Mailroom Reloaded: appendix
Gate configuration · deferral policy · verification scope
Supporting tables for the Mailroom Reloaded case study. These are configuration values, not measured outcomes.
Routing gate
| Setting | Value |
|---|---|
| High-confidence band | 0.95 |
| Low-confidence band | 0.70 |
| Maximum retries | 2 |
| Judge high band | 0.95 |
A document above the high band can take the deterministic path; below the low band it goes to the LLM or human review. The bands are configuration, and a real deployment has to fit them on its own validation split.
Deferral policy
Contract and merger-agreement documents defer by class, regardless of confidence. The classifier work showed why: subclass heads for those classes are the weakest, so a confident-looking answer is not a reliable one.
Opt-in Jev gate
The Jev decision model is off by default and only acts on the medium-confidence band once models/jev_calibration.json exists. It never overrides a hard rule. Its threshold is 0.8. All of this is opt-in; without the calibration file the gate is inert.
Verification scope
The case study and technical report preserve a dated verification run. They are not a current passing-test total, because the suite has changed since. The repository has no hosted CI; the local suite is the gate.
Screenshots in the report come from a seeded demo run with the mock provider. They show the interface, not document quality, and /ui and /tui have changed since the capture.