The cheapest decision call may cost more per accepted case.

Video caption

The cheapest decision call may cost more per accepted case. A fast Choice can still require a second model and review. Choose a route by completed work under your error budget. #EricFieldNotes

Full written post / accessibility read

Jev returns typed judgments, while a small generative model can also draft the explanation. Comparing their API prices per call misses what the customer actually receives. The denominator should be cases accepted after validation, not raw model responses.

Suppose a support decision needs a chosen route and a customer-ready explanation. A typed judge returns the route, then a writer drafts text. A small model might return both. Either path can win, depending on error rate, validation, latency and escalation. You cannot infer the winner from the first call.

Freeze the same owner-labeled cases and source packet for both routes. Record total spend and time until the exact deliverable is accepted. Keep wrong-but-accepted decisions as a separate harm-weighted line, rather than hiding them inside a low cost average.

Compare cost per independently accepted case, tail latency and wrong-branch severity. Do this because a cheap first answer that triggers extra writing or repair is not necessarily a cheap service.

#EricFieldNotes

Four-beat scene transcript

1. The cheapest decision call may cost more per accepted case.

Jev returns typed judgments, while a small generative model can also draft the explanation. Comparing their API prices per call misses what the customer actually receives. The denominator should be cases accepted after validation, not raw model responses.

Visual: Retries, explanation and review change the denominator.

2. One path takes an extra writer.

Suppose a support decision needs a chosen route and a customer-ready explanation. A typed judge returns the route, then a writer drafts text. A small model might return both. Either path can win, depending on error rate, validation, latency and escalation. You cannot infer the winner from the first call.

Visual: A fast Choice can still require a second model and review.

3. Run a matched-case ledger.

Freeze the same owner-labeled cases and source packet for both routes. Record total spend and time until the exact deliverable is accepted. Keep wrong-but-accepted decisions as a separate harm-weighted line, rather than hiding them inside a low cost average.

Visual: Sum every call, wait, retry, human minute and reversal.

4. Optimize the accepted outcome.

Compare cost per independently accepted case, tail latency and wrong-branch severity. Do this because a cheap first answer that triggers extra writing or repair is not necessarily a cheap service.

Visual: Choose a route by completed work under your error budget.

Research and claim limits

• TypeSafe AI: Introducing System One Models and Jev (S97)
• Anthropic: Models overview (S104)
• Qwen: Qwen3 model family (S105)
• TypeSafe AI: Jev interface (S176)

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