Dream-RSI 3.3.1

Dream-RSI is a bounded research platform for replay-based recursive exploration-policy improvement, autonomous experiment planning, coding-task evaluation, repository repair, persistent research memory, and reproducible multi-model experiments.

What is included

• Discovery Tree and replay-based research loop
• Independent fresh-world validation
• Structured LLM policy/research planning with JSON contracts
• Ollama, OpenAI-compatible, and Gemini providers
• Parallel experiment orchestration with deterministic result ordering
• Repeated-seed and multi-model research matrices
• Statistical aggregation and paired score deltas
• Failure diagnosis and bounded strategy mutation
• Failure-driven benchmark task generation
• Persistent SQLite research memory
• Repository-level patching, diagnostics, and repair
• Docker-first execution boundary with explicit resource/security policy
• Production FastAPI API with API-key authentication
• Redis Streams queue with retries, stale-message reclamation, and DLQ
• PostgreSQL or SQLite persistence
• Kubernetes and Helm deployment manifests
• Health/readiness endpoints and Prometheus metrics
• Atomic autonomous-run checkpoints
• Automated regression suite

Release status

This archive is a single clean release: 3.3.1. Historical versioned implementation files and runtime database artifacts are intentionally excluded from the release archive.

The project is production-oriented, but high-assurance execution of hostile model-generated code still requires an external microVM/gVisor worker boundary. The application does not pretend that an in-process integration stub is a security boundary.

Quick start

python -m venv .venv
# Linux/macOS
source .venv/bin/activate
# Windows PowerShell: .venv\\Scripts\\Activate.ps1

python -m pip install -e '.[dev]'
python -m pytest -q
python -m dream_rsi.cli.run_demo

LLM coding research

Local Ollama:

python -m pip install -e '.[dev]'
python -m dream_rsi.cli.llm_coding \
  --provider ollama \
  --model qwen3.5:9b \
  --name addition \
  --prompt "Read two integers from stdin and print their sum." \
  --case "2 3" "5" \
  --case "10 7" "17" \
  --iterations 5 \
  --output artifacts/llm-addition.json

Remote providers require their corresponding credentials in the environment.

Autonomous research

dream-rsi-autonomous \
  --provider ollama \
  --model qwen3.5:9b \
  --task sum-two \
  --cycles 3 \
  --memory artifacts/research.sqlite3 \
  --output artifacts/autonomous-report.json

The autonomous loop is bounded by an explicit cycle count and does not recursively execute unrestricted LLM output.

Multi-model research matrix

The matrix CLI is fully executable rather than a placeholder:

dream-rsi-research-matrix \
  --provider ollama \
  --models qwen3.5:9b,qwen3.5:9b \
  --tasks sum-two,max-list \
  --seeds 0,1 \
  --iterations 2 \
  --workers 2 \
  --output artifacts/research-matrix.json

Use distinct model names in --models when comparing providers/models. The output contains trials, descriptive aggregates, and paired score deltas; it does not embed a model ranking.

Production deployment

cp .env.example .env
# Set strong, unique values for every production secret.
docker compose -f docker-compose.production.yml up --build -d

Production mode requires API authentication and disables local execution fallback by default. Generated/untrusted code must run in the Docker boundary; for high-assurance deployments use a separately deployed gVisor/Firecracker worker.

Security model

The default execution policy denies network access, uses a read-only root filesystem, drops Linux capabilities, enables no-new-privileges, and limits CPU, memory, PID count, runtime, and output. Local subprocess fallback is development-only and is automatically disabled when DREAM_RSI_ENV=production unless an operator explicitly overrides it.

The Firecracker integration is deliberately an RPC boundary, not a fake in-process VM. A production high-assurance deployment should place privileged microVM lifecycle management in a separate worker/service.

Testing and release verification

make test
make check-release
python -m compileall -q dream_rsi

A release is considered clean only when tests, compilation, CLI import/smoke checks, version consistency, and archive hygiene all pass.

Research interpretation

The bundled toy and coding benchmarks validate the framework and experimental protocol. They are not evidence of general intelligence, AGI, or superiority on real-world software engineering benchmarks. Real task adapters and independently held-out evaluation environments should be supplied for substantive research claims.