Jev vs Rule-Based Code
Use deterministic rules for computable facts and semantic judgments for meaning that is difficult to encode.
On this page
Start with what you can computeWhere semantic judgment helpsA hybrid pipelineTest rule and model failures separatelyWhen rules are enoughStart with what you can compute
Rules such as amount > 100, country == "US", and date < deadline are exact and testable. A database lookup or arithmetic function should not become a model question merely because a model is available.
Where semantic judgment helps
“Does this customer sound urgent?”, “Which department owns this issue?”, and “Does this passage support the claim?” depend on language meaning. Jev can be useful when brittle keyword rules fail to express that meaning and a bounded answer is enough.
| Input | Deterministic part | Semantic part |
|---|---|---|
| Refund ticket | Check payment IDs and amount | Detect explicit refund request |
| Tool proposal | Enforce allowlist and ownership | Assess contextual risk |
| Natural-language date | Calendar arithmetic | Select the referenced weekday |
| Invoice email | Parse exact candidate addresses | Select the billing contact |
A hybrid pipeline
First use code to parse and validate. Send only the relevant evidence to Jev. Then use code to interpret the typed result, apply policy, and execute permitted actions. This keeps precise facts precise while using semantic judgment only where it adds value.
Test rule and model failures separately
A parser can miss a candidate even when the model would select it correctly. A classifier can choose the wrong candidate even when extraction is perfect. Measure both stages; an end-to-end score alone does not tell you which boundary needs fixing.
When rules are enough
If your decision has a stable exact condition, keep the rule. It is easier to explain, cheaper to execute, and reproducible. Use Jev when you can clearly name the semantic ambiguity and have examples that demonstrate why simpler rules are inadequate.