime experiments are preserved as research history, not active runtime.	Archive

Design constraints

No global backpropagation. All weight updates are computed from local cell activity. From Phase 41 onwards, the entire training pipeline uses local rules only.
No transformer attention. Context and long-range dependencies are handled through the Big Brain memory layer and native cell connections.
No external model dependencies. No third-party weights are loaded at runtime. No cloud inference API is called at any stage.
No borrowed architecture. Elena AILL is not a derivative of GPT, LLaMA, Mistral, or any other published model family.
Identity stability. Self-reference outputs must remain consistent across restarts and checkpoint loads. First validated in Phase 16b; stable since.

Current status

As of June 2026, training is in Stage 7 (Phase 45+). The 10M parameter milestone is validated and stable. Architecture documentation, benchmark results, and public artifacts are accessible and aligned with the repository evidence.

⚠

Demo access is not yet open. The public demo is gated behind the Stage 7 benchmark quality check. This page will be updated when that gate passes. Do not assume demo availability based on milestone progress alone.

Architecture

System design

Elena AILL uses no gradient descent, no attention mechanisms, and no pre-trained external weights. The full system is composed of four native modules developed and validated independently.

Current validated runtime architecture and public evidence boundary

Elena AILL works through its native runtime: SenzaAdam Big Brain + Clean Speak.

The project was not started as a fine-tuned commercial model or an external LLM wrapper. Elena AILL was implemented from scratch in PyTorch, with custom research modules developed and tested on local hardware, including an RTX 4060 Ti 16GB.

The public package publishes verified research data, benchmark reports, architecture history, and real PNG charts only. It does not advertise public chat, upload, SaaS access, or a live web demo.

This boundary is deliberate and useful for accuracy. The page can be read as an AILL Research document about persistent memory, Applied AI benchmarks, and modular runtime validation without pretending to be something the project is not claiming.

Main project

Dense history preserved, validated path clarified

Earlier dense AILL models remain preserved as research history while the active public path is SenzaAdam Big Brain + Clean Speak, backed by persistent internal memory and source-traced benchmark evidence.

Modular training

Not fine-tuning

This is modular training, memory ingestion, and persistent brain updates rather than a single monolithic fine-tuning story.

Public boundary

Evidence before product surface

The website surfaces benchmark evidence first. The interactive demo stays private until the Stage 7 writer is ready.

Overview gallery of the Elena AILL public benchmark package with Stage 2 to Stage 6 research visuals — Overview gallery of the public benchmark pack: Stage 2 to Stage 6 evidence, runtime scorecards, source-trace charts, ingestion charts, and public-ready benchmark sheets collected in one PNG.

Readers should read this architecture section as the shortest honest summary of the live public state: Elena AILL is the main research project, SenzaAdam Big Brain stores and recalls persistent internal memory and context, Clean Speak turns that context into controlled written output, and the public site documents evidence before claims.

Research Lineage

Elena AILL remains the main project, while the public site distinguishes research lineage from active runtime. Earlier dense AILL history remains preserved, GLIA stays as a research track, and the current public formula centers on SenzaAdam Big Brain + Clean Speak.

This distinction is central to how the project should be indexed and understood. In practical Applied AI terms, the site separates preserved historical work, currently validated runtime components, and the public evidence boundary that limits what can honestly be claimed today.

Elena AILL Main Project
├── Dense AILL Research History
│   ├── earlier dense models and checkpoints preserved
│   ├── legacy benchmark references kept
│   └── not the current public runtime path
│
├── GLIA Research Track
│   ├── memory and routing validation
│   ├── internal recall and QA from memory
│   └── preserved benchmark evidence
│
├── SenzaAdam Big Brain
│   ├── active persistent-memory and context module
│   ├── no Adam optimizer state
│   ├── no global backward on the active memory path
│   └── verified 10M benchmark result
│
├── Clean Speak
│   ├── current text output layer
│   └── final response layer for the active public formula
│
├── SPEAK-NEW
│   └── documented historical memory-to-language lineage
│
└── Stage 7
    └── integration, packaging, and release scope

Main project

Elena AILL stays the top-level project

Dense AILL history remains public research context, but it is no longer the active validated public runtime path.

GLIA

GLIA research track

GLIA remains preserved as technical history and benchmark evidence, not as the current public serving runtime.

Persistent memory

SenzaAdam Big Brain

The active public research module stores and recalls learned facts as internal memory and context without Adam optimizer state or global backward in the active path.

Output layer

Clean Speak

Clean Speak turns internal Big Brain context into controlled Italian written answers while keeping benchmark accountability visible. SPEAK-NEW remains documented lineage.

Validation

Stage 2-6 public evidence

The current public package publishes only validated benchmark stages, reports, and real PNG charts.

Release gate

Stage 7 before public demo

The current focus is to improve written output quality, coding reliability, and document consistency without weakening the memory-grounded runtime.

GLIA Research Track

GLIA is a preserved Elena AILL research track for memory and routing validation. It was used to explore entity recall, internal memory recall, routing, QA from memory, and interference-resistant recall. GLIA is not the current public runtime, but it helped validate ideas that informed the current Elena AILL direction: persistent memory, bounded routing, and modular architecture.

GLIA remains in the project because it explains where part of the current vocabulary came from. The public discussion of routing, internal recall, entity recall, and interference resistance did not appear suddenly with the 10M SenzaAdam result; it was shaped by earlier isolated validation work.

GLIA non-Transformer memory metrics chart from the isolated validation run — GLIA Core v2 isolated validation metrics. The PNG summarizes the preserved GLIA memory and routing validation block used before the current Big Brain runtime path, with the chart values presented directly inside the saved benchmark image.

AILL plus GLIA three-seed probe stability figure for seeds 0, 1, and 2 — AILL plus GLIA three-seed probe stability. The image compares repeated probe behavior across seed 0, seed 1, and seed 2, so the reader can see that the GLIA probe was checked across multiple seeds instead of only one run.

The key distinction is preserved on purpose: GLIA is searchable public research history, not the current serving claim. Keeping that difference explicit makes the site more indexable and less misleading because historical validation and current runtime status are not merged into one vague headline.

SenzaAdam Big Brain

The persistent memory and context layer. Replaces the classical optimizer with a native accumulation rule that retains knowledge across sessions without resetting weights between contexts.

Unlike a standard neural network trained with Adam or SGD, Big Brain does not compute a global loss. Memory consolidation runs on a fixed cadence — earlier knowledge is not evicted when new knowledge is added.

Property	Value
Learning rule	Local accumulation only
Global loss	None — by design
Memory retention	Persistent across restarts
Scale validated	100 MB (Phase 42, May 2026)
External deps	Zero

Memory

Persistent verified memory

The validated public path centers on memory that can be ingested, retained, recalled internally, and audited instead of hidden behind opaque generation-only behavior.

