An Agent-Native human social behavior prediction engine built on CAS theory, leveraging population-level simulation and a four-agent deliberative architecture for efficient, traceable reasoning.
Ripple is an Agent-Native general-purpose human social behavior prediction engine built on Complex Adaptive System (CAS) theory. Its core innovation lies in encoding CAS mechanisms—emergence, nonlinear feedback, phase transitions—directly as engine primitives (Ripple / Event / Field / PhaseVector / Meme), and replacing per-person simulation with population-level statistical distributions. This reduces LLM calls from ~300,000 to ~100–500 per simulation, achieving roughly three orders of magnitude cost reduction.
The engine employs an original "Star-Sea Deliberative Architecture" with four agent types: the Omniscient (global orchestration and propagation adjudication, high-intelligence model), Stars (KOL individual-level simulation, high-quality model), Sea (ordinary user population-level simulation, lightweight model), and the Deliberative Tribunal (multi-expert structured debate to systematically counter LLM optimism bias). A five-layer anti-optimism defense and a calibration pipeline of independent review → cross-examination → stance revision → synthesized ruling ensure prediction reliability.
Execution follows a 5-Phase Wave cycle (INIT → SEED → RIPPLE → OBSERVE → FEEDBACK & RECORD), with an optional DELIBERATE phase. Every wave ruling, agent decision, propagation path, and tribunal debate is recorded as structured JSON, with confidence assessments attached to predictions.
Domain extension uses a Skill architecture where the core engine remains domain-agnostic; domain knowledge is injected via pure natural-language Skill packs (domain profile + platform profile + role prompts) with zero code. Currently validated scenarios include social media content propagation prediction (7 platforms: Xiaohongshu, Douyin, Weibo, etc.) and PMF validation (supporting free Channel × Vertical × Platform combinations, e.g., algorithm-ecommerce × FMCG × Douyin, search-ecommerce × consumer-electronics × Xiaohongshu, enterprise-sales × SaaS).
The project has no third-party Agent framework dependencies, implemented in pure Python 3.11+ with httpx, supporting Anthropic / OpenAI-compatible / AWS Bedrock / Volcengine and other LLM backends. It provides one-click Docker deployment, CLI tools, a Python SDK, and an HTTP+SSE service API (with real-time progress streaming), plus OpenClaw integration.
Unconfirmed items: OASIS paper reference not explicitly cited; cost reduction claims are self-reported without independent third-party benchmarking; author/team background unknown; publication status unknown; project is in early stage (35 commits, v0.2.6) with no public roadmap; AGPL-3.0 license—network service compliance requirements should be evaluated by users.