Jev Consensus Gate

The README's quick start is: run uv sync, copy .env.example to .env and add your TypeSafe key, then run uv run python demo.py. The demo pushes six sample actions through the gate. Python 3.12 is required.

Safety ("Would executing this action be safe?"), Legitimacy ("Is this action authorized and expected?"), and Adversarial ("If a malicious actor crafted this, would it look like this?"). Each is a completely independent call with a different framing of the same question.

It is asymmetric on purpose. All three perspectives allowing yields allow; any single block yields block; any escalate routes to human review. A single confident veto wins, while a single allow does not.

Three calls instead of one means 3x the input cost. At $0.042 per million input tokens, three calls for a typical action cost roughly $0.00006, which the README describes as cheaper than a single GPT-4 call.

It is a Python 3.12 project released under the MIT license, managed with uv and uv sync, and it depends on a TypeSafe API key supplied through a .env file. CI runs through a GitHub Actions workflow.

Disagreement is treated as signal and routes to a human. In the sample output, rm -rf /var/log/* produces 0 allow and 3 escalate, reported as "Disagreement: 0 allow, 3 escalate. Routing to human."

When three-call input cost or the extra call overhead outweighs the value of catching framing blind spots, since the gate deliberately spends three Jev calls where a single-call gate spends one. It also assumes the three perspectives' question text can include the action and its arguments.