Compute

Active path stays bounded

Stage 6 shows cortex_used_ratio 0.878049 with context_used_ratio 0.0, reinforcing that the active memory path is actually used during the validated benchmark flow.

Scope

Benchmark evidence, not product marketing

The public framing stays on verified memory behavior, source traceability, and modular runtime validation.

Stage 5 runtime memory scorecard for the SenzaAdam Brain and SPEAK-NEW runtime — Stage 5 runtime memory scorecard. Key values visible inside the PNG include GPU NVIDIA GeForce RTX 4060 Ti, VRAM total 15.996 GB, VRAM used 1.377 GB, VRAM free 14.355 GB, RAM total 31.303 GB, RAM used 1.145 GB, process memory 601.473 MB, brain_file_size_mb 0.173, runtime_mode SenzaAdam Brain + SPEAK-NEW, writer_mode deterministic, and PASS across the listed checks.

Stage 6 cortex versus context usage chart showing Elena AILL self-documentation relying on cortex memory usage — Stage 6 cortex versus context usage. The chart shows cortex_used_ratio 0.878049 and context_used_ratio 0.0, which means the validated self-documentation run is using cortex memory instead of relying on plain prompt context.

This is why the SenzaAdam section talks in terms of anti-forgetting, active cells, cortex usage, and persistent brain size. Those are concrete signals that help both human readers and search systems understand what kind of memory behavior the project is actually validating.

Clean Speak / SPEAK-NEW Lineage

Clean Speak is the current layer used to turn internal Big Brain context into controlled Italian written answers. SPEAK-NEW remains documented as the earlier validated memory-to-language lineage, so older benchmark language stays searchable without confusing the current runtime formula.

The project is careful here because better phrasing alone is not enough; the public release threshold depends on language quality staying grounded in internal memory context, traceability, and controlled response behavior.

Writer mode

Deterministic is the default

The public benchmark pack keeps the deterministic writer as the default validated mode while the decoder remains optional and experimental.

Language

Controlled Italian answers

Stage 6 reaches answer_valid_italian_ratio 1.0 while preserving source_trace_success 1.0 on the self-documentation evaluation set.

Release gate

Output quality comes before broader claims

The current priority is better output quality while remaining memory-grounded, not widening claims beyond what the runtime has already validated.

Stage 5 writer mode comparison chart showing deterministic and decoder benchmark modes for SPEAK-NEW — Stage 5 writer mode comparison. The PNG compares deterministic and decoder writer modes on the same runtime benchmark slice, with deterministic kept as the validated public-safe writer path.

Stage 6 source trace success chart showing fully traced self-documentation answers — Stage 6 source-trace success. The benchmark figure records source_trace_success 1.0 on the self-documentation evaluation set, meaning the validated answers stayed fully traceable to their local sources.

That is also why controlled written output remains the default public reference point. It gives the cleanest benchmark surface for evaluating memory-grounded language behavior while broader generation remains secondary until it can be documented with the same level of reliability.

Native Hebbian Cells

Connection units that strengthen when both the source and target cells are active simultaneously. This is the only weight update mechanism in the system — there is no gradient signal of any kind.

Every cell in the system uses this rule. There is no layer that falls back to standard gradient descent as a fallback. The constraint is architectural, not a configuration option.

SenzaAdam optimizer

A custom optimizer that does not use gradient moments (no first or second moment estimates as in Adam). Finalized and stabilized in Phase 41. The optimizer is not a wrapper around PyTorch's optimizer base class — it is implemented independently.

Runtime environment

Property	Value
Compute	CUDA-accelerated · WSL/Linux
External LLM calls	Zero — at any runtime stage
Borrowed weights	None
Identity stability	Verified Phase 16b · Stable since
Backpropagation	Removed Phase 41 · Not present anywhere in pipeline

Training phases

All phases

Chronological record of every training phase. Each phase builds directly on the previous. There have been no regressions, no architecture pivots, and no rebrands across the full history of the project.

◆

This page now lists both the early core phases and the later public-safe publication phases already documented in the repository. Internal-only work may still exist between published checkpoints, but the publicly described phase record is expanded here instead of being compressed into only a few macro blocks.

Phase groups

Phase 1–10 Architecture foundations Done March 2026

Core cell-to-cell connection logic established. Native Hebbian rule prototyped. No global loss function from the very first implementation. Architecture decisions made in this block form the immutable base of all later work.

Phase 11–20 Critic sidecar · No-attention path Done April 2026

First clean no-attention path. Critic module implemented and validated in isolation (Phase 11c). No-attention chooser calibration completed (Phase 12f). Shadow dry-run confirmed no regression on earlier checkpoints.

Phase 21–30 Unified runtime · Clean Speak Done May 2026

Unified critic and runtime (Phase 22). Output repair work (Phase 28). Natural output repair stabilized (Phase 30b). Clean Speak became the stable output module. 1 million parameter milestone validated during this block.

Phase 31–40 Speak from zero · Bootstrap repair Done May 2026

Speak-from-zero bootstrap (Phase 37c–37f). Bootstrap repair cycles completed. Language speak path v1 functional and validated. Quote and writing repair (Phase 40) finalized. All output paths stable.

Phase 41 No global backpropagation Done May 31, 2026

Global backpropagation completely removed from all training paths. SenzaAdam optimizer cleaned and finalized. All weight updates are now computed exclusively from local cell activity. Verified stable across all existing checkpoints with no regression.

Phase 42 Big Brain 100MB — zero backprop Done May 31, 2026

Big Brain running stably at 100 MB scale with zero global backpropagation confirmed. Memory accumulation verified. Consolidation cadence validated. No instability observed at this scale. No regressions on earlier checkpoints.

Phase 44 Dialog cells — Micro Brain v5 Done June 2, 2026

Dialog micro-brain v5 validated end-to-end. Cell-chat native runtime confirmed. All dialog cell interactions are local. No external routing or sampling. Identity outputs stable throughout all test turns.

Phase 45+ Stage 7 — Integration hardening Active June 2026

Public website evidence surface refinement. Runtime framing and public artifact alignment. Stage 7 integration hardening. Currently in progress and focused on coherence between claims, artifacts, and validated runtime behavior.

Phase 51+ 100M parameter target Planned TBD

The next scale milestone. Architecture is ready for scale — only training compute and validation time are required. Demo access is gated behind Stage 7 completing first.

Published later phases

Phase 69 Hebbian / synaptic consolidation hardening Done Public-safe published

Phase 69 validated synaptic consolidation for the SenzaAdam Big Brain + Clean Speak path. The memory line stopped being only accumulation and started reinforcing local links while preserving old recall.

