Int Crucible

A general multi-agent reasoning system built on world-model construction, candidate generation, stress-testing, and the intelligence metric I = P / R.

⸻

Overview

Int Crucible is a domain-agnostic reasoning engine. It builds structured world models, generates solution candidates, stress-tests them with scenario suites, evaluates them through constraint-weighted scoring, and ranks them using the intelligence metric:

Intelligence = Prediction / Resources (I = P / R)

Unlike research-only systems, Int Crucible is built for any domain: • policy / constitutions • engineering • algorithms • system design • theoretical reasoning • creative or conceptual exploration

The system is fully modular, transparent, and designed for multi-turn refinement, lineage tracking, and human feedback.

⸻

High-Level Architecture

Int Crucible uses a multi-agent pipeline: 1. ProblemSpec Agent Parses user intent into constraints (with weights), goals, and required resolution. 2. WorldModeller Builds a structured world model at appropriate fidelity. Later versions may include research and knowledge-graph enrichment. 3. Designers Generate diverse solution candidates from the world model. 4. ScenarioGenerator Produces synthetic test cases that probe fragile assumptions and constraint boundaries. 5. Evaluators Run scenario reasoning, compute P (prediction quality), R (resource/complexity cost), and constraint satisfaction. 6. I-Ranker Computes I = P / R, applies weighted constraints, and produces a ranked candidate list. 7. Provenance Tracker Maintains lineage, parentage, origin tags, and full transformation history. 8. FeedbackAgent Allows human users to flag issues, diagnose model errors or missing constraints, and trigger selective re-runs.

The system is designed to become self-improving through pipeline-level introspection (future phase).

⸻

Repository Layout

The repo is structured to allow incremental development and clean isolation between Crucible components and external dependencies:

/core/ # Crucible agents (ProblemSpec, Modeller, Designers, Evaluators, Ranker) /world_model/ # World model schema, nodes, templates /design/ # Candidate-generation logic /evaluation/ # Scenario generation, test runners, scoring /ui/ # Optional UI or CLI wrappers /meta/ # System-level evaluation, logs, self-improvement /vendor/kosmos/ # Kosmos code imported via git subtree (infrastructure only)

The Kosmos subtree provides long-horizon orchestration, knowledge-graph tooling, and sandbox execution, but all science-specific agents and pipelines will be replaced or removed.

⸻

Setup

This project uses the Cursor Project Bootstrapper: https://github.com/copperdogma/cursor-project-bootstrapper

After cloning:

./bootstrap.sh

Then add the Kosmos subtree:

git remote add kosmos https://github.com/jimmc414/Kosmos.git git subtree add --prefix=vendor/kosmos kosmos main --squash

Backend Setup

The Int Crucible backend integrates with Kosmos for agent orchestration and infrastructure.

Quick Setup:

# Automated setup (recommended)
./setup_backend.sh

# Or manual setup:
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -e vendor/kosmos  # Install Kosmos first
pip install -e .  # Install Int Crucible backend
cp .env.example .env  # Edit with your configuration

Configuration

Edit .env file with your settings:

• DATABASE_URL: Database connection string (default: sqlite:///crucible.db)
• LOG_LEVEL: Logging verbosity (default: INFO)
• API_HOST / API_PORT: API server settings
• LLM Provider settings (see .env.example for options):
  • Anthropic Claude: LLM_PROVIDER=anthropic, ANTHROPIC_API_KEY=...
  • OpenAI: LLM_PROVIDER=openai, OPENAI_API_KEY=...
  • Local (Ollama): LLM_PROVIDER=openai, OPENAI_BASE_URL=http://localhost:11434/v1

Testing the Integration

Quick Verification (Recommended):

# Run the automated verification script (checks installation)
./verify_setup.sh

# Then start the server (in a separate terminal or after verification)
./start_server.sh

The verification script checks that everything is installed correctly. The start script actually starts the API server.

Manual Testing:

# Activate virtual environment
source venv/bin/activate

# Test Kosmos integration
crucible kosmos-test

# List available Kosmos agents
crucible kosmos-agents

# Start the API server
python -m crucible.api.main
# Or use uvicorn directly:
uvicorn crucible.api.main:app --reload

The API will be available at http://127.0.0.1:8000 with:

• GET / - Root endpoint
• GET /health - Health check
• GET /kosmos/agents - List available Kosmos agents
• POST /kosmos/test - Test Kosmos integration

API documentation is available at http://127.0.0.1:8000/docs when the server is running.

Need Help? See VERIFICATION_GUIDE.md for detailed step-by-step instructions.

⸻

Development Goals (MVP → Advanced)

MVP • Structured ProblemSpec → WorldModel → Designers → Evaluators → Ranking • No external research; world models built from prompt context • Synthetic scenario generation • Lineage tracking for all candidates • Eval-only mode for user-supplied ideas • Seeded search mode for refining user ideas

Phase 2 • Research-enhanced world models (web, corpora, APIs) • World model stored in a typed knowledge graph • Real simulation/test execution where applicable

Phase 3 • Designer ↔ Modeller kickback loops (challenge assumptions) • Scenario test suite inheritance across runs • High-resolution constraint modeling • Implementer modeling (actors + hard constraints)

Phase 4 • SystemEvaluator agent analyzes logs and proposes pipeline improvements • Self-refining world model schemas • Multi-run architecture evolution guided by I=P/R at system level

⸻

Why “Int Crucible”?

A crucible is a vessel for refining raw material under heat and stress. This system does the same for ideas: generate → test → break → refine → converge.

The “Int” prefix nods both to intelligence and to the INT ability score from D&D—appropriate for a system whose core metric is I=P/R.

⸻

Status

Early development. Architecture is complete; infrastructure integration and agent implementations are under construction.

⸻

License

To be determined.