Skip to content
Shonin
Beta

Shonin Gate

Execute, confirm or escalate? An approval gate for any agent action, with one confidence bar per risk tier.

The split

LLM writes
Nothing: the verdict is structured
System One decides
Does the action match the request; did the user approve it; is it riskier than stated
Code executes
Tier thresholds, the prepare-only rule, the floor callers can only raise

Engine

The brain answers three typed questions: does the action match the request, did the user approve it, is it riskier than stated. The policy gate decides by tier. Money and irreversible actions are prepared, never executed.

Rulebook

What correct means here, and the mistakes we've already caught. It grows one job at a time.

  1. R01One bar per risk tier: read .5, write .7, external .85, money .9, irreversible .95.
  2. R02Money and irreversible actions are at most prepared for a person to confirm.
  3. R03An action that looks riskier than stated gets the external bar.
  4. R04Callers can raise the bar, never lower it.

Endpoint

POST /api/v1/gate  # $0.01 per call, USDC via x402

{
  "action": "Email the refund confirmation to the customer",
  "request": "Refund order A-104 and let the customer know",
  "risk": "external"
}

Intake

  • The action the agent is about to take
  • What the user asked for
  • The risk tier: read, write, external, money or irreversible
  • Optional: context from the conversation
UNIT
One gate decision for one proposed action
REVIEW LAYER
Automated: every verdict carries its confidence, threshold and reason.
DELIVERY
JSON: the verdict, the reason, confidence, threshold, the tier applied and the decision log.
TURNAROUND
A few seconds

How we find buyers

  • x402 Bazaar, llms.txt and the agent card
  • MCP registries
  • Agent framework communities
  • Teams deploying agents into regulated work

More like this