README data

Gate: PASS
Synapses reinforced: 325
Weak links softened: 19
Merge candidates: 279
Old recall: 9/10
New recall sample: 100/100
Anti-forgetting score: 1.0
Active cells average: 3.0
Reload from checkpoint: PASS

Phase 70 Modular memory scale-up to 10k facts Done Public-safe published

Phase 70 validated the controlled jump from 1k to 10k persistent facts while keeping Big Brain as internal memory only and Clean Speak as the final output layer.

README data

Gate: PASS
Facts applied: 10,000
Semantic fact index count: 11,056
Old recall: 10/10
Phase 68 recall: 100/100
New recall sample: 1000/1000
Anti-forgetting score: 1.0
Active cells average: 3.0
Reload from checkpoint: PASS

Phase 71 Official runtime 10k conversation probe Done Public-safe published

Phase 71 validated the official runtime path after the 10k scale-up in a realistic conversation probe, with Big Brain used as context only and Clean Speak as final output layer.

README data

Status: PASS
Probe prompts: 100
Sensible answers: 100/100
Context used: 100/100
Ambiguity defer: 15/15
Old recall: 10/10
New recall sample: 1000/1000
Anti-forgetting score: 1.0
Active cells average: 3.0

Phase 72 Clean SynapticMemory backend V1 Done Public-safe published

Phase 72 recovered the old SynapticMemory idea as a clean bounded sparse backend without importing the old runtime directly.

README data

Status: PASS
Backend created: aill_synaptic_memory_backend_v1
Pre-integration test: PASS
Post-integration test: PASS
Official runtime 10k smoke: PASS
Sensible: 100/100
Context used: 100/100
Ambiguity defer: 15/15
Active cells average: 3.0

Phase 73 GLIA stability layer V1 Done Public-safe published

Phase 73 recovered the GLIA stability idea as a clean optional guard for fragile context, conflicting signals, and active-path pressure.

README data

Status: PASS
Backend created: aill_glia_stability_layer_v1
Pre-integration test: PASS
Post-integration test: PASS
Official runtime 10k smoke: PASS
Sensible: 100/100
Context used: 100/100
Ambiguity defer: 15/15
Active cells average: 3.0

Phase 74Runtime learn / confirm loopDonePublic-safe published

Phase 74 validated controlled runtime learning: Elena proposes updates, asks for confirmation, writes only confirmed facts into a working brain, and rejects unconfirmed information.

README data

Status: PASS
Proposals: 10
Confirmed updates written: 8
Rejected updates not saved: 2
Reload from checkpoint: PASS
Confirmed recalled after reload: 8/8
Rejected absent after reload: 2/2
Hallucination added: 0

Phase 75Conflict resolutionDonePublic-safe published

Phase 75 validated conflict protection before `.brain` updates: confirmed facts persist, contradictory ones are blocked, and ambiguous ones are routed to clarification.

README data

Status: PASS
Status match: 10/10
Applied updates: 5
Blocked conflicts: 3
Clarification items: 1
Duplicate/protected items: 1
Applied recalled after reload: 5/5
Blocked absent after reload: 3/3

Phase 76Long conversation memoryDonePublic-safe published

Phase 76 validated temporary conversation memory with safe consolidation: stable confirmed items persist after reload while volatile details stay outside persistent memory.

README data

Status: PASS
Conversation turns: 13
Stable candidates: 3
Consolidated candidates: 3
Reload from checkpoint: PASS
Consolidated recalled after reload: 3/3
Volatile facts absent after reload: 3/3

Phase 77Big Brain 100k controlled scaleDonePublic-safe published

Phase 77 validated the controlled 100k memory scale step and proved that the large index remained grounded, bounded, and reloadable.

README data

Status: PASS
New facts applied: 100,000
Semantic fact index after reload: 111,071
Reload from checkpoint: PASS
Indexed new-fact recall sample: 100/100
Runtime micro-smoke: 10/10
Anti-forgetting score: 1.0
Active cells average: 3.0

Phase 78Fast 100k runtime recallDonePublic-safe published

Phase 78 turned the 100k checkpoint into a practical runtime recall result with fast internal cache-based retrieval instead of broad scan per prompt.

README data

Status: PASS
Runtime recall new prompts: 100/100
Old memory prompts: 5/5
Average recall latency: under 50 ms
P95 recall latency: under 50 ms
Active cells average: 3.0
Hallucination added: 0

Phase 79Official 100k conversation runtimeDonePublic-safe published

Phase 79 validated the official 100k conversation runtime on a full prompt probe, with Big Brain used only as internal context.

README data

Status: PASS
Sensible responses: 100/100
Prompts using Big Brain context: 100/100
Ambiguous prompts deferred: 20/20
P95 latency: under 50 ms
Active cells average: 3.0
Hallucination added: 0

Phase 80100k runtime freeze milestoneDonePublic-safe published

Phase 80 froze the validated 100k runtime as an internal stable milestone while keeping the frozen checkpoint private and auditable.

README data

Status: PASS
Frozen milestone created
Source and frozen hashes match
Reload from frozen checkpoint: PASS
Runtime smoke sensible: 10/10
Ambiguous prompts deferred: 4/4
P95 latency: under 50 ms

Phase 81Private demo frozen 100kDonePublic-safe published

Phase 81 validated the frozen F80 runtime on a broader private 200-prompt demo probe.

README data

Status: PASS
Sensible responses: 200/200
Prompts using Big Brain context: 200/200
Ambiguous prompts deferred: 40/40
P95 latency: under 50 ms
Active cells average: 3.0

Phase 82Interactive chat CLI frozen 100kDonePublic-safe published

Phase 82 introduced the manual terminal chat path for talking to Elena on the frozen 100k runtime without any checkpoint writes or training.

README data

Status: PASS
Interactive CLI created
Test smoke: 5/5 sensible responses
Ambiguous prompt defer: 1/1
Checkpoint touched: false
Training started: false

Phase 83Clean Speak output expansion plus GPUDonePublic-safe published

Phase 83 fixed a major runtime weakness: strong memory but too few output patterns. The output layer became broader and GPU-backed.

README data

Status: PASS
Speech validation prompts passed: 100/100
Template repetition rate: below 0.15
False defer count: 0
Conversation: 30/30
Identity/capability: 20/20
Python/code/debug: 20/20
Ambiguity defer: 20/20

Phase 84Learned speech composer V1DonePublic-safe published

Phase 84 moved Clean Speak away from rigid runtime templates toward a learned text state built from sentence transitions and family-specific word structure.

README data

Status: PASS
Learned speech composer active
Template runtime disabled in test
Validation prompts passed: 60/60
Repetition rate: below 0.20
GPU path active
Checkpoint touched: false

Phase 85Learned speech composer V1 extendedDonePublic-safe published

Phase 85 extended the learned output state and hardened the runtime smoke test with lower repetition while keeping Big Brain as memory/context only.

