Do you need a decision or a sentence?

Video caption

Do you need a decision or a sentence? A correct Choice and a wrong explanation can coexist. Give each model a job its output can finish. #EricFieldNotes

Full written post / accessibility read

Jev is designed to answer bounded typed questions. It does not write the customer explanation. If the job is mostly a useful paragraph, adding a typed judge before a writer may create another moving part without improving the delivered work.

Consider a fictional warranty triage. The typed route chooses repair, then a writer says replacement and invents a delivery date. The system now has two artifacts to reconcile. A small generative model could also fail, but its single artifact makes a different validation path possible.

Create three slices. For decision-only tasks, test typed answers against owner labels. For explanation tasks, test a small writer directly. For combined tasks, test both architectures with consistency checks and human review. Count final accepted cases, not which model sounded clever.

Use a typed judge when a typed decision is the valuable unit, and add generation only when the product needs it. Do this because architecture cost follows the complete artifact and the checks needed to trust it.

#EricFieldNotes

Four-beat scene transcript

1. Do you need a decision or a sentence?

Jev is designed to answer bounded typed questions. It does not write the customer explanation. If the job is mostly a useful paragraph, adding a typed judge before a writer may create another moving part without improving the delivered work.

Visual: A typed model can be excellent at the wrong artifact.

2. The route can double the review surface.

Consider a fictional warranty triage. The typed route chooses repair, then a writer says replacement and invents a delivery date. The system now has two artifacts to reconcile. A small generative model could also fail, but its single artifact makes a different validation path possible.

Visual: A correct Choice and a wrong explanation can coexist.

3. Route by deliverable shape.

Create three slices. For decision-only tasks, test typed answers against owner labels. For explanation tasks, test a small writer directly. For combined tasks, test both architectures with consistency checks and human review. Count final accepted cases, not which model sounded clever.

Visual: Decision-only, text-only and combined cases differ.

4. Do not pay for an unused primitive.

Use a typed judge when a typed decision is the valuable unit, and add generation only when the product needs it. Do this because architecture cost follows the complete artifact and the checks needed to trust it.

Visual: Give each model a job its output can finish.

Research and claim limits

• TypeSafe AI: Jev documentation (S98)
• TypeSafe AI: Jev 1.13 jaggedness (S99)
• Anthropic: Models overview (S104)
• Qwen: Qwen3 model family (S105)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.