Precedent: human-approved operating rules and deterministic action gates for AI agents

Precedent turns how a team already works into operating rules that AI agents follow. It mines question-and-answer threads from team chat (WhatsApp, Slack, JSONL) into draft rules grounded in the source messages, requires a person to approve every change, and gates each agent action through deterministic checks outside the model (coverage, judgement, grounding, risk-tiered confidence, risk), returning act, needs-approval or escalate. A second, execution gate re-checks each step and the result after the action is done, and halts a running action that goes beyond its decision, recording which rule, which run and who halted it. Every escalation becomes a draft rule and a regression test; drift detection flags stale, overridden, unused and duplicate rules. It runs as a Python library, CLI, MCP server, or compiles to Agent Skills and AGENTS.md.