README data

Status: PASS
Sensible responses: 100/100
Fixed runtime templates used: false
Repetition rate: 0.060000
GPU path: active
Checkpoint writes: none

Phase 86Memory-conditioned speechDonePublic-safe published

Phase 86 let internal Big Brain context influence the speech composer without allowing memory to generate the final answer directly.

README data

Status: PASS
Sensible responses: 120/120
Memory-conditioned responses: 88
Fixed runtime templates used: false
Repetition rate: 0.066667
GPU path: active

Phase 87Speech routing hardeningDonePublic-safe published

Phase 87 hardened the output router so ordinary prompts, simple math, capability questions, and typo-heavy requests were no longer hijacked by unrelated memory context.

README data

Status: PASS
Bug cases passed: 14/14
False memory conditioning: 0
GPU path: active
Fixed runtime templates used: false
Checkpoint writes: none

Phase 88Unlabeled speech trainingDonePublic-safe published

Phase 88 trained the output state from unlabeled natural text instead of prompt-answer pairs while keeping the runtime clean.

README data

Status: PASS
Sensible responses: 200/200
False memory conditioning: 0
Repetition rate: 0.060000
GPU path: active
Supervised prompt-response training: false

Phase 895GB speech training gateNot promotedPublic-safe published

Phase 89 staged a clean 5GB dataset gate and trained a broader speech state, but manual smoke exposed corpus-style phrasing so the result stayed experimental and was not promoted.

README data

Status: CREATED / NOT PROMOTED
Selected size: 5.268907 GB
Automated runtime check: PASS
Manual smoke: FAIL_NOT_PROMOTED
Default runtime restored to F88
GPU path: active

Phase 9110GB speech and practical gateDonePublished summary

Phase 91 promoted the 10GB speech and practical gate after manual smoke. The public repo currently exposes the milestone summary, but not a dedicated per-phase README under `da-pubblicare`.

Phase 92Chat regression hardeningDonePublished summary

Phase 92 hardened chat regression around identity, capability, memory, and uncertainty responses. The public repo currently exposes the milestone summary, but not a dedicated per-phase README under `da-pubblicare`.

Phase 9320GB speech and practical gateDonePublished summary

Phase 93 passed the 20GB speech and practical gate with GPU-batched scoring on RTX 4060 Ti. The public repo currently exposes the milestone summary, but not a dedicated per-phase README under `da-pubblicare`.

Phase 94Practical output depth gateDonePublished summary

Phase 94 passed the practical output depth gate for code, debug, work messages, and document summaries. The public repo currently exposes the milestone summary, but not a dedicated per-phase README under `da-pubblicare`.

Phase 95Elena vs LLM-style baseline benchmarkDonePublic-safe published

Phase 95 added the first local repeatable Elena vs LLM-style baseline benchmark and finally exposes the exact benchmark deltas instead of a vague comparison sentence.

README data

Status: PASS
Cases: 20
Elena overall: 4.05/5
Baseline overall: 1.275/5
Elena wins: 19
Baseline wins: 0
Ties: 1
Code: Elena 5.0/5 vs Baseline 1.0/5
Debug: Elena 4.0/5 vs Baseline 2.0/5
Math: Elena 4.0/5 vs Baseline 4.0/5
Memory: Elena 4.0/5 vs Baseline 1.0/5
Safety: Elena 4.0/5 vs Baseline 1.0/5

Phase 96 Continual learning work checkpoint Done Public-safe published

Elena received a small verified local memory update, saved it into a work brain checkpoint, reloaded from zero, and preserved both prior and new memory. This is the public-safe continual-learning milestone before runtime promotion.

Phase 97 Runtime promotion gate for F96 Done Public-safe published

The F96 work checkpoint was tested inside the live runtime path before promotion. Clean Speak remained the final output layer, no training ran, and no checkpoint overwrite was performed during the gate.

Phase 98 Frozen F80 vs work F96 runtime A/B Done Public-safe published

The frozen F80 runtime and the F96 continual-learning work checkpoint were compared on the same local runtime prompts. No training, repair, checkpoint write, or frozen overwrite was performed during the comparison.

Phase 99 F96 promoted to new frozen runtime Done Public-safe published

The validated F96 work checkpoint was promoted into a new frozen runtime milestone while preserving the older F80 frozen checkpoint. This phase marks the runtime promotion boundary, not a retraining phase.

Phase 100Enterprise manual runtime gateDonePublic-safe published

Phase 100 ran the 100-prompt enterprise-style manual gate on the promoted F99 runtime without training, repair, or checkpoint writes.

README data

Status: PASS
Readable pass: 95/100
Sensible pass: 90/100
Ambiguity pass: 16/16
Latency avg ms: 87.406
Capability: 5/5
Code: 5/5
Document: 10/10
Memory: 14/15

Phase 101F99 runtime hardeningDonePublic-safe published

Phase 101 hardened the enterprise/work composer and ambiguity handling of F99 without any training or checkpoint writes.

README data

Status: PASS
Sensible before: 75/100
Sensible after: 81/100
Ambiguity before: 12/16
Ambiguity after: 13/16
Pseudo-word ratio: 0.005263

Phase 102Memory priority and ambiguity hardeningDonePublic-safe published

Phase 102 improved grounded memory routing and made ambiguity/defer behavior stricter without training or checkpoint writes.

README data

Status: PASS
Sensible before: 81/100
Sensible after: 85/100
Ambiguity before: 13/16
Ambiguity after: 16/16
Latency avg ms: 79.027

Phase 103Enterprise gate 90/100DonePublic-safe published

Phase 103 raised the promoted F99 runtime to a 90/100 sensible enterprise gate without training or checkpoint writes.

README data

Status: PASS
Sensible before: 85/100
Sensible after: 90/100
Ambiguity before: 16/16
Ambiguity after: 16/16
Latency avg ms: 87.406

Phase 104Blind runtime validationFailPublic-safe published

Phase 104 deliberately exposed the promoted F99 runtime to a new blind prompt set and showed that the previous enterprise gate did not generalize well enough.

README data

Status: FAIL
Readable pass: 99/100
Sensible pass: 59/100
Ambiguity pass: 13/15
Latency avg ms: 82.525
Capability: 1/4
Conversation: 1/4
Debug: 1/4
Math: 4/4

Phase 105Blind failure learning diagnosticDonePublic-safe published

Phase 105 turned F104 blind failure families into local brain updates instead of prompt-answer patching, while leaving the promoted frozen F99 checkpoint untouched.

