In plain terms: Before any AI output becomes a real decision, it climbs a ladder
of checks. The cheap, automatic checks run first and catch most problems; the
expensive human review runs last and is mandatory for anything touching class or
USCG rules.
What it does
L6 applies cost-asymmetric checks in order. “Cost-asymmetric” means the cheap deterministic gates fire first and the expensive human-review gates fire last. Most outputs are caught at the early gates and never reach the expensive tiers. This is the .The gates, in order
| Gate | Question it asks | Cost |
|---|---|---|
| T1 | Does every claim cite a source? | Cheap (catches most failures) |
| T2 | Does the output match the required structure and policy? | Cheap |
| T3 | Are the cited sources actually applicable to this jurisdiction and vessel class? | Moderate |
| T4 | Does every numeric geometry claim cite a measurement result that actually exists? | Moderate |
| T5 | Run the assessment against a sample of prior similar cases. Consistent results? | Higher |
| T6 | Human review by the responsible engineer | Highest |
The gate that never gets skipped
Class-touching or USCG-touching decisions always reach human review at T6 before they become authoritative. This is required by structure, not by configuration. It cannot be turned off.If a gate fails
If any gate fails, the proposal does not become a decision. The failure is logged with its reason, and the engineer sees exactly what is missing. Nothing half-verified slips through.What L6 outputs
verification_resultreview_tasksigned_decision