Policy driven AI middleware and agent system with deterministic routing structured tool orchestration SPO based fact extraction crosssource conflict detection agreement scoring bounded reconciliation loops memory guided strategy bias and execution aware planning transforming LLM outputs into a controlled self correcting verifiable reasoning and decision engine across dynamic real world tasks
user query → policy decision engine → tool routing → multi-source orchestration → structured extraction (SPO + numerics + entities) → conflict detection (numeric + semantic) → evaluation + reconciliation controller → memory guided synthesis → final answer
| Feature | Traditional RAG + Embeddings | Claude Code (Opus 4.8) | Neura v0.14 |
|---|---|---|---|
| 🧠 Long-term memory | ❌ Usually external vector DB | ✅ Persistent semantic memory graph | |
| 📂 Repository understanding | 📄 Stores code chunks | 📖 Reads repository when needed | Builds an architectural understanding of the repository |
| 🔍 Retrieval | Embedding similarity | Context + search + tools | Concept-first retrieval + graph expansion + embeddings |
| 🕸️ Knowledge structure | Independent chunks | Mostly implicit inside model context | Typed knowledge graph (Supports, DependsOn, PartOf, etc.) |
| 📈 Relationships | Similarity only | Inferred during reasoning | Explicit, persistent semantic relationships |
| 🧩 Architecture awareness | Low | High during current session | Persistent architectural model |
| 📚 Engineering history | None | Conversation history | Decisions, evidence, verification, evolution history |
| 🧪 Impact analysis | None | Model estimates likely impact | Graph-based dependency analysis with uncertainty estimates |
| 🎯 Planning | Prompt-based | Strong session planning | Persistent engineering plans connected to knowledge graph |
| 🔧 Tool integration | Usually external orchestration | Excellent integrated tool use | Tools operate through semantic knowledge and feed back into it |
| ✅ Verification | User initiated | Can run builds/tests | Verification becomes permanent engineering evidence |
| 📖 Explainability | Retrieved chunks | Natural-language explanation | Graph path + evidence chain + confidence + reasoning trace |
| 🔄 Learning after execution | Usually none | Mostly session-local | Reflections become evidence that improves future reasoning |
| 💤 Offline maintenance | None | None | Sleep consolidates concepts, refreshes embeddings, updates confidence |
| 🌳 Knowledge evolution | Static until re-indexed | Rebuild context next session | Continuously evolves as repositories and projects change |
| 📊 Confidence model | Retrieval score | Model confidence (implicit) | Evidence-backed confidence for concepts and relationships |
| 🏛️ Architectural decisions | None | Conversation only | First-class persistent engineering decisions |
| 📦 Multiple repositories | Separate indexes | Open multiple projects | Unified semantic graph spanning repositories |
| 🔄 Repository updates | Re-index changed files | Reads changed files | Incremental concept refresh with graph updates |
| 🤝 Cross-source reasoning | Difficult | Possible via prompt | Native across repositories, docs, conversations, evidence, and history |
| 📜 Engineering evidence | None | Conversation artifacts | Append-only evidence ledger with validation and reflection |
| 🎯 Prediction | None | Can estimate during reasoning | Predicts impact, compares with reality, calibrates over time |
| ❤️ Project understanding over months | Minimal | Depends on available context | Designed to continuously improve as the project evolves |
| 🧭 Primary philosophy | "Find relevant text." | "Read, reason, and use tools." | "Continuously build and refine an internal understanding of the software." |
gpt-oss-20b (patched llama-server, n-cpu-moe) companion 1.5B (jlens, PyTorch)
│ residual @ layer L (real) │ residual @ layer L
▼ ▼
in-engine LOGIT LENS ──────────┐ JACOBIAN LENS (future dir)
(current belief, entropy) │ │
▼ ▼
┌────── STATE ESTIMATOR (cogstate.py) ──────┐
│ EMA/Kalman-lite smoothing over layers │
└──────────────────┬──────────────────────────┘
▼
TRAJECTORY ANALYSIS
entropy · layer-agreement(KL) · convergence · hypothesis tracks ·
current-vs-future divergence (uncertainty)
▼
LIVE UI (dashboard) + ADAPTIVE ORCHESTRATION
token-by-token cognition panel (soft triggers: verify / retrieve /
tool / early-stop — opt-in, gated)
Unlike traditional agent architectures that treat the language model as a black box, Neura continuously observes, estimates, and reasons about the model's evolving cognitive state.
flowchart TB
%% =====================================================
%% INPUT
%% =====================================================
U([User])
API[Neura Runtime]
MEM[(Adaptive Memory<br/>Knowledge Graph<br/>Repair Learning)]
CTX[Context Compiler<br/>Rehydration Engine]
U --> API
MEM --> CTX
CTX --> API
%% =====================================================
%% MODEL ORCHESTRATION
%% =====================================================
subgraph MODELS["Model Layer"]
direction LR
LOCAL["Neura-OSS-20B<br/>Patched llama.cpp"]
REMOTE["GPT-5.5 / Frontier Model"]
COMP["Companion 1.5B<br/>Jacobian Lens"]
end
API --> LOCAL
API --> REMOTE
%% =====================================================
%% RESIDUAL STREAM
%% =====================================================
subgraph RESIDUAL["Residual Stream Observer"]
direction TB
L0["Embedding"]
L1["Layer 1"]
L2["Layer 2"]
L3["..."]