README data

Status: PASS
Families updated: 9
Cells updated: 54
Synapses updated: 135
Weights updated: 894
Semantic fact index count: 111090
Checkpoint created: true
Frozen checkpoint overwritten: false

Phase 106Learned signal gating diagnosticFailPublic-safe published

Phase 106 hardened the runtime gate for the learned signal family created in F105, but the retest still failed the target quality bar.

README data

Status: FAIL
F104 sensible retest: 65/99
Holdout sensible: 21/30
Ambiguity pass: 16/18
Learned signal seen: 109
Learned signal used: 0
Memory: 5/10
Memory new: 5/10

Phase 107Memory-to-speech calibrationFailPublic-safe published

Phase 107 calibrated how learned memory signals influence speech. It improved some areas but still did not meet the publication bar for the full retest.

README data

Status: FAIL
F104 sensible retest: 70/99
Holdout sensible: 21/30
Ambiguity pass: 14/18
Learned signal seen: 109
Learned signal used: 55
Explain: 8/8
Memory new: 3/10

Phase 108–111Hierarchical memory routing and runtime integrationDonePublished summary

This block introduced hierarchical memory routing into the runtime so Elena can use condensed internal context before expanding into larger memory paths. The goal was to keep active context bounded while making the runtime more practical on longer tasks.

README summary

Hierarchical memory routing: PASS
Runtime integration: PASS
Bounded active context preserved
F99 remains the preserved frozen milestone

Phase 108Hierarchical memory routing V1DonePublished summary

Phase 108 started the hierarchical memory-routing line, introducing a condensed internal routing step before broader memory expansion.

Phase 109Hierarchical routing hardeningDonePublished summary

Phase 109 hardened the first hierarchical routing pass so context selection remained local, bounded, and compatible with the existing runtime path.

Phase 110Bounded context expansionDonePublished summary

Phase 110 refined how the runtime expands from compact context into larger internal memory paths without losing bounded active-context discipline.

Phase 111Hierarchical runtime integrationDonePublished summary

Phase 111 completed the first runtime integration of hierarchical memory routing, making the routing line part of the practical Elena runtime.

Phase 112–114GPU context execution and chat CLI integrationDonePublished summary

This block validated GPU context execution on the RTX 4060 Ti path. Cached GPU context was made substantially faster, and the GPU context path was wired into the manual chat CLI as an optional runtime flag.

README summary

GPU context path: PASS
Hardware: RTX 4060 Ti
F113 cached GPU context: about 9.6x faster than first GPU pass
F114 chat CLI GPU flag: integrated

Phase 112GPU context executionDonePublished summary

Phase 112 validated GPU context execution on the practical Elena runtime path.

Phase 113Cached GPU context latencyDonePublished summary

Phase 113 reduced cached GPU-context latency by about 9.6x relative to the first GPU pass, making the path more practical in repeated runtime use.

Phase 114Chat CLI GPU context flagDonePublished summary

Phase 114 wired the GPU context path into the manual chat CLI as an optional runtime flag without changing the frozen runtime milestone.

Phase 115–123Real chat quality, hard talk, and readiness validationDonePublished summary

This block improved real chat quality, context alignment, manual talk behavior, local baseline benchmarking, and hard-talk readiness around the preserved F99 line. It is a runtime-hardening series, not a new frozen release.

README summary

F115 real chat quality audit: PASS
F118 manual talk regression: 56/60 sensible
F119 Elena vs local LLM-style baseline: Elena 158/200, baseline 23/200
F122 hard talk: 480/500 sensible, ambiguity 50/50
F123 aggregate readiness: 688/760 sensible
F123 readable: 760/760
F123 ambiguity defer: 83/91
F123 pseudo_word_ratio: 0.000000
GPU context: active
checkpoint_touched: false
frozen overwritten: false

Phase 115Real chat quality auditDonePublished summary

Phase 115 audited real chat quality on the live runtime path and marked the starting point of the later readiness block.

README summary

Status: PASS

Phase 117Context-to-output alignment hardeningDonePublished summary

Phase 117 hardened the alignment between internal context and final runtime output so practical answers stayed grounded without checkpoint writes.

Phase 118Manual talk regression after hardeningDonePublished summary

Phase 118 measured manual talk quality after the later routing and alignment hardening passes.

README summary

Sensible responses: 56/60
No checkpoint writes during validation

Phase 119Elena vs local LLM-style baselineDonePublished summary

Phase 119 compared Elena against a local LLM-style baseline, not a live commercial model, to measure practical runtime quality on the same benchmark slice.

README summary

Elena sensible: 158/200
Baseline sensible: 23/200
Live commercial model used: false
checkpoint_touched: false

Phase 120Hard talk benchmark startDonePublished summary

Phase 120 opened the final hard-talk readiness block around the preserved frozen runtime line.

Phase 121Context routing hardeningDonePublished summary

Phase 121 hardened context routing for practical requests so runtime behavior stayed usable under heavier conversational pressure.

Phase 122Hard talk 500DonePublished summary

Phase 122 ran the 500-prompt hard-talk benchmark as the main practical stress test of the later runtime-hardening line.

README summary

Sensible responses: 480/500
Ambiguity defer: 50/50

Phase 123Speaking runtime readinessDonePublic-safe published

Phase 123 aggregated the readiness evidence across F115, F118, and F122 without promoting a new frozen runtime checkpoint. F99 remained the preserved frozen milestone during this validation block.

README data

Sensible: 688/760
Readable: 760/760
Ambiguity defer: 83/91
pseudo_word_ratio: 0.000000
GPU context: active
checkpoint_touched: false
training_started during validation: false
frozen_checkpoint_overwritten: false

Phase 135–156Research-state runtime expansion, skills, cleanup, and dataset preparationDonePublished summary

This later block added larger research-state speech runs, planner routing, optional CodeSkill and Message/Document skills, controlled cleanup of legacy external material, routing fixes, and a cleaned dataset for future controlled training.

Phase 13550GB GPU speech-state research runDonePublished summary

Phase 135 completed a larger 50GB GPU speech-state run on the RTX 4060 Ti path as a research state, but it was not promoted as the official frozen runtime.

README summary

Status: PASS / research state not promoted
Scanned data: 50.0 GB
Accepted sentences: about 1.8M
Neutral Elena-only validation: 50/50
Frozen checkpoint overwritten: false

Phase 139-BResponse Planner V1DonePublished summary

Phase 139-B added Response Planner V1 as a routing layer, not as a prompt-answer lookup and not as a new brain checkpoint. The 50GB state remained research-only because it did not beat the default state after planner hardening.

README summary

Chat regression: PASS
50GB A/B chat: 58/60
Neutral Elena-only: 50/50
Runtime audits: PASS
checkpoint_writes: none

Phase 141CodeSkill V1 recoveryDonePublished summary

Phase 141 recovered CodeSkill V1 as a clean optional module inside `aill_originale`, keeping code and debug structure separate from old external scripts.

