In plain terms: Here is the whole life of one decision, start to finish.
Suppose a yard is reviewing a fuel system layout for a new build, and wants to
know whether it meets the current Coast Guard fuel-line clearance requirements.
The walkthrough
Trigger
The reviewing engineer opens the fuel system design and asks Forge for a
“rule applicability assessment.”
The system gathers what it needs
Forge pulls the relevant USCG sections, the current fuel-line geometry (as
actual measurements, not guesses), the prior similar systems on this vessel
class, and the yard’s own internal standards.
It drafts a proposal
The AI drafts an assessment: which USCG sections apply, which measurements
support the clearance claim, and which prior decisions on similar systems it
is relying on as precedent.
Confidence check
If the AI is not confident enough (for example, it cannot find a measurement
for a critical clearance), it stops and asks the engineer to supply the
missing data. It does not guess. This is working as intended.
The verification ladder runs, cheap checks first
The proposal then passes through an ordered set of checks. The cheap ones run
first and catch most problems; the expensive ones run last.
in action:
| Gate | Question it asks | Cost |
|---|---|---|
| T1 | Does every claim cite a source? | Cheap |
| T2 | Does the output match the required structure and policy? | Cheap |
| T3 | Are the cited sources actually applicable to this vessel and jurisdiction? | Moderate |
| T4 | Does every numeric clearance claim cite a measurement that really exists? | Moderate |
| T5 | Run it against a sample of past similar cases. Consistent results? | Higher |
| T6 | Human review by the responsible engineer | Highest |
Decision signed
Once the human approves at T6, the assessment becomes a recorded decision
with an , and the full reasoning behind it is preserved.