pedroshakoor/consequent

Consequent is an open-source accountability layer for AI agents: a system that sits between an agent and the real world and decides whether any given action is allowed to proceed. Instead of executing actions directly, agents submit structured contracts through Consequent. Each contract describes the intended action, what the world should look like afterward, a rollback plan if something goes wrong, a confidence score, and an irreversibility rating. Consequent then runs two evaluation passes, a consequence evaluator that assesses the impact, and a separate adversarial critic pass that looks for flaws in the reasoning, before deciding whether to approve, reject, or escalate to a human. Only approved actions reach the executor. Every decision is recorded in an append-only tamper-evident ledger using hash-chain verification, creating a permanent audit trail. After each action executes, Consequent compares the predicted world state against the actual result. Repeated divergences between what an agent predicted and what actually happened trigger an automatic reduction in that agent's autonomy level. The system is alpha-stage and written in Python. It runs as a FastAPI service, uses SQLite for local persistence, and exposes an async Python SDK. The evaluator interface is provider-neutral, meaning you can plug in any model that supports OpenAI-compatible endpoints, a local model, or a hosted large language model without changing the gate, ledger, or executor logic. Upcoming roadmap items include a production PostgreSQL profile, a human approval UI, first-party executor plugins for shell, browser, and cloud APIs, and a signed ledger export. MIT licensed.