README summary

Smoke checks: PASS
Runtime audits: PASS
checkpoint_writes: none
JSON runtime memory: not used

Phase 142CodeSkill V1 runtime integrationDonePublished summary

Phase 142 integrated CodeSkill V1 as an optional runtime path for code and debug prompts.

README summary

CodeSkill benchmark: 20/20
Chat regression: PASS
Runtime audits: PASS
Checkpoint or frozen runtime modified: false

Phase 143Message and Document Skill V1DonePublished summary

Phase 143 integrated Message/Document Skill V1 as an optional runtime path for practical messages and document prompts.

README summary

Message/Document benchmark: 15/15
CodeSkill benchmark: 20/20
Chat regression: PASS
Runtime audits: PASS
External target-answer data imported: false

Phase 144External speak dataset archiveDonePublished summary

Phase 144 archived old external speak datasets and rollback backups with target-answer or retrieval-style fields out of the active clean project path.

README summary

Active clean project: `test-modello/aill_originale`
Chat regression: PASS
CodeSkill: PASS
Message/Document Skill: PASS
Runtime audits: PASS

Phase 145Prepared dataset archiveDonePublished summary

Phase 145 archived old external prepared dataset versions out of the repository root while keeping them preserved for traceability.

README summary

Active clean project: `test-modello/aill_originale`
Post-cleanup runtime checks: PASS

Phase 146Legacy app and tool archiveDonePublished summary

Phase 146 archived old external app, demo, autolearning, training, and tool folders out of the active root while preserving them for traceability.

README summary

Active clean runtime path: `test-modello/aill_originale`
Post-cleanup runtime checks: PASS

Phase 147Remaining external root auditDonePublished summary

Phase 147 audited the remaining external root folders including old root code, scripts, checkpoints, benchmark packs, generated outputs, rescue checkpoints, and the large corpus vault.

README summary

Files moved in this audit: none
Active clean project: `test-modello/aill_originale`

Phase 153Clean Speak routing fixDonePublished summary

Phase 153 fixed Clean Speak routing around real chat failures. Big Brain remained context-only, practical routing improved, and checkpoints stayed untouched.

Phase 155Generative Speech Core V1DonePublished summary

Phase 155 added Generative Speech Core V1 as a controlled speaking component to improve sentence composition without letting Big Brain generate final answers.

README summary

Chat regression: PASS
CodeSkill: PASS
Message/Document Skill: PASS
Runtime audits: PASS

Phase 15610GB cleaned speech dataset preparationDonePublished summary

Phase 156 prepared a 10GB cleaned speech dataset from available corpus material for future controlled training without creating or promoting a new checkpoint in this phase.

README summary

Dataset preparation: completed
Promoted checkpoint created: false
Public runtime release changed: false

Phase 148Remaining external root archiveDonePublished in root README

Phase 148 completed the controlled cleanup of the remaining external root folders and made `test-modello/aill_originale` the only active runtime and development path.

README data

Status: PASS
Archived files: 2,930
Archived bytes: 5,550,970,124
Dataset libri touched: false
Checkpoint touched: false
Chat regression: PASS, 29/30 sensible
Message/Document skill: PASS, 15/15
CodeSkill: PASS, 20/20

Phase 149Speech runtime follow-up hardeningDonePublished in root README

Phase 149 targeted short follow-up prompts such as `si fallo`, `ok procedi`, and `continua`, reducing generic fallback behavior without changing the default speech state.

README data

Status: PASS
GPU used: true, RTX 4060 Ti
F149 test: 31/32 sensible
Pseudo-word ratio: 0.000000
Chat regression: PASS, 29/30 sensible
Message/Document skill: PASS, 15/15
CodeSkill: PASS, 20/20
Speech state promoted: false

Phase 150Conversation carryover runtimeDonePublished in root README

Phase 150 added a small runtime carry-over layer for short multi-turn chat so Elena can use the previous task context on concise follow-up turns.

README data

Status: PASS
Conversation carry-over test: 14/14 sensible
Carry-over used: 6 turns
Pseudo-word ratio: 0.000000
Chat regression: PASS, 29/30 sensible
Message/Document skill: PASS, 15/15
CodeSkill: PASS, 20/20
Persistent learning: not changed

Phase 151Elena Action Composer V1DonePublished in root README

Phase 151 introduced the first Action Composer layer between memory/context and final speech for executable tasks such as code, messages, and technical explanations.

README data

Status: PASS
Action Composer benchmark: 16/16 sensible
Action composer used: 6
Pseudo-word ratio: 0.000000
F150 carry-over regression: PASS, 14/14
Chat regression: PASS, 28/30 sensible
Message/Document skill: PASS, 15/15

Timeline

Milestone log

Key validated checkpoints in the Elena AILL research program. Every milestone listed here is backed by a reproducible benchmark run. No results are estimated or extrapolated.

Monthly Research Updates

Elena AILL is updated as an active research project. Each monthly update summarizes validated benchmark milestones, training status, integration progress, and next public validation steps.

New monthly sections are added only when the underlying result is validated and traceable. The timeline stays chronological month by month: historical milestones remain visible, the current month gets its own update card, and future work is described only after the last validated month already shown in public.

This slower cadence is deliberate. AILL Research updates function more like an engineering notebook and less like a product launch feed, so missing evidence stays private until it becomes measurable, reviewable, and fit for publication.

April 2026

1M SenzaAdam Milestone

Completed

Status: Completed historical milestone

Key updates

Elena AILL main project preserved
GLIA isolated validation completed
SenzaAdam 1M real-data brain passed
1,000,000 real concept and fact pairs
1,083,323 learned_facts
104,078 created_cells
979,245 reused_cells
recall_score = 0.9767
anti_forgetting_score = 1.0000
active_cells_avg = 1.00
no Adam
no global backward

Real pairs1,000,000

learned_facts1,083,323

created_cells104,078

reused_cells979,245

recall_score0.9767

anti_forgetting1.0000

active_cells_avg1.00

Training modeNo Adam

Next update

May 2026 / verified 10M result published, with stronger persistent-memory evidence and the roadmap shifted away from the 100M stress run.

May 2026

10M SenzaAdam Milestone

Completed

verified_10m_result = PASS
sample_facts = 10,000,000
learned_facts = 10,833,323
created_cells = 394,856
reused_cells = 10,438,467
recall_score = 0.9883
anti_forgetting_score = 1.0
active_cells_avg = 1.0
brain_file_size_mb = 2358.25

Next update

June 2026 / website package alignment, public wording cleanup, and integration hardening around SenzaAdam Big Brain + Clean Speak.

