Skip to content

Jev vs an LLM baseline

To put Jev's accuracy in context, the same states and questions were answered by Claude Haiku 4.5, which you might otherwise use for a classification step. Each request was a headless Claude Code call with no tools, a minimal system prompt and extended thinking disabled. Haiku returned the same typed answer space as JSON (a probability for Nouls, an option and confidence for Choices, a level and confidence for Scores).

Accuracy

Hand cases, and the test split of the generated cases:

Scenario Question Primitive Jev Haiku 4.5
1. Ward (hand) new_confusion Noul 100% 100%
1. Ward (hand) infection Noul 90% 85%
1. Ward (hand) concern Score 85% · MAE 0.18 95% · MAE 0.05
1. Ward (hand) pattern Choice 95% 100%
2. Discharge (hand) med_status Choice 98% 99%
2. Discharge (hand) allergy_conflict Noul 100% 90%
2. Discharge (hand) duplicate_therapy Noul 70% 85%
2. Discharge (hand) interaction Noul 75% 90%
2. Discharge (hand) justification Score 55% · MAE 0.46 55% · MAE 0.45
3. Inbox (hand) route Choice 70% 85%
3. Inbox (hand) urgency Score 70% · MAE 0.32 75% · MAE 0.25
3. Inbox (hand) red_flag Noul 85% 90%
3. Inbox (hand) medication_issue Noul 95% 90%
3. Inbox (hand) safeguarding Noul 80% 95%
1. Ward (test) new_confusion Noul 95% 95%
1. Ward (test) infection Noul 100% 85%
1. Ward (test) concern Score 78% · MAE 0.25 80% · MAE 0.20
1. Ward (test) pattern Choice 92% 85%
2. Discharge (test) med_status Choice 98% 92%
2. Discharge (test) allergy_conflict Noul 100% 98%
2. Discharge (test) duplicate_therapy Noul 85% 80%
2. Discharge (test) interaction Noul 68% 75%
2. Discharge (test) justification Score 60% · MAE 0.47 42% · MAE 0.65
3. Inbox (test) route Choice 85% 92%
3. Inbox (test) urgency Score 80% · MAE 0.22 82% · MAE 0.17
3. Inbox (test) red_flag Noul 82% 85%
3. Inbox (test) medication_issue Noul 95% 90%
3. Inbox (test) safeguarding Noul 72% 100%

Note search (note-level F1): Jev 0.86, Haiku 0.83.

The two models are close on most questions, and neither dominates:

  • Haiku is stronger on routing and on the questions where Jev over-calls (inbox route, safeguarding, discharge interactions).
  • Jev is stronger on high-volume extraction over the generated plans (per-medication status 98% vs 92%), on the infection and allergy checks, on grading justifications, and on note search.

Latency and cost

Run Requests p50 ms (Jev / Haiku) p95 ms (Jev / Haiku) Cost (Jev / Haiku) Haiku cost multiple
s1 ward 20 334 / 1587 588 / 1952 $0.0009 / $0.0399 43×
s2 discharge 20 324 / 2692 388 / 3512 $0.0022 / $0.0875 41×
s3 inbox 20 333 / 1798 402 / 2792 $0.0007 / $0.0398 56×
s1 ward gen 80 321 / 1560 405 / 1732 $0.0034 / $0.1538 45×
s2 discharge gen 80 320 / 2512 367 / 3690 $0.0083 / $0.3361 41×
s3 inbox gen 80 318 / 1744 365 / 1922 $0.0029 / $0.1565 53×
s4 search 300 324 / 2216 436 / 2635 $0.0560 / $3.2938 59×

Haiku latency is the API time reported by the CLI (duration_api_ms), which excludes CLI start-up. Its cost is the CLI's list-price estimate, including prompt-cache writes. Haiku token counts aren't compared: the runner recorded only uncached input tokens (as few as 3 for a request carrying ten notes), so they understate Haiku's real input.

The trade-off: Jev answers in a fifth to an eighth of the time, at 1/40th to 1/60th of the cost, with lower accuracy on some judgments and higher on others. That suits the architecture used throughout this report: Jev for every judgment, and a reasoning model only for the uncertain slice.