LN["Final Layer"]
L0 --> L1 --> L2 --> L3 --> LN
end
LOCAL --> RESIDUAL
%% =====================================================
%% LOGIT LENS
%% =====================================================
subgraph LOGLENS["Native Logit Lens"]
direction TB
LOGITS["Current Belief"]
ENT["Entropy"]
TOPK["Top-K Tokens"]
PROBS["Probability Distribution"]
LOGITS --> ENT
LOGITS --> TOPK
LOGITS --> PROBS
end
%% =====================================================
%% JLENS
%% =====================================================
subgraph JLENS["Companion Jacobian Lens"]
direction TB
JRES["Residual Snapshot"]
JPRED["Future Direction"]
JCONF["Future Confidence"]
JRES --> JPRED
JPRED --> JCONF
end
RESIDUAL --> LOGLENS
RESIDUAL --> JRES
COMP --> JLENS
%% =====================================================
%% TOKEN GRAPH
%% =====================================================
subgraph TOKENGRAPH["Hierarchical Token Graph"]
direction TB
TOK["Token Nodes"]
EMB["Embedding Nodes"]
CON["Concept Nodes"]
HYP["Hypothesis Nodes"]
REL["Semantic Relationships"]
TOK --> EMB
EMB --> CON
CON --> HYP
HYP --> REL
end
LOGLENS --> TOKENGRAPH
JLENS --> TOKENGRAPH
%% =====================================================
%% GRAPH ANALYTICS
%% =====================================================
subgraph GRAPHOPS["Graph Analytics"]
direction TB
CENT["Centrality"]
COMM["Communities"]
PATH["Shortest Paths"]
PERSIST["Temporal Persistence"]
GRAPHENT["Graph Entropy"]
CENT --> GRAPHENT
COMM --> GRAPHENT
PATH --> GRAPHENT
PERSIST --> GRAPHENT
end
TOKENGRAPH --> GRAPHOPS
%% =====================================================
%% TRAJECTORY
%% =====================================================
subgraph TRAJ["Trajectory Engine"]
direction TB
VEL["Velocity"]
ACC["Acceleration"]
CURV["Curvature"]
DRIFT["Semantic Drift"]
ATTR["Attractors"]
VEL --> ACC
ACC --> CURV
CURV --> DRIFT
DRIFT --> ATTR
end
RESIDUAL --> TRAJ
%% =====================================================
%% COGNITIVE STATE
%% =====================================================
subgraph COGSTATE["Cognitive State Estimator"]
direction TB
OBS["Observation Fusion"]
EMA["EMA"]
KAL["Kalman-lite"]
BAYES["Bayesian Updates"]
POST["Posterior State"]
OBS --> EMA
OBS --> KAL
OBS --> BAYES
EMA --> POST
KAL --> POST
BAYES --> POST
end
LOGLENS --> OBS
JLENS --> OBS
GRAPHOPS --> OBS
TRAJ --> OBS
%% =====================================================
%% METRICS
%% =====================================================
subgraph METRICS["Inference Metrics"]
direction TB
CONF["Confidence"]
UNC["Uncertainty"]
CONV["Convergence"]
KL["Layer KL"]
AGREE["Lens Agreement"]
STAB["Belief Stability"]
ENT2["Entropy"]
POST --> CONF
POST --> UNC
POST --> CONV
POST --> KL
POST --> AGREE
POST --> STAB
POST --> ENT2
end
%% =====================================================
%% HYPOTHESIS TRACKING
%% =====================================================
subgraph HYPTRACK["Hypothesis Evolution"]
direction TB
BIRTH["Birth"]
GROW["Growth"]
MERGE["Merge"]
SPLIT["Split"]
DECAY["Decay"]
FINAL["Final Decision"]
BIRTH --> GROW
GROW --> MERGE
MERGE --> SPLIT
SPLIT --> DECAY
DECAY --> FINAL
end
POST --> HYPTRACK
%% =====================================================
%% ORCHESTRATION
%% =====================================================
subgraph ORCH["Adaptive Orchestration"]
direction TB
VERIFY["Verification"]
RETRIEVE["Memory Retrieval"]
TOOLS["Tool Invocation"]
REROUTE["Model Routing"]
EARLY["Early Exit"]
MORE["Request More Reasoning"]
VERIFY
RETRIEVE
TOOLS
REROUTE
EARLY
MORE
end
CONF --> ORCH
UNC --> ORCH
CONV --> ORCH
AGREE --> ORCH
%% =====================================================
%% MEMORY
%% =====================================================
subgraph MEMORY["Persistent Cognitive Memory"]
direction TB
STATE["World State"]
USERMEM["User Memory"]
REPO["Repository State"]
REPAIR["Repair Motifs"]
GRAPHMEM["Knowledge Graph"]
TIMELINE["Reasoning Timeline"]
STATE --> GRAPHMEM
USERMEM --> GRAPHMEM
REPO --> GRAPHMEM
REPAIR --> GRAPHMEM
TIMELINE --> GRAPHMEM
end
ORCH --> MEMORY
HYPTRACK --> MEMORY
TOKENGRAPH --> MEMORY
MEMORY --> CTX
%% =====================================================
%% OUTPUT
%% =====================================================
subgraph OUTPUT["Presentation Layer"]
direction TB
UI["Live Cognition Dashboard"]
HEAT["Layer Heatmaps"]
GRAPHVIEW["Semantic Graph"]
TIMELINE2["Reasoning Timeline"]
TOKVIEW["Token Evolution"]
METRICVIEW["Confidence / Entropy"]
UI --> HEAT
UI --> GRAPHVIEW
UI --> TIMELINE2
UI --> TOKVIEW
UI --> METRICVIEW
end
POST --> OUTPUT
%% =====================================================
%% FINAL RESPONSE
%% =====================================================
REMOTE --> RESP([Final Response])
LOCAL --> RESP
ORCH --> RESP
RESP --> U