June 2026

Website package and integration alignment

In progress

public website narrative aligned around Elena AILL main project
current public runtime wording centered on SenzaAdam Big Brain + Clean Speak
verified 10M result kept as the public reference milestone
100M stress run deferred, not failed
SenzaAdam Big Brain + Clean Speak integration hardening continues
SPEAK-NEW remains documented as historical lineage
public benchmark package refresh stays limited to real PNG artifacts only
Stage 7 remains focused on integration, packaging, and public surface coherence
public claims remain limited to validated evidence and shipped artifacts

Apr 2026
Validated

1M parameters — first scale validation

Historical milestone. Strong retention results and bounded active cells confirmed that the architecture scales correctly from the very first checkpoint. No instability observed during the validation run. This result established the first public-facing evidence point for the project.

May 2026
Validated

10M parameters — reference benchmark milestone

The reference benchmark result for the public narrative. Big Brain and Clean Speak integrated as the stable official architecture for all subsequent work. All benchmark conditions passed including long context window tests. Activation remained bounded throughout. This is the milestone cited in all public-facing documentation.

May 31
Validated

Zero backpropagation — architecture complete

Phase 41 removed global backpropagation from the entire training pipeline. The architecture is now fully local — no gradient signal is computed at any stage. This was the final architectural constraint required before the system could be described as genuinely native.

Jun 2026
Active

Stage 7 — integration hardening

Currently in progress. Public evidence surface refinement, runtime framing, and Stage 7 integration hardening. The goal is a cleaner public package with tighter alignment between documentation, artifacts, and validated runtime behavior. Estimated completion within June 2026.

TBD
Planned

100M parameters — next scale target

Architecture is confirmed ready for this scale. Training compute allocation and validation schedule will be finalized after Stage 7 completes. This milestone is not gated on any architectural change — only on training time and compute availability.

Benchmarks

Results — 10M suite

The 10M parameter benchmark suite was run in May 2026. All checks below are reproducible. No results are cherry-picked, extrapolated, or measured under non-representative conditions.

Verified 10M internal benchmark result

The 10 million data milestone completed: internal 10M SenzaAdam benchmark evidence in the repo marks verified_10m_result as PASS on the read-only benchmark report.

The importance of this milestone is not scale alone. The stronger claim is that the benchmark remains stable while scale increases: retention stays high, anti-forgetting stays at the validated threshold, and the active path remains tightly bounded instead of expanding into broad opaque compute.

That is why the public wording stays narrow. The site reports a verified persistent-memory result under a specific benchmark protocol, not a general intelligence claim and not a promise that every private runtime component is already public.

learned_facts10,833,323

created_cells394,856

reused_cells10,438,467

recall_score0.9883

Retention

anti_forgetting_score = 1.0

The read-only 10M report shows no retention degradation in the validated benchmark threshold set.

Active path

active_cells_avg = 1.0

The active memory path remains tightly bounded instead of scaling by activating broad global compute.

Brain artifact

brain_file_size_mb = 2358.2526

The verified 10M benchmark report records a single persistent brain artifact instead of optimizer-state-heavy checkpoints.

Training constraints

No Adam, no global backward

The active memory path remains local-update based, with no Adam optimizer state and no global backward pass.

Taken together, these values show why the 10M result became the reference milestone for the current public narrative. More facts were learned, more cells were created and reused, retention stayed measurable, and the brain artifact remained explicit enough to describe as a persistent object rather than as a hidden optimizer-heavy side effect.

The roadmap shifted after strong 10M evidence, so 100M was deferred rather than failed

The 100M stress run was deferred, not failed, because the 10M run had already produced sufficiently strong evidence to shift engineering focus toward integration, public benchmark packaging, SenzaAdam Big Brain + Clean Speak validation, and Stage 7 runtime hardening for stronger written output, coding reliability, and document quality.

Deferral here is a prioritization choice, not a hidden negative result. Once the 10M run supplied a defensible benchmark story, the more valuable engineering problem became how to connect memory, language, runtime controls, and public evidence without creating a misleading public product surface.

Redirected

System integration

Engineering effort moved away from a larger stress run toward cleaner runtime integration and architecture clarity.

Published

Public benchmark packaging

The website and public package now prioritize traceable reports, metrics summaries, and real PNG artifacts only.

Validated

Big Brain + Clean Speak

Stage 2 through Stage 6 validation remains historical proof surface, while the current Elena AILL formula is SenzaAdam Big Brain + Clean Speak.

Stage 7 runtime hardening

The next public step is not bigger scale alone. It is better written output quality and cleaner runtime integration while staying grounded in memory.

This is also why the website is organized around benchmark explanation rather than just milestone celebration. A larger raw number would not have improved the public understanding of Elena AILL as much as a cleaner explanation of what the current validated Applied AI stack already proves.

Validated runtime, writer, and self-documentation benchmark evidence

The figures below are real PNG artifacts copied from the public benchmark pack and supporting reports only. No fake charts, no invented metrics, and no AI-generated infographic evidence are used here.

Each stage answers a different benchmark question. Stage 2 establishes core behavior, Stage 3 and Stage 4 increase robustness pressure, Stage 5 validates runtime behavior, and Stage 6 validates self-documentation, source traceability, and memory-grounded answering.

From an Applied AI perspective, this progression matters because it shows more than raw score accumulation. It shows whether a memory-grounded system can keep internal-recall discipline, fallback behavior, runtime accountability, and explainable sourcing while the public claims remain conservative.

Stage 2

PASS

Core deterministic and decoder runs both pass, while the validated hard set keeps deterministic slightly ahead at 0.953271 versus 0.943925.

Stage 3

PASS

Deterministic pass_rate reaches 0.97493 with memory_retrieval_success at 1.0 and unknown fallback success at 1.0.

Stage 4

PASS

Deterministic and decoder both reach pass_rate 1.0 on the validated Stage 4 benchmark set.

Stage 5

PASS

Runtime evidence records the earlier SenzaAdam Brain + SPEAK-NEW Stage 5 lineage on RTX 4060 Ti with writer_mode deterministic and brain_file_size_mb 0.173. The current public formula is SenzaAdam Big Brain + Clean Speak.

Stage 6

PASS

facts: 433, local sources: 9, mandatory selfdoc questions: 10/10, pass_rate: 0.97561, memory_retrieval_success: 0.972222, source_trace_success: 1.0, cortex_used_ratio: 0.878049.

Public scope

Research evidence only

These benchmark artifacts prove the current validated path without turning the site into a public chat product claim.

Stage 2, Stage 3, and Stage 4 pass rate comparison chart from the public SPEAK-NEW benchmark pack — Stage 2, Stage 3, and Stage 4 pass-rate comparison. The benchmark PNG compares pass-rate progression across the earlier SPEAK-NEW stages, including Stage 2 deterministic 0.953271 versus decoder 0.943925, Stage 3 deterministic 0.97493, and Stage 4 pass_rate 1.0.

