FIELD NOTE / TIKTOK
The Jev-plus-LLM tax is easy to hide.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved TikTok edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
The Jev-plus-LLM tax is easy to hide.
Video caption
The Jev-plus-LLM tax is easy to hide. Status read, decision, wording and verification can disagree. Measure accepted explanations after every hop. #EricFieldNotes
Full written post / accessibility read
A customer asks: my refund was approved, but the money never arrived. Jev can choose a workflow label. The customer still needs current payment state, a clear explanation, and possibly an escalation.
The status read says the processor has not settled. Jev selects a category. A writer then produces reassuring language. If the wording ignores the live status, someone must catch it before the customer sees it. The handoffs are part of the design cost.
Run the same cases through Jev plus writer plus verifier, and through a constrained small-model reply plus the same verifier. Read payment state before either path. Keep the refund authorization rule in code. Compare accepted replies, exception handling, latency and cost.
Use the split when its control or quality gain exceeds its extra latency and review work. Otherwise simplify. Do this because optimizing the decision API alone can make the customer-facing service slower and harder to audit.
#EricFieldNotes
Four-beat scene transcript
1. The Jev-plus-LLM tax is easy to hide.
A customer asks: my refund was approved, but the money never arrived. Jev can choose a workflow label. The customer still needs current payment state, a clear explanation, and possibly an escalation.
Visual: A choice is not the customer's explanation.
2. A split path creates extra seams.
The status read says the processor has not settled. Jev selects a category. A writer then produces reassuring language. If the wording ignores the live status, someone must catch it before the customer sees it. The handoffs are part of the design cost.
Visual: Status read, decision, wording and verification can disagree.
3. Compare two complete workflows.
Run the same cases through Jev plus writer plus verifier, and through a constrained small-model reply plus the same verifier. Read payment state before either path. Keep the refund authorization rule in code. Compare accepted replies, exception handling, latency and cost.
Visual: One small model may classify and explain in one pass.
4. Pay for a split only when it earns it.
Use the split when its control or quality gain exceeds its extra latency and review work. Otherwise simplify. Do this because optimizing the decision API alone can make the customer-facing service slower and harder to audit.
Visual: Measure accepted explanations after every hop.
Research and claim limits
- TypeSafe AI: Jev 1.13 jaggedness (S99)
- Anthropic: Models overview (S104)
- TypeSafe AI: Jev with coding agents (S106)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.