The system works, but there are still real limits to what it can do safely and reliably.

Resolver decisions are still prompt-dependent

The job-record resolver is structured and constrained, but it is still driven by model output.

That means the quality of downstream create, update, or skip decisions still depends on how clearly the email expresses the role, company, and stage in the process.

Ambiguous job matching is still a hard problem

If multiple roles exist for the same company and the email does not clearly identify which one it refers to, the system can still run into ambiguity.

That is especially true when the email only talks about a process step and does not mention the actual role clearly enough.

The system depends heavily on external services

This workflow depends on:

If one of those systems is unavailable or misconfigured, the service can only be as reliable as the remaining pieces allow.

It is intentionally narrow

This project is designed for one workflow: