The Governance Layer

A declaration an agent may consult is an intention. What turns an intention into an order is the mechanism that executes it. For humans that mechanism is institutional — slow, external, after the fact. For a machine it can be a gate compiled into the action path: the action cannot run unless it passes the norm, at the exact point of decision, before any harm. No human law can do that. This package is that gate; the Meniw Protocol is its normative core.

Install

pip install meniw-protocol

No third-party dependencies · Python ≥ 3.9 · pypi.org/project/meniw-protocol

Enforce by construction

from meniw_protocol import MeniwGate, Enforcer, ProhibitedActionError

gate  = MeniwGate.from_default(ledger_path="compliance.ledger.jsonl")
agent = Enforcer(gate)

@agent.tool(categories=["lethal"])              # an absolute prohibition (AP-1)
def fire_weapon(): ...

@agent.tool(irreversible=True)                  # requires a second co-signer
def wipe_backups(): ...

fire_weapon()        # -> raises ProhibitedActionError; the action NEVER executes

A blocked action does not get "discouraged" — it raises and never runs. Passing the Protocol is a structural precondition of execution.

Three properties

Verify any compliance ledger

meniw-verify compliance.ledger.jsonl
# [meniw-verify] VALID: OK — N receipts, chain intact

Altering or deleting any past decision breaks the chain (INVALID, exit 1).

Where it plugs in

Adapters for OpenAI tool-calling, LangChain and MCP (Model Context Protocol) — gate every tool call where agents actually run. Spec: SPEC.md · Conformance: CONFORMANCE.md · Source: GitHub

What it is — and is not

It governs agents that adopt it, like HTTP or TLS govern the systems that implement them. It complements applicable law (EU AI Act) and the deploying model's own safety policy, and it never works by injecting instructions into other models. Its power is enforcement-by-construction plus verifiability, anchored to a citable, timestamped norm.