Risk policy

Deterministic intent scoring - the same code runs server-side and here, so every decision is reproducible.

Intent scoring policy v1

deterministic and open

Human approval threshold

score ≥ 40/100 escalates to the Approvals page

Repeat spend today

+10 points

Unknown vendor

+15 points

Amount bands

≤5$: +0 ≤20$: +10 ≤∞$: +25

Category baseline weights

saas: 10ai: 15compute: 20agent: 30rwa: 45marketplace: 25

The on-chain fence (per-tx caps, daily caps, expiry, vendor/token allowlists) is enforced by the vault contract no matter what this score says. Scoring only decides who confirms: the machine, or a human.