AWE -- Artificial Wisdom Engine

The AI industry built intelligence. It forgot to build judgment.

AWE is a Rust engine that transmutes raw LLM intelligence into calibrated wisdom -- consequence-aware, confidence-tracked, and compounding with every decision. It is model-agnostic, privacy-first, and designed to get wiser the more you use it.

The Transmutation Pipeline

Intelligence In --> Wisdom Out

  Receive --> Interrogate --> Calibrate --> Consequent --> Synthesize --> Judge --> Articulate
     |             |              |             |              |           |           |
  Parse the    Surface        Assign       Model 1st/      Connect     Weigh      Deliver
  decision     hidden       calibrated     2nd/3rd order   cross-     trade-offs  articulated
  context      assumptions  confidence     consequences    domain     and timing  judgment
                                                           patterns

Requirements

AWE requires an LLM to deliver accurate wisdom. The 7-stage pipeline uses an LLM for contextual analysis, consequence modeling, precedent synthesis, and calibrated judgment. Without one, AWE runs in limited algorithmic mode with generic heuristics.

Provider	Setup	Quality
Anthropic Claude	export ANTHROPIC_API_KEY=sk-ant-...	Best — full contextual wisdom
Ollama (local)	export AWE_OLLAMA_MODEL=llama3.2	Good — free, private, offline
None	—	Limited — generic heuristics only

Quick Start

cargo install --path crates/awe-cli

# Configure your LLM (required for accurate wisdom)
export ANTHROPIC_API_KEY=sk-ant-...

# Transmute a decision
awe transmute "Should we rewrite our monolith as microservices?"

# Transmute with domain context
awe transmute "Should we scale to Kubernetes?" --domain infrastructure

# Feed back outcomes (closes the learning loop)
awe feedback dc_01JQ... --outcome confirmed --details "Strangler approach worked"

# View calibration accuracy for a domain
awe calibration --domain software-engineering

Architecture

AWE is a Rust workspace with 12 crates -- 11 libraries composing the wisdom pipeline, plus the CLI.

crates/
  awe-core/          Foundation types, traits, error handling
                     The periodic table of wisdom elements (Dc, Cq, Pr, As, Cf, Tf, Pv, Cx, Fb, Ap)

  awe-epistemic/     Layer 1: Confidence calibration, knowledge boundaries, contradiction detection
  awe-consequence/   Layer 2: Consequence chain modeling, reversibility scoring
  awe-memory/        Layer 3: Wisdom Graph persistence (SQLite), pattern extraction
  awe-judgment/      Layer 4: Trade-off mapping, timing intelligence, values reasoning
  awe-synthesis/     Layer 5: Cross-domain analogy, meta-pattern recognition
  awe-socratic/      Layer 6: Assumption surfacing, question reformulation, bias detection

  awe-engine/        The complete 7-stage transmutation pipeline orchestrator
  awe-protocol/      Wire format, message types, versioning
  awe-cli/           CLI binary (`awe`)

Architecture Diagram

flowchart TB
    subgraph "Context Assembly"
        A["4DA developer_dna<br/>(stack, concerns, gaps)"] --> C[identity.json]
        B["System Probe<br/>OS / CPU / RAM / Models"] --> C
        D["Wisdom Graph Stats<br/>decisions, feedback %"] --> C
        C --> E[context.json]
    end

    subgraph "MCP Server"
        F[awe_transmute] --> G{"buildContextFile()"}
        G --> E
        F --> H["AWE CLI Binary<br/>(awe transmute --context_file)"]
        I[awe_set_identity] --> C
        J[awe_scan_all] --> H
        K[awe_feedback] --> H
    end

    subgraph "7-Stage Pipeline"
        H --> L["1. Receive<br/>Parse query + detect domain"]
        L --> M["2. Interrogate<br/>Bias detection + assumptions"]
        M --> N["3. Calibrate<br/>Brier score + confidence"]
        N --> O["4. Consequent<br/>DAG + reversibility scoring"]
        O --> P["5. Synthesize<br/>Precedents + principles"]
        P --> Q["6. Judge<br/>Trade-offs + timing"]
        Q --> R["7. Articulate<br/>Final wisdom output"]
    end

    subgraph "Compounding Loop"
        R --> S[("Wisdom Graph<br/>(SQLite)")]
        S --> T["Principle Extraction<br/>7+ evidence threshold"]
        T --> S
        K --> S
        S --> N
        S --> P
    end

    subgraph "Data Harvest"
        U[Git Repos] --> V["awe scan<br/>Decision Detection"]
        V --> S
        W["awe scan --infer<br/>Outcome Inference"] --> S
    end

