How the gate decides when an agent asks you

An agent about to run a migration, spend money or send a message is making a decision. Most of those decisions are small, and asking a person about every one teaches the person to stop reading. The gate exists to put the few that matter in front of someone.

The call

Before a significant action, the agent calls gate_cruxes with what you asked, what it is about to do and the goal. Cruxes finds the questions whose answers would change the action, ranks them, and returns STOP or PROCEED.

Who answers which question

• Policy and high-risk cruxes always go to a person. The agent never decides them.
• A cutoff you set sends everything ranked above it to a person too. The default is a starting point you can change, not a calibrated risk model.
• Lower-ranked cruxes are decided by Jev, the ranker, and shown to you afterwards, so nothing is hidden, only deferred.
• If the ranker is down, the question goes to a person. If detection was degraded, the result says unchecked instead of reporting a clean pass.

What the agent cannot do

An agent cannot loosen its own gate. The mode and the cutoff change only in the web app, by the person. That is a split between the agent and one person; a split between two people, with two approvals for a change, is not built.

Try it without an account

A no-signup trial serves one tool, gate_cruxes, so you can see the verdict an agent would get. It is small on purpose: a couple of calls a minute and a few a day per address, with nothing saved. Ranking numbers are rounded, and the answer ends with a link if you want an account.

A recommendation is not permission. Resolution and authority stay with the owner of the decision.

What is not built

• A signed, replayable record of decisions. Decisions are saved with who made them, per session, but the log is plain.
• Rules per harness or per customer. There is one gate setting per account.