DreamForge AI

DreamForge is a research and visualization simulator. It does not measure brains, diagnose conditions, predict dreams, infer psychological meaning, or provide medical advice.

An open-source, deterministic, offline sandbox for simulating explicit, configurable, non-clinical proxies related to sleep regulation, sleep-stage dynamics, normalized neuromodulatory patterns, synthetic-memory graph selection, and structured dream-context features — under active construction toward the milestone plan in MASTER_PROMPT.md.

Status

First execution slice + M1 (complete) and M2 groundwork: deterministic core with an 8-hour (960 × 30 s epoch) offline trace, event sourcing, DQCJ-1 canonical serialization, hash-verified exports, structured dream context/features/score, and the mandatory offline narrative provider with labeled report blocks. See ARCHITECTURE.md, RESEARCH.md, LIMITATIONS.md, and docs/.

Not yet built (by explicit scope control): API, plugins, notebooks, streaming/tool-use adapter features.

Optional: local / OpenAI-compatible narrative provider (opt-in)

The offline mock is always the default. To use a local model via Ollama's OpenAI-compatible endpoint (or any /chat/completions service), construct the adapter explicitly in your own script — nothing is enabled by configuration alone:

from dreamforge.core.providers.narrative import NarrativeRequest
from dreamforge.integrations.openai_compat import OpenAICompatConfig, OpenAICompatProvider
from dreamforge.integrations.transport import UrllibTransport

config = OpenAICompatConfig(
    base_url="http://127.0.0.1:11434/v1",  # Ollama loopback example
    model="llama3.2:3b",
    timeout_seconds=20,
    max_retries=2,
)
provider = OpenAICompatProvider(config, UrllibTransport())
response = provider.generate(request)  # request = validated NarrativeRequest

Guarantees (ADR 0005): allowlisted projection only; strict response schema; per-attempt timeout; bounded retries then fail-closed; errors redacted to status code + response hash; every response labeled generative_interpretation.

An Anthropic-native adapter with the same guarantees ships alongside it (AnthropicCompatProvider — messages API, x-api-key + anthropic-version headers, content-block extraction); both share one vetted bounded-retry/redaction implementation (integrations/retry.py).

Dashboard (optional extras)

".venv/Scripts/python.exe" -m pip install -c constraints.txt streamlit plotly
".venv/Scripts/python.exe" -m streamlit run src/dreamforge/visualization/dashboard.py -- exports/demo_8h

Renders verified exports only; accessibility/theme documentation in DASHBOARD.md.

What it simulates (and what that means)

• Conceptual two-process-inspired sleep-regulation equations (homeostatic "S" pressure + sinusoidal circadian proxy) — mathematical constructs, not physiology.
• A semi-Markov Wake/N1/N2/N3/REM stage process at 30-second resolution — explicitly not PSG scoring.
• Four normalized [0,1] neuromodulatory proxy indices — qualitative synthetic values, never concentrations or measurements.
• Selection over a synthetic directed weighted memory graph — graph selection, not neural replay.
• Deterministic structured features and scores derived only from the above.

All parameters carry evidence grades (assumption, synthetic_demo, …) in docs/scientific_model/claim_registry.yaml. No empirical or clinical claim is made anywhere in this repository.

Quick start (offline demo)

Requires Python 3.11+ (developed on 3.12, Windows). No network egress at runtime; install needs PyPI once.

python -m venv .venv
".venv/Scripts/python.exe" -m pip install -c constraints.txt -e .   # Windows path; use .venv/bin elsewhere
".venv/Scripts/python.exe" -m dreamforge.demo                       # runs examples/configs/demo_8h.json

The demo validates its configuration, runs 960 epochs, verifies stage-transition legality, writes an export under exports/demo_8h/ (events.ndjson + manifest.json), re-imports it, and re-verifies all hashes.

Development checks

".venv/Scripts/python.exe" -m pytest -q                                            # tests
".venv/Scripts/python.exe" -m pytest -q --cov=src/dreamforge --cov-report=term-missing
".venv/Scripts/python.exe" -m ruff check src tests examples                        # lint
".venv/Scripts/python.exe" -m black --check src tests examples                     # format-check
".venv/Scripts/python.exe" -m mypy src/dreamforge/core                             # typecheck

Repository layout

See ARCHITECTURE.md for the diagram and component table.

License

MIT — see LICENSE. Copyright (c) 2026 JToSound.