Loading

Data Flow

1. Context injection: 4DA surfaces developer DNA (primary stack, domain concerns, knowledge gaps) which is written to identity.json. The MCP server's buildContextFile() merges this with system probe data (OS, CPU, RAM, local models) and Wisdom Graph statistics (decision count, feedback coverage) into context.json. This file is passed to the AWE CLI via --context_file, where it deserializes into DeveloperContext and personalizes every pipeline stage.

2. Compounding loop: Each transmutation produces a Decision stored in the Wisdom Graph. When the user provides feedback (awe feedback), outcomes close the learning loop. The principle extraction system (awe-memory) continuously analyzes decision-outcome pairs, promoting patterns with 7+ evidence pairs to validated Principles. These principles feed back into the Calibrate and Synthesize stages, making future transmutations wiser.

3. Auto-detection: awe scan walks git repositories, detecting architectural decisions from commit messages and dependency changes. awe scan --infer checks whether those decisions persisted (dependency still in manifest? branch merged or abandoned?) and auto-generates feedback. This closes the compounding loop without requiring manual input.

Wisdom Elements

Every transmutation produces structured elements, not prose:

Symbol	Element	Description
Dc	Decision	A choice point -- action taken, deferred, or rejected
Cq	Consequence	An effect chain (1st/2nd/3rd order)
Pr	Principle	A pattern extracted from 3+ decision-outcome pairs
As	Assumption	Something treated as true without verification
Cf	Confidence	Calibrated probability tracked against reality
Tf	Trade-off	Two or more values in tension
Pv	Perspective	A stakeholder viewpoint changing what "good" means
Cx	Context	Temporal, situational, cultural factors
Fb	Feedback	Outcome data that closes the learning loop
Ap	Anti-pattern	A decision pattern that reliably produces bad outcomes

Key Properties

Model-agnostic -- works with Claude, GPT, Ollama, or no LLM at all. Without an LLM, AWE runs in pure algorithmic mode using pattern matching and heuristics against the Wisdom Graph.

Privacy-first -- BYOK (Bring Your Own Key). All data stays local. The Wisdom Graph lives on your machine. API keys are never stored remotely.

Compounding -- AWE gets wiser with use. Every decision and its outcome feed the Wisdom Graph. Principles require 3+ evidence pairs. Anti-patterns require 3+ failure cases. Confidence is tracked against reality, not declared.

Configuration

Set your API key via environment variable:

export ANTHROPIC_API_KEY=sk-ant-...

Or configure via ~/.awe/config.toml:

[llm]
provider = "Anthropic"
model = "claude-sonnet-4-20250514"

[engine]
max_consequence_order = 3
socratic_by_default = true
min_calibration_samples = 10

Supported providers: Anthropic, OpenAi, Ollama, Custom.

For Ollama (fully local, no API key needed):

[llm]
provider = "Ollama"
model = "llama3"
base_url = "http://localhost:11434"

Development

cargo build                        # Build all crates
cargo test                         # Run all tests
cargo test -p awe-core             # Test a specific crate
cargo clippy --workspace           # Lint
cargo doc --workspace --no-deps    # Generate docs

Minimum supported Rust version: 1.85 (edition 2024).

Invariants

These are non-negotiable properties of AWE:

1. Every Confidence value is tracked against outcomes -- no decorative confidence
2. Principles require 3+ evidence pairs -- no premature generalization
3. Anti-patterns require 3+ failure cases -- no single-incident overreaction
4. The Wisdom Graph is append-mostly -- decisions and feedback are never deleted
5. AWE is model-agnostic -- works with any LLM or no LLM
6. All data stays local -- BYOK for any API keys

Specifications

• spec/ontology.md -- Canonical element definitions
• spec/protocol.md -- AWE Protocol specification

License

FSL-1.1-Apache-2.0

Copyright 2026 4DA Systems Pty Ltd. Source-available under the Functional Source License. Converts to Apache 2.0 after two years.