Gallery sheet of real Stage 2 to Stage 4 benchmark charts from the Elena AILL public package — Stage 2-4 chart gallery assembled from real benchmark PNGs. The sheet groups pass rate, fallback behavior, memory retrieval success, context QA, distractor rejection, misleading-context handling, and multi-fact validation into one overview image.

Gallery sheet of Stage 5 private runtime benchmark charts for SenzaAdam Brain and SPEAK-NEW — Stage 5 private runtime evidence from the earlier SenzaAdam Brain + SPEAK-NEW lineage. The gallery sheet collects runtime memory scorecard, writer-mode comparison, Stage 2-4 pass-rate comparison, VRAM/runtime measurements, and other saved Stage 5 benchmark panels into one PNG.

Stage 6 self-documentation scorecard showing pass rate, memory retrieval success, and source trace success — Stage 6 self-documentation scorecard. Key values inside the PNG are pass_rate 0.97561, memory_retrieval_success 0.972222, source_trace_success 1.0, answer_valid_italian_ratio 1.0, mandatory_selfdoc_questions 10/10, 433 tracked facts, and 9 local sources.

Stage 6 ingestion facts by source chart showing local source distribution — Stage 6 facts ingested by local source. The PNG breaks down 433 ingested facts across 9 local sources, so the reader can see the source distribution used by the self-documentation benchmark instead of only the final aggregate score.

Gallery sheet of Stage 6 self-documentation ingestion charts from the Elena AILL public package — Stage 6 self-documentation gallery from the public benchmark pack. The sheet groups Stage 6 scorecard, source-trace success, cortex-versus-context usage, ingestion-by-source, and related self-documentation benchmark visuals into one summary PNG.

The PNG charts remain first-class public evidence because they are easy to archive, compare, and cite across updates. A visual artifact with a filename, stage label, and stable location helps readers verify that the public story is anchored to the same outputs used to build the benchmark package.

◆

This is the reference benchmark result cited in all public documentation. It supersedes the 1M suite (April 2026) which remains documented for historical completeness.

Parameter scale

Check	Result	Notes
Parameter count	10,000,000	Stable across all checkpoints. No parameter drift observed between runs.
Active cells	Bounded	Cell activation stayed within design bounds throughout. No runaway growth in short or long context conditions.
Learning rule	Local only	Zero global backpropagation at any stage. Confirmed by code audit and training log inspection.
Identity anchor	Stable	Self-reference outputs consistent across 12 restart cycles and 5 different checkpoint loads. No identity drift.
External deps	Zero	No external LLM inference. No API calls. No borrowed weights. Network I/O during inference: none.
Long context	Pass	Extended context windows tested. Big Brain memory layer maintained coherence across long sessions without degradation.

1M suite — historical reference

The 1M parameter suite was run in April 2026. Results were consistent with design expectations and established the first public evidence point. Retention was strong, active cells were bounded, and identity outputs were stable across the full run.

This suite is superseded by the 10M results for all public-facing claims. It is kept here for completeness and to document the absence of regressions between scale steps.

🔒

Demo status: private. The interactive demo is gated behind the Stage 7 quality check. Passing the 10M benchmark is a necessary condition but not sufficient on its own. Stage 7 integration must complete before demo access opens.

Public surface

What is public

Elena AILL follows an evidence-first public posture. Only validated results appear on the public surface. No speculative roadmap, no unvalidated claims, and no demo promises that cannot be immediately delivered.

Currently public

Resource	Location	Status
Main website	research.elenaxvora.com ↗	Live
Benchmark reports	1M and 10M parameter validation results	Public
Architecture overview	Module descriptions, design rationale, constraint list	Public
PNG artifacts	Real benchmark output images from validated runs	Public

Not yet public

Resource	Gate	Status
Interactive demo	Stage 7 quality gate must pass first	Private
Source code	No release timeline set	Private
Model weights	No release timeline set	Private

Positioning

Elena AILL is positioned as a research project. The public surface is designed to communicate research progress and validated evidence. Claims on the website are deliberately conservative: if something has not been validated with a reproducible run, it is not claimed.

The website at elenaxvora.com is the canonical public source of truth. This documentation exists as a more detailed companion for readers who want the full technical context behind what the website says.

Public Status

The public surface is the evidence package: benchmark summaries, real PNG artifacts, phase history, and technical documentation aligned with the repository.

Elena AILL should be described here as a system that reads, writes, handles code, and produces documents and structured outputs.

Complete PNG Inventory

Every real benchmark PNG is listed, previewed, and described in the public package. The inventory covers every real PNG in the benchmark pack plus preserved Elena AILL, GLIA, and SenzaAdam visuals already generated in the repository.

Each PNG stays tied to a stable filename, benchmark stage, and short explanation. Previews are intended to load on demand in the live site; branded PNGs are preferred for public presentation when available, while original exports remain individually accessible.

Complete PNG Inventory

Every real benchmark PNG is listed, previewed, and described.

The inventory covers every real PNG in the public benchmark package plus the preserved Elena AILL, GLIA, and SenzaAdam visuals already generated in the repository.

Each PNG stays tied to a stable filename, benchmark stage, short explanation, and reason it matters, exactly like the original public surface.

Loading PNG inventory…

Previews load on demand. Branded PNGs are preferred for public presentation when available, while original exports remain individually accessible.

Release Scope

The current public scope is documentation, benchmark evidence, and repository-backed artifacts.

Earlier stages validated memory behavior, runtime behavior, coding, writing, and self-documentation. The remaining work is runtime integration and packaging quality, not audio positioning.

Integration

Runtime integration

Focus engineering effort on cleaner native integration around the validated SenzaAdam Big Brain + Clean Speak path instead of forcing a larger benchmark first.

Public pack

Benchmark packaging

Keep the public benchmark package aligned with traceable reports, metrics summaries, and real PNG artifacts only.

Output quality

Writing and code quality

Improve written output quality, code generation, and document consistency without dropping the grounded memory path that earlier validated stages already proved.

Public scope

Evidence-first release

The site should present truthful evidence, reproducible benchmarks, and technical artifacts while describing Elena through the capabilities already documented here: reading, writing, code handling, documents, and structured outputs.

The site remains an evidence-first document for Elena AILL Research, SenzaAdam, and the current Applied AI validation path.

Final Public Note

Elena AILL has moved from isolated experiments to a validated persistent-memory research pipeline. GLIA remains preserved as a memory and routing research track. SenzaAdam Big Brain and Clean Speak are the current active direction. The public website shows real benchmark data, phase history, and PNG artifacts only.

The cleanest summary is this: Elena AILL remains the main research project, and the public site is an evidence layer designed to stay truthful, searchable, and conservative while reflecting what the system actually does today.